├── tools ├── __init__.py ├── dist_train.sh ├── dist_test.sh ├── slurm_test.sh ├── slurm_train.sh ├── coco_eval.py ├── publish_model.py ├── upgrade_model_version.py ├── get_flops.py ├── voc_eval.py ├── divide_dataset.py ├── train.py └── detectron2pytorch.py ├── .gitignore ├── mmdet_extra ├── ps_roi_align_ori │ ├── __init__.py │ ├── vision.cpp │ ├── vision.h │ ├── setup.py │ ├── PSROIAlign.h │ └── ps_roi_align.py ├── __init__.py ├── shared_fc_bbox_head_modified.py └── utils_extra.py ├── mmdet ├── models │ ├── shared_heads │ │ ├── __init__.py │ │ └── res_layer.py │ ├── roi_extractors │ │ └── __init__.py │ ├── necks │ │ ├── __init__.py │ │ ├── hrfpn.py │ │ └── bfp.py │ ├── plugins │ │ ├── __init__.py │ │ └── non_local.py │ ├── backbones │ │ └── __init__.py │ ├── bbox_heads │ │ └── __init__.py │ ├── registry.py │ ├── utils │ │ ├── scale.py │ │ ├── __init__.py │ │ ├── conv_ws.py │ │ ├── weight_init.py │ │ └── norm.py │ ├── mask_heads │ │ ├── __init__.py │ │ ├── htc_mask_head.py │ │ └── fused_semantic_head.py │ ├── detectors │ │ ├── fcos.py │ │ ├── retinanet.py │ │ ├── __init__.py │ │ ├── faster_rcnn.py │ │ ├── mask_rcnn.py │ │ ├── fast_rcnn.py │ │ ├── single_stage.py │ │ ├── reppoints_detector.py │ │ └── rpn.py │ ├── anchor_heads │ │ ├── __init__.py │ │ └── retina_head.py │ ├── losses │ │ ├── mse_loss.py │ │ ├── accuracy.py │ │ ├── __init__.py │ │ ├── smooth_l1_loss.py │ │ ├── balanced_l1_loss.py │ │ ├── focal_loss.py │ │ ├── utils.py │ │ └── cross_entropy_loss.py │ ├── __init__.py │ └── builder.py ├── ops │ ├── nms │ │ ├── __init__.py │ │ ├── src │ │ │ ├── nms_cuda.cpp │ │ │ └── nms_cpu.cpp │ │ └── nms_wrapper.py │ ├── roi_pool │ │ ├── __init__.py │ │ ├── gradcheck.py │ │ ├── roi_pool.py │ │ └── src │ │ │ └── roi_pool_cuda.cpp │ ├── roi_align │ │ ├── __init__.py │ │ ├── gradcheck.py │ │ ├── src │ │ │ └── roi_align_cuda.cpp │ │ └── roi_align.py │ ├── masked_conv │ │ ├── __init__.py │ │ ├── src │ │ │ └── masked_conv2d_cuda.cpp │ │ └── masked_conv.py │ ├── sigmoid_focal_loss │ │ ├── __init__.py │ │ ├── sigmoid_focal_loss.py │ │ └── src │ │ │ └── sigmoid_focal_loss.cpp │ ├── dcn │ │ └── __init__.py │ ├── __init__.py │ └── context_block.py ├── __init__.py ├── datasets │ ├── registry.py │ ├── loader │ │ ├── __init__.py │ │ └── build_loader.py │ ├── cityscapes.py │ ├── voc.py │ ├── __init__.py │ ├── pipelines │ │ ├── __init__.py │ │ ├── compose.py │ │ └── test_aug.py │ ├── wider_face.py │ ├── builder.py │ ├── dataset_wrappers.py │ └── xml_style.py ├── core │ ├── mask │ │ ├── __init__.py │ │ ├── utils.py │ │ └── mask_target.py │ ├── fp16 │ │ ├── __init__.py │ │ └── utils.py │ ├── utils │ │ ├── __init__.py │ │ ├── misc.py │ │ └── dist_utils.py │ ├── bbox │ │ ├── assigners │ │ │ ├── base_assigner.py │ │ │ ├── __init__.py │ │ │ └── assign_result.py │ │ ├── samplers │ │ │ ├── combined_sampler.py │ │ │ ├── __init__.py │ │ │ ├── sampling_result.py │ │ │ ├── pseudo_sampler.py │ │ │ ├── instance_balanced_pos_sampler.py │ │ │ ├── random_sampler.py │ │ │ ├── ohem_sampler.py │ │ │ └── base_sampler.py │ │ ├── __init__.py │ │ ├── assign_sampling.py │ │ ├── geometry.py │ │ └── bbox_target.py │ ├── __init__.py │ ├── post_processing │ │ ├── __init__.py │ │ ├── bbox_nms.py │ │ └── merge_augs.py │ ├── anchor │ │ ├── __init__.py │ │ ├── point_generator.py │ │ └── anchor_generator.py │ └── evaluation │ │ ├── __init__.py │ │ └── bbox_overlaps.py ├── utils │ ├── __init__.py │ └── registry.py └── apis │ ├── __init__.py │ └── env.py ├── requirements.txt ├── scripts ├── calculate_flops.sh └── test.sh └── README.md /tools/__init__.py: -------------------------------------------------------------------------------- 1 | -------------------------------------------------------------------------------- /.gitignore: -------------------------------------------------------------------------------- 1 | *.DS_Store -------------------------------------------------------------------------------- /mmdet_extra/ps_roi_align_ori/__init__.py: -------------------------------------------------------------------------------- 1 | -------------------------------------------------------------------------------- /mmdet/models/shared_heads/__init__.py: -------------------------------------------------------------------------------- 1 | from .res_layer import ResLayer 2 | 3 | __all__ = ['ResLayer'] 4 | -------------------------------------------------------------------------------- /mmdet/ops/nms/__init__.py: -------------------------------------------------------------------------------- 1 | from .nms_wrapper import nms, soft_nms 2 | 3 | __all__ = ['nms', 'soft_nms'] 4 | -------------------------------------------------------------------------------- /mmdet/ops/roi_pool/__init__.py: -------------------------------------------------------------------------------- 1 | from .roi_pool import RoIPool, roi_pool 2 | 3 | __all__ = ['roi_pool', 'RoIPool'] 4 | -------------------------------------------------------------------------------- /mmdet/__init__.py: -------------------------------------------------------------------------------- 1 | from .version import __version__, short_version 2 | 3 | __all__ = ['__version__', 'short_version'] 4 | -------------------------------------------------------------------------------- /mmdet/ops/roi_align/__init__.py: -------------------------------------------------------------------------------- 1 | from .roi_align import RoIAlign, roi_align 2 | 3 | __all__ = ['roi_align', 'RoIAlign'] 4 | -------------------------------------------------------------------------------- /mmdet/models/roi_extractors/__init__.py: -------------------------------------------------------------------------------- 1 | from .single_level import SingleRoIExtractor 2 | 3 | __all__ = ['SingleRoIExtractor'] 4 | -------------------------------------------------------------------------------- /mmdet/datasets/registry.py: -------------------------------------------------------------------------------- 1 | from mmdet.utils import Registry 2 | 3 | DATASETS = Registry('dataset') 4 | PIPELINES = Registry('pipeline') 5 | -------------------------------------------------------------------------------- /mmdet/ops/masked_conv/__init__.py: -------------------------------------------------------------------------------- 1 | from .masked_conv import MaskedConv2d, masked_conv2d 2 | 3 | __all__ = ['masked_conv2d', 'MaskedConv2d'] 4 | -------------------------------------------------------------------------------- /mmdet/models/necks/__init__.py: -------------------------------------------------------------------------------- 1 | from .bfp import BFP 2 | from .fpn import FPN 3 | from .hrfpn import HRFPN 4 | 5 | __all__ = ['FPN', 'BFP', 'HRFPN'] 6 | -------------------------------------------------------------------------------- /mmdet/core/mask/__init__.py: -------------------------------------------------------------------------------- 1 | from .mask_target import mask_target 2 | from .utils import split_combined_polys 3 | 4 | __all__ = ['split_combined_polys', 'mask_target'] 5 | -------------------------------------------------------------------------------- /mmdet/ops/sigmoid_focal_loss/__init__.py: -------------------------------------------------------------------------------- 1 | from .sigmoid_focal_loss import SigmoidFocalLoss, sigmoid_focal_loss 2 | 3 | __all__ = ['SigmoidFocalLoss', 'sigmoid_focal_loss'] 4 | -------------------------------------------------------------------------------- /requirements.txt: -------------------------------------------------------------------------------- 1 | mmcv==0.2.13 2 | numpy==1.17 3 | matplotlib 4 | six 5 | terminaltables 6 | pycocotools 7 | torch>=1.1 8 | torchvision 9 | imagecorruptions 10 | pillow==6.1 -------------------------------------------------------------------------------- /mmdet/models/plugins/__init__.py: -------------------------------------------------------------------------------- 1 | from .generalized_attention import GeneralizedAttention 2 | from .non_local import NonLocal2D 3 | 4 | __all__ = ['NonLocal2D', 'GeneralizedAttention'] 5 | -------------------------------------------------------------------------------- /mmdet/utils/__init__.py: -------------------------------------------------------------------------------- 1 | from .flops_counter import get_model_complexity_info 2 | from .registry import Registry, build_from_cfg 3 | 4 | __all__ = ['Registry', 'build_from_cfg', 'get_model_complexity_info'] 5 | -------------------------------------------------------------------------------- /mmdet/core/fp16/__init__.py: -------------------------------------------------------------------------------- 1 | from .decorators import auto_fp16, force_fp32 2 | from .hooks import Fp16OptimizerHook, wrap_fp16_model 3 | 4 | __all__ = ['auto_fp16', 'force_fp32', 'Fp16OptimizerHook', 'wrap_fp16_model'] 5 | -------------------------------------------------------------------------------- /mmdet/datasets/loader/__init__.py: -------------------------------------------------------------------------------- 1 | from .build_loader import build_dataloader 2 | from .sampler import DistributedGroupSampler, GroupSampler 3 | 4 | __all__ = ['GroupSampler', 'DistributedGroupSampler', 'build_dataloader'] 5 | -------------------------------------------------------------------------------- /mmdet/models/backbones/__init__.py: -------------------------------------------------------------------------------- 1 | from .hrnet import HRNet 2 | from .resnet import ResNet, make_res_layer 3 | from .resnext import ResNeXt 4 | from .ssd_vgg import SSDVGG 5 | 6 | __all__ = ['ResNet', 'make_res_layer', 'ResNeXt', 'SSDVGG', 'HRNet'] 7 | -------------------------------------------------------------------------------- /mmdet/core/utils/__init__.py: -------------------------------------------------------------------------------- 1 | from .dist_utils import DistOptimizerHook, allreduce_grads 2 | from .misc import multi_apply, tensor2imgs, unmap 3 | 4 | __all__ = [ 5 | 'allreduce_grads', 'DistOptimizerHook', 'tensor2imgs', 'unmap', 6 | 'multi_apply' 7 | ] 8 | -------------------------------------------------------------------------------- /tools/dist_train.sh: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env bash 2 | 3 | PYTHON=${PYTHON:-"python"} 4 | 5 | CONFIG=$1 6 | GPUS=$2 7 | 8 | $PYTHON -m torch.distributed.launch --nproc_per_node=$GPUS \ 9 | $(dirname "$0")/train.py $CONFIG --launcher pytorch --seed 2 ${@:3} 10 | -------------------------------------------------------------------------------- /mmdet/core/bbox/assigners/base_assigner.py: -------------------------------------------------------------------------------- 1 | from abc import ABCMeta, abstractmethod 2 | 3 | 4 | class BaseAssigner(metaclass=ABCMeta): 5 | 6 | @abstractmethod 7 | def assign(self, bboxes, gt_bboxes, gt_bboxes_ignore=None, gt_labels=None): 8 | pass 9 | -------------------------------------------------------------------------------- /tools/dist_test.sh: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env bash 2 | 3 | PYTHON=${PYTHON:-"python"} 4 | 5 | CONFIG=$1 6 | CHECKPOINT=$2 7 | GPUS=$3 8 | 9 | $PYTHON -m torch.distributed.launch --nproc_per_node=$GPUS \ 10 | $(dirname "$0")/test.py $CONFIG $CHECKPOINT --launcher pytorch ${@:4} 11 | -------------------------------------------------------------------------------- /mmdet/models/bbox_heads/__init__.py: -------------------------------------------------------------------------------- 1 | from .bbox_head import BBoxHead 2 | from .convfc_bbox_head import ConvFCBBoxHead, SharedFCBBoxHead 3 | from .double_bbox_head import DoubleConvFCBBoxHead 4 | 5 | __all__ = [ 6 | 'BBoxHead', 'ConvFCBBoxHead', 'SharedFCBBoxHead', 'DoubleConvFCBBoxHead' 7 | ] 8 | -------------------------------------------------------------------------------- /mmdet/datasets/cityscapes.py: -------------------------------------------------------------------------------- 1 | from .coco import CocoDataset 2 | from .registry import DATASETS 3 | 4 | 5 | @DATASETS.register_module 6 | class CityscapesDataset(CocoDataset): 7 | 8 | CLASSES = ('person', 'rider', 'car', 'truck', 'bus', 'train', 'motorcycle', 9 | 'bicycle') 10 | -------------------------------------------------------------------------------- /mmdet/models/registry.py: -------------------------------------------------------------------------------- 1 | from mmdet.utils import Registry 2 | 3 | BACKBONES = Registry('backbone') 4 | NECKS = Registry('neck') 5 | ROI_EXTRACTORS = Registry('roi_extractor') 6 | SHARED_HEADS = Registry('shared_head') 7 | HEADS = Registry('head') 8 | LOSSES = Registry('loss') 9 | DETECTORS = Registry('detector') 10 | -------------------------------------------------------------------------------- /mmdet/core/__init__.py: -------------------------------------------------------------------------------- 1 | from .anchor import * # noqa: F401, F403 2 | from .bbox import * # noqa: F401, F403 3 | from .evaluation import * # noqa: F401, F403 4 | from .fp16 import * # noqa: F401, F403 5 | from .mask import * # noqa: F401, F403 6 | from .post_processing import * # noqa: F401, F403 7 | from .utils import * # noqa: F401, F403 8 | -------------------------------------------------------------------------------- /mmdet/core/post_processing/__init__.py: -------------------------------------------------------------------------------- 1 | from .bbox_nms import multiclass_nms 2 | from .merge_augs import (merge_aug_bboxes, merge_aug_masks, 3 | merge_aug_proposals, merge_aug_scores) 4 | 5 | __all__ = [ 6 | 'multiclass_nms', 'merge_aug_proposals', 'merge_aug_bboxes', 7 | 'merge_aug_scores', 'merge_aug_masks' 8 | ] 9 | -------------------------------------------------------------------------------- /mmdet/models/utils/scale.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import torch.nn as nn 3 | 4 | 5 | class Scale(nn.Module): 6 | 7 | def __init__(self, scale=1.0): 8 | super(Scale, self).__init__() 9 | self.scale = nn.Parameter(torch.tensor(scale, dtype=torch.float)) 10 | 11 | def forward(self, x): 12 | return x * self.scale 13 | -------------------------------------------------------------------------------- /scripts/calculate_flops.sh: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env sh 2 | 3 | 4 | # calculate flops of TinyDet-M 5 | python tools/get_flops.py configs/tinydet_M.py --shape 320 6 | 7 | # calculate flops of TinyDet-S 8 | # python tools/get_flops.py configs/tinydet_S.py --shape 320 9 | 10 | # # calculate flops of TinyDet-L 11 | # python tools/get_flops.py configs/tinydet_L.py --shape 512 -------------------------------------------------------------------------------- /mmdet/models/mask_heads/__init__.py: -------------------------------------------------------------------------------- 1 | from .fcn_mask_head import FCNMaskHead 2 | from .fused_semantic_head import FusedSemanticHead 3 | from .grid_head import GridHead 4 | from .htc_mask_head import HTCMaskHead 5 | from .maskiou_head import MaskIoUHead 6 | 7 | __all__ = [ 8 | 'FCNMaskHead', 'HTCMaskHead', 'FusedSemanticHead', 'GridHead', 9 | 'MaskIoUHead' 10 | ] 11 | -------------------------------------------------------------------------------- /mmdet_extra/ps_roi_align_ori/vision.cpp: -------------------------------------------------------------------------------- 1 | // Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved. 2 | #include "PSROIAlign.h" 3 | 4 | PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) { 5 | 6 | m.def("ps_roi_align_forward", &PSROIAlign_forward, "PSROIAlign_forward"); 7 | m.def("ps_roi_align_backward", &PSROIAlign_backward, "PSROIAlign_backward"); 8 | 9 | } 10 | -------------------------------------------------------------------------------- /mmdet/core/bbox/assigners/__init__.py: -------------------------------------------------------------------------------- 1 | from .approx_max_iou_assigner import ApproxMaxIoUAssigner 2 | from .assign_result import AssignResult 3 | from .base_assigner import BaseAssigner 4 | from .max_iou_assigner import MaxIoUAssigner 5 | from .point_assigner import PointAssigner 6 | 7 | __all__ = [ 8 | 'BaseAssigner', 'MaxIoUAssigner', 'ApproxMaxIoUAssigner', 'AssignResult', 9 | 'PointAssigner' 10 | ] 11 | -------------------------------------------------------------------------------- /mmdet/apis/__init__.py: -------------------------------------------------------------------------------- 1 | from .env import get_root_logger, init_dist, set_random_seed 2 | from .inference import (inference_detector, init_detector, show_result, 3 | show_result_pyplot) 4 | from .train import train_detector 5 | 6 | __all__ = [ 7 | 'init_dist', 'get_root_logger', 'set_random_seed', 'train_detector', 8 | 'init_detector', 'inference_detector', 'show_result', 'show_result_pyplot' 9 | ] 10 | -------------------------------------------------------------------------------- /mmdet/core/anchor/__init__.py: -------------------------------------------------------------------------------- 1 | from .anchor_generator import AnchorGenerator 2 | from .anchor_target import anchor_inside_flags, anchor_target 3 | from .guided_anchor_target import ga_loc_target, ga_shape_target 4 | from .point_generator import PointGenerator 5 | from .point_target import point_target 6 | 7 | __all__ = [ 8 | 'AnchorGenerator', 'anchor_target', 'anchor_inside_flags', 'ga_loc_target', 9 | 'ga_shape_target', 'PointGenerator', 'point_target' 10 | ] 11 | -------------------------------------------------------------------------------- /mmdet/models/utils/__init__.py: -------------------------------------------------------------------------------- 1 | from .conv_module import ConvModule, build_conv_layer 2 | from .conv_ws import ConvWS2d, conv_ws_2d 3 | from .norm import build_norm_layer 4 | from .scale import Scale 5 | from .weight_init import (bias_init_with_prob, kaiming_init, normal_init, 6 | uniform_init, xavier_init) 7 | 8 | __all__ = [ 9 | 'conv_ws_2d', 'ConvWS2d', 'build_conv_layer', 'ConvModule', 10 | 'build_norm_layer', 'xavier_init', 'normal_init', 'uniform_init', 11 | 'kaiming_init', 'bias_init_with_prob', 'Scale' 12 | ] 13 | -------------------------------------------------------------------------------- /mmdet/models/detectors/fcos.py: -------------------------------------------------------------------------------- 1 | from ..registry import DETECTORS 2 | from .single_stage import SingleStageDetector 3 | 4 | 5 | @DETECTORS.register_module 6 | class FCOS(SingleStageDetector): 7 | 8 | def __init__(self, 9 | backbone, 10 | neck, 11 | bbox_head, 12 | train_cfg=None, 13 | test_cfg=None, 14 | pretrained=None): 15 | super(FCOS, self).__init__(backbone, neck, bbox_head, train_cfg, 16 | test_cfg, pretrained) 17 | -------------------------------------------------------------------------------- /mmdet/models/detectors/retinanet.py: -------------------------------------------------------------------------------- 1 | from ..registry import DETECTORS 2 | from .single_stage import SingleStageDetector 3 | 4 | 5 | @DETECTORS.register_module 6 | class RetinaNet(SingleStageDetector): 7 | 8 | def __init__(self, 9 | backbone, 10 | neck, 11 | bbox_head, 12 | train_cfg=None, 13 | test_cfg=None, 14 | pretrained=None): 15 | super(RetinaNet, self).__init__(backbone, neck, bbox_head, train_cfg, 16 | test_cfg, pretrained) 17 | -------------------------------------------------------------------------------- /mmdet/ops/roi_pool/gradcheck.py: -------------------------------------------------------------------------------- 1 | import os.path as osp 2 | import sys 3 | 4 | import torch 5 | from torch.autograd import gradcheck 6 | 7 | sys.path.append(osp.abspath(osp.join(__file__, '../../'))) 8 | from roi_pool import RoIPool # noqa: E402, isort:skip 9 | 10 | feat = torch.randn(4, 16, 15, 15, requires_grad=True).cuda() 11 | rois = torch.Tensor([[0, 0, 0, 50, 50], [0, 10, 30, 43, 55], 12 | [1, 67, 40, 110, 120]]).cuda() 13 | inputs = (feat, rois) 14 | print('Gradcheck for roi pooling...') 15 | test = gradcheck(RoIPool(4, 1.0 / 8), inputs, eps=1e-5, atol=1e-3) 16 | print(test) 17 | -------------------------------------------------------------------------------- /mmdet/core/bbox/samplers/combined_sampler.py: -------------------------------------------------------------------------------- 1 | from ..assign_sampling import build_sampler 2 | from .base_sampler import BaseSampler 3 | 4 | 5 | class CombinedSampler(BaseSampler): 6 | 7 | def __init__(self, pos_sampler, neg_sampler, **kwargs): 8 | super(CombinedSampler, self).__init__(**kwargs) 9 | self.pos_sampler = build_sampler(pos_sampler, **kwargs) 10 | self.neg_sampler = build_sampler(neg_sampler, **kwargs) 11 | 12 | def _sample_pos(self, **kwargs): 13 | raise NotImplementedError 14 | 15 | def _sample_neg(self, **kwargs): 16 | raise NotImplementedError 17 | -------------------------------------------------------------------------------- /mmdet/models/anchor_heads/__init__.py: -------------------------------------------------------------------------------- 1 | from .anchor_head import AnchorHead 2 | from .fcos_head import FCOSHead 3 | from .ga_retina_head import GARetinaHead 4 | from .ga_rpn_head import GARPNHead 5 | from .guided_anchor_head import FeatureAdaption, GuidedAnchorHead 6 | from .reppoints_head import RepPointsHead 7 | from .retina_head import RetinaHead 8 | from .rpn_head import RPNHead 9 | from .ssd_head import SSDHead 10 | 11 | __all__ = [ 12 | 'AnchorHead', 'GuidedAnchorHead', 'FeatureAdaption', 'RPNHead', 13 | 'GARPNHead', 'RetinaHead', 'GARetinaHead', 'SSDHead', 'FCOSHead', 14 | 'RepPointsHead' 15 | ] 16 | -------------------------------------------------------------------------------- /tools/slurm_test.sh: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env bash 2 | 3 | set -x 4 | 5 | PARTITION=$1 6 | JOB_NAME=$2 7 | CONFIG=$3 8 | CHECKPOINT=$4 9 | GPUS=${GPUS:-8} 10 | GPUS_PER_NODE=${GPUS_PER_NODE:-8} 11 | CPUS_PER_TASK=${CPUS_PER_TASK:-5} 12 | PY_ARGS=${@:5} 13 | SRUN_ARGS=${SRUN_ARGS:-""} 14 | 15 | srun -p ${PARTITION} \ 16 | --job-name=${JOB_NAME} \ 17 | --gres=gpu:${GPUS_PER_NODE} \ 18 | --ntasks=${GPUS} \ 19 | --ntasks-per-node=${GPUS_PER_NODE} \ 20 | --cpus-per-task=${CPUS_PER_TASK} \ 21 | --kill-on-bad-exit=1 \ 22 | ${SRUN_ARGS} \ 23 | python -u tools/test.py ${CONFIG} ${CHECKPOINT} --launcher="slurm" ${PY_ARGS} 24 | -------------------------------------------------------------------------------- /mmdet/ops/dcn/__init__.py: -------------------------------------------------------------------------------- 1 | from .deform_conv import (DeformConv, DeformConvPack, ModulatedDeformConv, 2 | ModulatedDeformConvPack, deform_conv, 3 | modulated_deform_conv) 4 | from .deform_pool import (DeformRoIPooling, DeformRoIPoolingPack, 5 | ModulatedDeformRoIPoolingPack, deform_roi_pooling) 6 | 7 | __all__ = [ 8 | 'DeformConv', 'DeformConvPack', 'ModulatedDeformConv', 9 | 'ModulatedDeformConvPack', 'DeformRoIPooling', 'DeformRoIPoolingPack', 10 | 'ModulatedDeformRoIPoolingPack', 'deform_conv', 'modulated_deform_conv', 11 | 'deform_roi_pooling' 12 | ] 13 | -------------------------------------------------------------------------------- /mmdet/core/bbox/samplers/__init__.py: -------------------------------------------------------------------------------- 1 | from .base_sampler import BaseSampler 2 | from .combined_sampler import CombinedSampler 3 | from .instance_balanced_pos_sampler import InstanceBalancedPosSampler 4 | from .iou_balanced_neg_sampler import IoUBalancedNegSampler 5 | from .ohem_sampler import OHEMSampler 6 | from .pseudo_sampler import PseudoSampler 7 | from .random_sampler import RandomSampler 8 | from .sampling_result import SamplingResult 9 | 10 | __all__ = [ 11 | 'BaseSampler', 'PseudoSampler', 'RandomSampler', 12 | 'InstanceBalancedPosSampler', 'IoUBalancedNegSampler', 'CombinedSampler', 13 | 'OHEMSampler', 'SamplingResult' 14 | ] 15 | -------------------------------------------------------------------------------- /tools/slurm_train.sh: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env bash 2 | 3 | set -x 4 | 5 | PARTITION=$1 6 | JOB_NAME=$2 7 | CONFIG=$3 8 | WORK_DIR=$4 9 | GPUS=${5:-8} 10 | GPUS_PER_NODE=${GPUS_PER_NODE:-8} 11 | CPUS_PER_TASK=${CPUS_PER_TASK:-5} 12 | SRUN_ARGS=${SRUN_ARGS:-""} 13 | PY_ARGS=${PY_ARGS:-"--validate"} 14 | 15 | srun -p ${PARTITION} \ 16 | --job-name=${JOB_NAME} \ 17 | --gres=gpu:${GPUS_PER_NODE} \ 18 | --ntasks=${GPUS} \ 19 | --ntasks-per-node=${GPUS_PER_NODE} \ 20 | --cpus-per-task=${CPUS_PER_TASK} \ 21 | --kill-on-bad-exit=1 \ 22 | ${SRUN_ARGS} \ 23 | python -u tools/train.py ${CONFIG} --work_dir=${WORK_DIR} --launcher="slurm" ${PY_ARGS} 24 | -------------------------------------------------------------------------------- /mmdet/ops/nms/src/nms_cuda.cpp: -------------------------------------------------------------------------------- 1 | // Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved. 2 | #include 3 | 4 | #define CHECK_CUDA(x) AT_CHECK(x.type().is_cuda(), #x, " must be a CUDAtensor ") 5 | 6 | at::Tensor nms_cuda(const at::Tensor boxes, float nms_overlap_thresh); 7 | 8 | at::Tensor nms(const at::Tensor& dets, const float threshold) { 9 | CHECK_CUDA(dets); 10 | if (dets.numel() == 0) 11 | return at::empty({0}, dets.options().dtype(at::kLong).device(at::kCPU)); 12 | return nms_cuda(dets, threshold); 13 | } 14 | 15 | PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) { 16 | m.def("nms", &nms, "non-maximum suppression"); 17 | } -------------------------------------------------------------------------------- /mmdet_extra/__init__.py: -------------------------------------------------------------------------------- 1 | from .shared_fc_bbox_head_modified import SharedFCBBoxHeadModified 2 | from .single_level_modified import SingleRoIExtractorModified 3 | from .flops_counter_modified import get_model_complexity_info 4 | 5 | from .ps_roi_align_ori.ps_roi_align import PSROIAlign_ori 6 | from mmdet import ops 7 | ops.PSROIAlign_ori = PSROIAlign_ori 8 | 9 | 10 | from .utils_extra import convert_sync_batchnorm, replace_relu6_with_relu, \ 11 | remove_relu_series, get_backbone_madds, AttentionModule 12 | 13 | 14 | from .tinyhrdet import TinyHRDet 15 | from .mobilenetv3_bc import MobileNetV3BC 16 | from .mobilenetv3_d import MobileNetV3D 17 | from .tinyfpn import TinyFPN 18 | from .tinyrpn import TinyRPN 19 | 20 | 21 | -------------------------------------------------------------------------------- /mmdet/models/losses/mse_loss.py: -------------------------------------------------------------------------------- 1 | import torch.nn as nn 2 | import torch.nn.functional as F 3 | 4 | from ..registry import LOSSES 5 | from .utils import weighted_loss 6 | 7 | mse_loss = weighted_loss(F.mse_loss) 8 | 9 | 10 | @LOSSES.register_module 11 | class MSELoss(nn.Module): 12 | 13 | def __init__(self, reduction='mean', loss_weight=1.0): 14 | super().__init__() 15 | self.reduction = reduction 16 | self.loss_weight = loss_weight 17 | 18 | def forward(self, pred, target, weight=None, avg_factor=None): 19 | loss = self.loss_weight * mse_loss( 20 | pred, 21 | target, 22 | weight, 23 | reduction=self.reduction, 24 | avg_factor=avg_factor) 25 | return loss 26 | -------------------------------------------------------------------------------- /mmdet/core/fp16/utils.py: -------------------------------------------------------------------------------- 1 | from collections import abc 2 | 3 | import numpy as np 4 | import torch 5 | 6 | 7 | def cast_tensor_type(inputs, src_type, dst_type): 8 | if isinstance(inputs, torch.Tensor): 9 | return inputs.to(dst_type) 10 | elif isinstance(inputs, str): 11 | return inputs 12 | elif isinstance(inputs, np.ndarray): 13 | return inputs 14 | elif isinstance(inputs, abc.Mapping): 15 | return type(inputs)({ 16 | k: cast_tensor_type(v, src_type, dst_type) 17 | for k, v in inputs.items() 18 | }) 19 | elif isinstance(inputs, abc.Iterable): 20 | return type(inputs)( 21 | cast_tensor_type(item, src_type, dst_type) for item in inputs) 22 | else: 23 | return inputs 24 | -------------------------------------------------------------------------------- /mmdet/datasets/voc.py: -------------------------------------------------------------------------------- 1 | from .registry import DATASETS 2 | from .xml_style import XMLDataset 3 | 4 | 5 | @DATASETS.register_module 6 | class VOCDataset(XMLDataset): 7 | 8 | CLASSES = ('aeroplane', 'bicycle', 'bird', 'boat', 'bottle', 'bus', 'car', 9 | 'cat', 'chair', 'cow', 'diningtable', 'dog', 'horse', 10 | 'motorbike', 'person', 'pottedplant', 'sheep', 'sofa', 'train', 11 | 'tvmonitor') 12 | 13 | def __init__(self, **kwargs): 14 | super(VOCDataset, self).__init__(**kwargs) 15 | if 'VOC2007' in self.img_prefix: 16 | self.year = 2007 17 | elif 'VOC2012' in self.img_prefix: 18 | self.year = 2012 19 | else: 20 | raise ValueError('Cannot infer dataset year from img_prefix') 21 | -------------------------------------------------------------------------------- /mmdet/datasets/__init__.py: -------------------------------------------------------------------------------- 1 | from .builder import build_dataset 2 | from .cityscapes import CityscapesDataset 3 | from .coco import CocoDataset 4 | from .custom import CustomDataset 5 | from .dataset_wrappers import ConcatDataset, RepeatDataset 6 | from .extra_aug import ExtraAugmentation 7 | from .loader import DistributedGroupSampler, GroupSampler, build_dataloader 8 | from .registry import DATASETS 9 | from .voc import VOCDataset 10 | from .wider_face import WIDERFaceDataset 11 | from .xml_style import XMLDataset 12 | 13 | __all__ = [ 14 | 'CustomDataset', 'XMLDataset', 'CocoDataset', 'VOCDataset', 15 | 'CityscapesDataset', 'GroupSampler', 'DistributedGroupSampler', 16 | 'build_dataloader', 'ConcatDataset', 'RepeatDataset', 'ExtraAugmentation', 17 | 'WIDERFaceDataset', 'DATASETS', 'build_dataset' 18 | ] 19 | -------------------------------------------------------------------------------- /scripts/test.sh: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env sh 2 | 3 | #test TinyDet-M 4 | python -m torch.distributed.launch --nproc_per_node=1 ./tools/test.py \ 5 | ./configs/tinydet_M.py \ 6 | ./pth_file/tinydet_M.pth \ 7 | --launcher pytorch 8 | 9 | 10 | # #test TinyDet-S 11 | # python -m torch.distributed.launch --nproc_per_node=1 ./tools/test.py \ 12 | # ./configs/tinydet_S.py \ 13 | # ./pth_file/tinydet_S.pth \ 14 | # --launcher pytorch 15 | 16 | 17 | # #test TinyDet-L 18 | # python -m torch.distributed.launch --nproc_per_node=1 ./tools/test.py \ 19 | # ./configs/tinydet_L.py \ 20 | # ./pth_file/tinydet_L.pth \ 21 | # --launcher pytorch 22 | -------------------------------------------------------------------------------- /mmdet/datasets/pipelines/__init__.py: -------------------------------------------------------------------------------- 1 | from .compose import Compose 2 | from .formating import (Collect, ImageToTensor, ToDataContainer, ToTensor, 3 | Transpose, to_tensor) 4 | from .loading import LoadAnnotations, LoadImageFromFile, LoadProposals 5 | from .test_aug import MultiScaleFlipAug 6 | from .transforms import (Expand, MinIoURandomCrop, Normalize, Pad, 7 | PhotoMetricDistortion, RandomCrop, RandomFlip, Resize, 8 | SegResizeFlipPadRescale) 9 | 10 | __all__ = [ 11 | 'Compose', 'to_tensor', 'ToTensor', 'ImageToTensor', 'ToDataContainer', 12 | 'Transpose', 'Collect', 'LoadAnnotations', 'LoadImageFromFile', 13 | 'LoadProposals', 'MultiScaleFlipAug', 'Resize', 'RandomFlip', 'Pad', 14 | 'RandomCrop', 'Normalize', 'SegResizeFlipPadRescale', 'MinIoURandomCrop', 15 | 'Expand', 'PhotoMetricDistortion' 16 | ] 17 | -------------------------------------------------------------------------------- /mmdet/models/detectors/__init__.py: -------------------------------------------------------------------------------- 1 | from .base import BaseDetector 2 | from .cascade_rcnn import CascadeRCNN 3 | from .double_head_rcnn import DoubleHeadRCNN 4 | from .fast_rcnn import FastRCNN 5 | from .faster_rcnn import FasterRCNN 6 | from .fcos import FCOS 7 | from .grid_rcnn import GridRCNN 8 | from .htc import HybridTaskCascade 9 | from .mask_rcnn import MaskRCNN 10 | from .mask_scoring_rcnn import MaskScoringRCNN 11 | from .reppoints_detector import RepPointsDetector 12 | from .retinanet import RetinaNet 13 | from .rpn import RPN 14 | from .single_stage import SingleStageDetector 15 | from .two_stage import TwoStageDetector 16 | 17 | __all__ = [ 18 | 'BaseDetector', 'SingleStageDetector', 'TwoStageDetector', 'RPN', 19 | 'FastRCNN', 'FasterRCNN', 'MaskRCNN', 'CascadeRCNN', 'HybridTaskCascade', 20 | 'DoubleHeadRCNN', 'RetinaNet', 'FCOS', 'GridRCNN', 'MaskScoringRCNN', 21 | 'RepPointsDetector' 22 | ] 23 | -------------------------------------------------------------------------------- /tools/coco_eval.py: -------------------------------------------------------------------------------- 1 | from argparse import ArgumentParser 2 | 3 | from mmdet.core import coco_eval 4 | 5 | 6 | def main(): 7 | parser = ArgumentParser(description='COCO Evaluation') 8 | parser.add_argument('result', help='result file path') 9 | parser.add_argument('--ann', help='annotation file path') 10 | parser.add_argument( 11 | '--types', 12 | type=str, 13 | nargs='+', 14 | choices=['proposal_fast', 'proposal', 'bbox', 'segm', 'keypoint'], 15 | default=['bbox'], 16 | help='result types') 17 | parser.add_argument( 18 | '--max-dets', 19 | type=int, 20 | nargs='+', 21 | default=[100, 300, 1000], 22 | help='proposal numbers, only used for recall evaluation') 23 | args = parser.parse_args() 24 | coco_eval(args.result, args.types, args.ann, args.max_dets) 25 | 26 | 27 | if __name__ == '__main__': 28 | main() 29 | -------------------------------------------------------------------------------- /mmdet/core/bbox/samplers/sampling_result.py: -------------------------------------------------------------------------------- 1 | import torch 2 | 3 | 4 | class SamplingResult(object): 5 | 6 | def __init__(self, pos_inds, neg_inds, bboxes, gt_bboxes, assign_result, 7 | gt_flags): 8 | self.pos_inds = pos_inds 9 | self.neg_inds = neg_inds 10 | self.pos_bboxes = bboxes[pos_inds] 11 | self.neg_bboxes = bboxes[neg_inds] 12 | self.pos_is_gt = gt_flags[pos_inds] 13 | 14 | self.num_gts = gt_bboxes.shape[0] 15 | self.pos_assigned_gt_inds = assign_result.gt_inds[pos_inds] - 1 16 | self.pos_gt_bboxes = gt_bboxes[self.pos_assigned_gt_inds, :] 17 | if assign_result.labels is not None: 18 | self.pos_gt_labels = assign_result.labels[pos_inds] 19 | else: 20 | self.pos_gt_labels = None 21 | 22 | @property 23 | def bboxes(self): 24 | return torch.cat([self.pos_bboxes, self.neg_bboxes]) 25 | -------------------------------------------------------------------------------- /mmdet/models/detectors/faster_rcnn.py: -------------------------------------------------------------------------------- 1 | from ..registry import DETECTORS 2 | from .two_stage import TwoStageDetector 3 | 4 | 5 | @DETECTORS.register_module 6 | class FasterRCNN(TwoStageDetector): 7 | 8 | def __init__(self, 9 | backbone, 10 | rpn_head, 11 | bbox_roi_extractor, 12 | bbox_head, 13 | train_cfg, 14 | test_cfg, 15 | neck=None, 16 | shared_head=None, 17 | pretrained=None): 18 | super(FasterRCNN, self).__init__( 19 | backbone=backbone, 20 | neck=neck, 21 | shared_head=shared_head, 22 | rpn_head=rpn_head, 23 | bbox_roi_extractor=bbox_roi_extractor, 24 | bbox_head=bbox_head, 25 | train_cfg=train_cfg, 26 | test_cfg=test_cfg, 27 | pretrained=pretrained) 28 | -------------------------------------------------------------------------------- /mmdet/models/losses/accuracy.py: -------------------------------------------------------------------------------- 1 | import torch.nn as nn 2 | 3 | 4 | def accuracy(pred, target, topk=1): 5 | assert isinstance(topk, (int, tuple)) 6 | if isinstance(topk, int): 7 | topk = (topk, ) 8 | return_single = True 9 | else: 10 | return_single = False 11 | 12 | maxk = max(topk) 13 | _, pred_label = pred.topk(maxk, dim=1) 14 | pred_label = pred_label.t() 15 | correct = pred_label.eq(target.view(1, -1).expand_as(pred_label)) 16 | 17 | res = [] 18 | for k in topk: 19 | correct_k = correct[:k].view(-1).float().sum(0, keepdim=True) 20 | res.append(correct_k.mul_(100.0 / pred.size(0))) 21 | return res[0] if return_single else res 22 | 23 | 24 | class Accuracy(nn.Module): 25 | 26 | def __init__(self, topk=(1, )): 27 | super().__init__() 28 | self.topk = topk 29 | 30 | def forward(self, pred, target): 31 | return accuracy(pred, target, self.topk) 32 | -------------------------------------------------------------------------------- /mmdet/core/bbox/samplers/pseudo_sampler.py: -------------------------------------------------------------------------------- 1 | import torch 2 | 3 | from .base_sampler import BaseSampler 4 | from .sampling_result import SamplingResult 5 | 6 | 7 | class PseudoSampler(BaseSampler): 8 | 9 | def __init__(self, **kwargs): 10 | pass 11 | 12 | def _sample_pos(self, **kwargs): 13 | raise NotImplementedError 14 | 15 | def _sample_neg(self, **kwargs): 16 | raise NotImplementedError 17 | 18 | def sample(self, assign_result, bboxes, gt_bboxes, **kwargs): 19 | pos_inds = torch.nonzero( 20 | assign_result.gt_inds > 0).squeeze(-1).unique() 21 | neg_inds = torch.nonzero( 22 | assign_result.gt_inds == 0).squeeze(-1).unique() 23 | gt_flags = bboxes.new_zeros(bboxes.shape[0], dtype=torch.uint8) 24 | sampling_result = SamplingResult(pos_inds, neg_inds, bboxes, gt_bboxes, 25 | assign_result, gt_flags) 26 | return sampling_result 27 | -------------------------------------------------------------------------------- /mmdet/core/bbox/assigners/assign_result.py: -------------------------------------------------------------------------------- 1 | import torch 2 | 3 | 4 | class AssignResult(object): 5 | 6 | def __init__(self, num_gts, gt_inds, max_overlaps, labels=None): 7 | self.num_gts = num_gts 8 | self.gt_inds = gt_inds 9 | # print('-------self.num_gts',self.num_gts) 10 | # print('-------(self.gt_inds > 0).sum()', (self.gt_inds>0).sum()) 11 | # print('-------(self.gt_inds==0).sum()', (self.gt_inds==0).sum()) 12 | 13 | self.max_overlaps = max_overlaps 14 | self.labels = labels 15 | 16 | def add_gt_(self, gt_labels): 17 | self_inds = torch.arange( 18 | 1, len(gt_labels) + 1, dtype=torch.long, device=gt_labels.device) 19 | self.gt_inds = torch.cat([self_inds, self.gt_inds]) 20 | self.max_overlaps = torch.cat( 21 | [self.max_overlaps.new_ones(self.num_gts), self.max_overlaps]) 22 | if self.labels is not None: 23 | self.labels = torch.cat([gt_labels, self.labels]) 24 | -------------------------------------------------------------------------------- /mmdet/models/__init__.py: -------------------------------------------------------------------------------- 1 | from .anchor_heads import * # noqa: F401,F403 2 | from .backbones import * # noqa: F401,F403 3 | from .bbox_heads import * # noqa: F401,F403 4 | from .builder import (build_backbone, build_detector, build_head, build_loss, 5 | build_neck, build_roi_extractor, build_shared_head) 6 | from .detectors import * # noqa: F401,F403 7 | from .losses import * # noqa: F401,F403 8 | from .mask_heads import * # noqa: F401,F403 9 | from .necks import * # noqa: F401,F403 10 | from .registry import (BACKBONES, DETECTORS, HEADS, LOSSES, NECKS, 11 | ROI_EXTRACTORS, SHARED_HEADS) 12 | from .roi_extractors import * # noqa: F401,F403 13 | from .shared_heads import * # noqa: F401,F403 14 | 15 | __all__ = [ 16 | 'BACKBONES', 'NECKS', 'ROI_EXTRACTORS', 'SHARED_HEADS', 'HEADS', 'LOSSES', 17 | 'DETECTORS', 'build_backbone', 'build_neck', 'build_roi_extractor', 18 | 'build_shared_head', 'build_head', 'build_loss', 'build_detector' 19 | ] 20 | -------------------------------------------------------------------------------- /mmdet/ops/__init__.py: -------------------------------------------------------------------------------- 1 | from .context_block import ContextBlock 2 | from .dcn import (DeformConv, DeformConvPack, DeformRoIPooling, 3 | DeformRoIPoolingPack, ModulatedDeformConv, 4 | ModulatedDeformConvPack, ModulatedDeformRoIPoolingPack, 5 | deform_conv, deform_roi_pooling, modulated_deform_conv) 6 | from .masked_conv import MaskedConv2d 7 | from .nms import nms, soft_nms 8 | from .roi_align import RoIAlign, roi_align 9 | from .roi_pool import RoIPool, roi_pool 10 | from .sigmoid_focal_loss import SigmoidFocalLoss, sigmoid_focal_loss 11 | 12 | __all__ = [ 13 | 'nms', 'soft_nms', 'RoIAlign', 'roi_align', 'RoIPool', 'roi_pool', 14 | 'DeformConv', 'DeformConvPack', 'DeformRoIPooling', 'DeformRoIPoolingPack', 15 | 'ModulatedDeformRoIPoolingPack', 'ModulatedDeformConv', 16 | 'ModulatedDeformConvPack', 'deform_conv', 'modulated_deform_conv', 17 | 'deform_roi_pooling', 'SigmoidFocalLoss', 'sigmoid_focal_loss', 18 | 'MaskedConv2d', 'ContextBlock' 19 | ] 20 | -------------------------------------------------------------------------------- /mmdet/ops/roi_align/gradcheck.py: -------------------------------------------------------------------------------- 1 | import os.path as osp 2 | import sys 3 | 4 | import numpy as np 5 | import torch 6 | from torch.autograd import gradcheck 7 | 8 | sys.path.append(osp.abspath(osp.join(__file__, '../../'))) 9 | from roi_align import RoIAlign # noqa: E402, isort:skip 10 | 11 | feat_size = 15 12 | spatial_scale = 1.0 / 8 13 | img_size = feat_size / spatial_scale 14 | num_imgs = 2 15 | num_rois = 20 16 | 17 | batch_ind = np.random.randint(num_imgs, size=(num_rois, 1)) 18 | rois = np.random.rand(num_rois, 4) * img_size * 0.5 19 | rois[:, 2:] += img_size * 0.5 20 | rois = np.hstack((batch_ind, rois)) 21 | 22 | feat = torch.randn( 23 | num_imgs, 16, feat_size, feat_size, requires_grad=True, device='cuda:0') 24 | rois = torch.from_numpy(rois).float().cuda() 25 | inputs = (feat, rois) 26 | print('Gradcheck for roi align...') 27 | test = gradcheck(RoIAlign(3, spatial_scale), inputs, atol=1e-3, eps=1e-3) 28 | print(test) 29 | test = gradcheck(RoIAlign(3, spatial_scale, 2), inputs, atol=1e-3, eps=1e-3) 30 | print(test) 31 | -------------------------------------------------------------------------------- /mmdet/core/evaluation/__init__.py: -------------------------------------------------------------------------------- 1 | from .class_names import (coco_classes, dataset_aliases, get_classes, 2 | imagenet_det_classes, imagenet_vid_classes, 3 | voc_classes) 4 | from .coco_utils import coco_eval, fast_eval_recall, results2json 5 | from .eval_hooks import (CocoDistEvalmAPHook, CocoDistEvalRecallHook, 6 | DistEvalHook, DistEvalmAPHook) 7 | from .mean_ap import average_precision, eval_map, print_map_summary 8 | from .recall import (eval_recalls, plot_iou_recall, plot_num_recall, 9 | print_recall_summary) 10 | 11 | __all__ = [ 12 | 'voc_classes', 'imagenet_det_classes', 'imagenet_vid_classes', 13 | 'coco_classes', 'dataset_aliases', 'get_classes', 'coco_eval', 14 | 'fast_eval_recall', 'results2json', 'DistEvalHook', 'DistEvalmAPHook', 15 | 'CocoDistEvalRecallHook', 'CocoDistEvalmAPHook', 'average_precision', 16 | 'eval_map', 'print_map_summary', 'eval_recalls', 'print_recall_summary', 17 | 'plot_num_recall', 'plot_iou_recall' 18 | ] 19 | -------------------------------------------------------------------------------- /mmdet/models/losses/__init__.py: -------------------------------------------------------------------------------- 1 | from .accuracy import Accuracy, accuracy 2 | from .balanced_l1_loss import BalancedL1Loss, balanced_l1_loss 3 | from .cross_entropy_loss import (CrossEntropyLoss, binary_cross_entropy, 4 | cross_entropy, mask_cross_entropy) 5 | from .focal_loss import FocalLoss, sigmoid_focal_loss 6 | from .ghm_loss import GHMC, GHMR 7 | from .iou_loss import BoundedIoULoss, IoULoss, bounded_iou_loss, iou_loss 8 | from .mse_loss import MSELoss, mse_loss 9 | from .smooth_l1_loss import SmoothL1Loss, smooth_l1_loss 10 | from .utils import reduce_loss, weight_reduce_loss, weighted_loss 11 | 12 | __all__ = [ 13 | 'accuracy', 'Accuracy', 'cross_entropy', 'binary_cross_entropy', 14 | 'mask_cross_entropy', 'CrossEntropyLoss', 'sigmoid_focal_loss', 15 | 'FocalLoss', 'smooth_l1_loss', 'SmoothL1Loss', 'balanced_l1_loss', 16 | 'BalancedL1Loss', 'mse_loss', 'MSELoss', 'iou_loss', 'bounded_iou_loss', 17 | 'IoULoss', 'BoundedIoULoss', 'GHMC', 'GHMR', 'reduce_loss', 18 | 'weight_reduce_loss', 'weighted_loss' 19 | ] 20 | -------------------------------------------------------------------------------- /mmdet/models/detectors/mask_rcnn.py: -------------------------------------------------------------------------------- 1 | from ..registry import DETECTORS 2 | from .two_stage import TwoStageDetector 3 | 4 | 5 | @DETECTORS.register_module 6 | class MaskRCNN(TwoStageDetector): 7 | 8 | def __init__(self, 9 | backbone, 10 | rpn_head, 11 | bbox_roi_extractor, 12 | bbox_head, 13 | mask_roi_extractor, 14 | mask_head, 15 | train_cfg, 16 | test_cfg, 17 | neck=None, 18 | shared_head=None, 19 | pretrained=None): 20 | super(MaskRCNN, self).__init__( 21 | backbone=backbone, 22 | neck=neck, 23 | shared_head=shared_head, 24 | rpn_head=rpn_head, 25 | bbox_roi_extractor=bbox_roi_extractor, 26 | bbox_head=bbox_head, 27 | mask_roi_extractor=mask_roi_extractor, 28 | mask_head=mask_head, 29 | train_cfg=train_cfg, 30 | test_cfg=test_cfg, 31 | pretrained=pretrained) 32 | -------------------------------------------------------------------------------- /mmdet_extra/ps_roi_align_ori/vision.h: -------------------------------------------------------------------------------- 1 | // Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved. 2 | #pragma once 3 | #include 4 | 5 | //// 6 | #include 7 | 8 | 9 | 10 | 11 | 12 | 13 | 14 | int PSROIAlignForwardLaucher(at::Tensor bottom_data, 15 | at::Tensor bottom_rois, 16 | at::Tensor top_data, 17 | at::Tensor argmax_data, 18 | float spatial_scale, 19 | int group_size, 20 | int sampling_ratio, 21 | cudaStream_t stream); 22 | 23 | 24 | int PSROIAlignBackwardLaucher(at::Tensor top_diff, 25 | at::Tensor argmax_data, 26 | at::Tensor bottom_rois, 27 | at::Tensor bottom_diff, 28 | float spatial_scale, 29 | int group_size, 30 | int sampling_ratio, 31 | cudaStream_t stream); -------------------------------------------------------------------------------- /mmdet/models/builder.py: -------------------------------------------------------------------------------- 1 | from torch import nn 2 | 3 | from mmdet.utils import build_from_cfg 4 | from .registry import (BACKBONES, DETECTORS, HEADS, LOSSES, NECKS, 5 | ROI_EXTRACTORS, SHARED_HEADS) 6 | 7 | 8 | def build(cfg, registry, default_args=None): 9 | 10 | if isinstance(cfg, list): 11 | modules = [ 12 | build_from_cfg(cfg_, registry, default_args) for cfg_ in cfg 13 | ] 14 | return nn.Sequential(*modules) 15 | else: 16 | return build_from_cfg(cfg, registry, default_args) 17 | 18 | 19 | def build_backbone(cfg): 20 | return build(cfg, BACKBONES) 21 | 22 | 23 | def build_neck(cfg): 24 | return build(cfg, NECKS) 25 | 26 | 27 | def build_roi_extractor(cfg): 28 | return build(cfg, ROI_EXTRACTORS) 29 | 30 | 31 | def build_shared_head(cfg): 32 | return build(cfg, SHARED_HEADS) 33 | 34 | 35 | def build_head(cfg): 36 | return build(cfg, HEADS) 37 | 38 | 39 | def build_loss(cfg): 40 | return build(cfg, LOSSES) 41 | 42 | 43 | def build_detector(cfg, train_cfg=None, test_cfg=None): 44 | return build(cfg, DETECTORS, dict(train_cfg=train_cfg, test_cfg=test_cfg)) 45 | -------------------------------------------------------------------------------- /tools/publish_model.py: -------------------------------------------------------------------------------- 1 | import argparse 2 | import subprocess 3 | 4 | import torch 5 | 6 | 7 | def parse_args(): 8 | parser = argparse.ArgumentParser( 9 | description='Process a checkpoint to be published') 10 | parser.add_argument('in_file', help='input checkpoint filename') 11 | parser.add_argument('out_file', help='output checkpoint filename') 12 | args = parser.parse_args() 13 | return args 14 | 15 | 16 | def process_checkpoint(in_file, out_file): 17 | checkpoint = torch.load(in_file, map_location='cpu') 18 | # remove optimizer for smaller file size 19 | if 'optimizer' in checkpoint: 20 | del checkpoint['optimizer'] 21 | # if it is necessary to remove some sensitive data in checkpoint['meta'], 22 | # add the code here. 23 | torch.save(checkpoint, out_file) 24 | sha = subprocess.check_output(['sha256sum', out_file]).decode() 25 | final_file = out_file.rstrip('.pth') + '-{}.pth'.format(sha[:8]) 26 | subprocess.Popen(['mv', out_file, final_file]) 27 | 28 | 29 | def main(): 30 | args = parse_args() 31 | process_checkpoint(args.in_file, args.out_file) 32 | 33 | 34 | if __name__ == '__main__': 35 | main() 36 | -------------------------------------------------------------------------------- /mmdet/core/bbox/__init__.py: -------------------------------------------------------------------------------- 1 | from .assigners import AssignResult, BaseAssigner, MaxIoUAssigner 2 | from .bbox_target import bbox_target 3 | from .geometry import bbox_overlaps 4 | from .samplers import (BaseSampler, CombinedSampler, 5 | InstanceBalancedPosSampler, IoUBalancedNegSampler, 6 | PseudoSampler, RandomSampler, SamplingResult) 7 | from .transforms import (bbox2delta, bbox2result, bbox2roi, bbox_flip, 8 | bbox_mapping, bbox_mapping_back, delta2bbox, 9 | distance2bbox, roi2bbox) 10 | 11 | from .assign_sampling import ( # isort:skip, avoid recursive imports 12 | assign_and_sample, build_assigner, build_sampler) 13 | 14 | __all__ = [ 15 | 'bbox_overlaps', 'BaseAssigner', 'MaxIoUAssigner', 'AssignResult', 16 | 'BaseSampler', 'PseudoSampler', 'RandomSampler', 17 | 'InstanceBalancedPosSampler', 'IoUBalancedNegSampler', 'CombinedSampler', 18 | 'SamplingResult', 'build_assigner', 'build_sampler', 'assign_and_sample', 19 | 'bbox2delta', 'delta2bbox', 'bbox_flip', 'bbox_mapping', 20 | 'bbox_mapping_back', 'bbox2roi', 'roi2bbox', 'bbox2result', 21 | 'distance2bbox', 'bbox_target' 22 | ] 23 | -------------------------------------------------------------------------------- /mmdet/datasets/pipelines/compose.py: -------------------------------------------------------------------------------- 1 | import collections 2 | 3 | from mmdet.utils import build_from_cfg 4 | from ..registry import PIPELINES 5 | 6 | 7 | @PIPELINES.register_module 8 | class Compose(object): 9 | 10 | def __init__(self, transforms): 11 | assert isinstance(transforms, collections.abc.Sequence) 12 | self.transforms = [] 13 | for transform in transforms: 14 | if isinstance(transform, dict): 15 | transform = build_from_cfg(transform, PIPELINES) 16 | self.transforms.append(transform) 17 | elif callable(transform): 18 | self.transforms.append(transform) 19 | else: 20 | raise TypeError('transform must be callable or a dict') 21 | 22 | def __call__(self, data): 23 | for t in self.transforms: 24 | data = t(data) 25 | if data is None: 26 | return None 27 | return data 28 | 29 | def __repr__(self): 30 | format_string = self.__class__.__name__ + '(' 31 | for t in self.transforms: 32 | format_string += '\n' 33 | format_string += ' {0}'.format(t) 34 | format_string += '\n)' 35 | return format_string 36 | -------------------------------------------------------------------------------- /mmdet/core/utils/misc.py: -------------------------------------------------------------------------------- 1 | from functools import partial 2 | 3 | import mmcv 4 | import numpy as np 5 | from six.moves import map, zip 6 | 7 | 8 | def tensor2imgs(tensor, mean=(0, 0, 0), std=(1, 1, 1), to_rgb=True): 9 | num_imgs = tensor.size(0) 10 | mean = np.array(mean, dtype=np.float32) 11 | std = np.array(std, dtype=np.float32) 12 | imgs = [] 13 | for img_id in range(num_imgs): 14 | img = tensor[img_id, ...].cpu().numpy().transpose(1, 2, 0) 15 | img = mmcv.imdenormalize( 16 | img, mean, std, to_bgr=to_rgb).astype(np.uint8) 17 | imgs.append(np.ascontiguousarray(img)) 18 | return imgs 19 | 20 | 21 | def multi_apply(func, *args, **kwargs): 22 | pfunc = partial(func, **kwargs) if kwargs else func 23 | map_results = map(pfunc, *args) 24 | return tuple(map(list, zip(*map_results))) 25 | 26 | 27 | def unmap(data, count, inds, fill=0): 28 | """ Unmap a subset of item (data) back to the original set of items (of 29 | size count) """ 30 | if data.dim() == 1: 31 | ret = data.new_full((count, ), fill) 32 | ret[inds] = data 33 | else: 34 | new_size = (count, ) + data.size()[1:] 35 | ret = data.new_full(new_size, fill) 36 | ret[inds, :] = data 37 | return ret 38 | -------------------------------------------------------------------------------- /mmdet/core/mask/utils.py: -------------------------------------------------------------------------------- 1 | import mmcv 2 | 3 | 4 | def split_combined_polys(polys, poly_lens, polys_per_mask): 5 | """Split the combined 1-D polys into masks. 6 | 7 | A mask is represented as a list of polys, and a poly is represented as 8 | a 1-D array. In dataset, all masks are concatenated into a single 1-D 9 | tensor. Here we need to split the tensor into original representations. 10 | 11 | Args: 12 | polys (list): a list (length = image num) of 1-D tensors 13 | poly_lens (list): a list (length = image num) of poly length 14 | polys_per_mask (list): a list (length = image num) of poly number 15 | of each mask 16 | 17 | Returns: 18 | list: a list (length = image num) of list (length = mask num) of 19 | list (length = poly num) of numpy array 20 | """ 21 | mask_polys_list = [] 22 | for img_id in range(len(polys)): 23 | polys_single = polys[img_id] 24 | polys_lens_single = poly_lens[img_id].tolist() 25 | polys_per_mask_single = polys_per_mask[img_id].tolist() 26 | 27 | split_polys = mmcv.slice_list(polys_single, polys_lens_single) 28 | mask_polys = mmcv.slice_list(split_polys, polys_per_mask_single) 29 | mask_polys_list.append(mask_polys) 30 | return mask_polys_list 31 | -------------------------------------------------------------------------------- /mmdet/core/bbox/assign_sampling.py: -------------------------------------------------------------------------------- 1 | import mmcv 2 | 3 | from . import assigners, samplers 4 | 5 | 6 | def build_assigner(cfg, **kwargs): 7 | 8 | if isinstance(cfg, assigners.BaseAssigner): 9 | return cfg 10 | elif isinstance(cfg, dict): 11 | return mmcv.runner.obj_from_dict(cfg, assigners, default_args=kwargs) 12 | else: 13 | raise TypeError('Invalid type {} for building a sampler'.format( 14 | type(cfg))) 15 | 16 | 17 | def build_sampler(cfg, **kwargs): 18 | if isinstance(cfg, samplers.BaseSampler): 19 | return cfg 20 | elif isinstance(cfg, dict): 21 | return mmcv.runner.obj_from_dict(cfg, samplers, default_args=kwargs) 22 | else: 23 | raise TypeError('Invalid type {} for building a sampler'.format( 24 | type(cfg))) 25 | 26 | def assign_and_sample(bboxes, gt_bboxes, gt_bboxes_ignore, gt_labels, cfg): 27 | bbox_assigner = build_assigner(cfg.assigner) 28 | bbox_sampler = build_sampler(cfg.sampler) 29 | assign_result = bbox_assigner.assign(bboxes, gt_bboxes, gt_bboxes_ignore, 30 | gt_labels) 31 | sampling_result = bbox_sampler.sample(assign_result, bboxes, gt_bboxes, 32 | gt_labels) 33 | 34 | return assign_result, sampling_result 35 | -------------------------------------------------------------------------------- /mmdet/models/mask_heads/htc_mask_head.py: -------------------------------------------------------------------------------- 1 | from ..registry import HEADS 2 | from ..utils import ConvModule 3 | from .fcn_mask_head import FCNMaskHead 4 | 5 | 6 | @HEADS.register_module 7 | class HTCMaskHead(FCNMaskHead): 8 | 9 | def __init__(self, *args, **kwargs): 10 | super(HTCMaskHead, self).__init__(*args, **kwargs) 11 | self.conv_res = ConvModule( 12 | self.conv_out_channels, 13 | self.conv_out_channels, 14 | 1, 15 | conv_cfg=self.conv_cfg, 16 | norm_cfg=self.norm_cfg) 17 | 18 | def init_weights(self): 19 | super(HTCMaskHead, self).init_weights() 20 | self.conv_res.init_weights() 21 | 22 | def forward(self, x, res_feat=None, return_logits=True, return_feat=True): 23 | if res_feat is not None: 24 | res_feat = self.conv_res(res_feat) 25 | x = x + res_feat 26 | for conv in self.convs: 27 | x = conv(x) 28 | res_feat = x 29 | outs = [] 30 | if return_logits: 31 | x = self.upsample(x) 32 | if self.upsample_method == 'deconv': 33 | x = self.relu(x) 34 | mask_pred = self.conv_logits(x) 35 | outs.append(mask_pred) 36 | if return_feat: 37 | outs.append(res_feat) 38 | return outs if len(outs) > 1 else outs[0] 39 | -------------------------------------------------------------------------------- /mmdet/core/anchor/point_generator.py: -------------------------------------------------------------------------------- 1 | import torch 2 | 3 | 4 | class PointGenerator(object): 5 | 6 | def _meshgrid(self, x, y, row_major=True): 7 | xx = x.repeat(len(y)) 8 | yy = y.view(-1, 1).repeat(1, len(x)).view(-1) 9 | if row_major: 10 | return xx, yy 11 | else: 12 | return yy, xx 13 | 14 | def grid_points(self, featmap_size, stride=16, device='cuda'): 15 | feat_h, feat_w = featmap_size 16 | shift_x = torch.arange(0., feat_w, device=device) * stride 17 | shift_y = torch.arange(0., feat_h, device=device) * stride 18 | shift_xx, shift_yy = self._meshgrid(shift_x, shift_y) 19 | stride = shift_x.new_full((shift_xx.shape[0], ), stride) 20 | shifts = torch.stack([shift_xx, shift_yy, stride], dim=-1) 21 | all_points = shifts.to(device) 22 | return all_points 23 | 24 | def valid_flags(self, featmap_size, valid_size, device='cuda'): 25 | feat_h, feat_w = featmap_size 26 | valid_h, valid_w = valid_size 27 | assert valid_h <= feat_h and valid_w <= feat_w 28 | valid_x = torch.zeros(feat_w, dtype=torch.uint8, device=device) 29 | valid_y = torch.zeros(feat_h, dtype=torch.uint8, device=device) 30 | valid_x[:valid_w] = 1 31 | valid_y[:valid_h] = 1 32 | valid_xx, valid_yy = self._meshgrid(valid_x, valid_y) 33 | valid = valid_xx & valid_yy 34 | return valid 35 | -------------------------------------------------------------------------------- /mmdet/datasets/pipelines/test_aug.py: -------------------------------------------------------------------------------- 1 | import mmcv 2 | 3 | from ..registry import PIPELINES 4 | from .compose import Compose 5 | 6 | 7 | @PIPELINES.register_module 8 | class MultiScaleFlipAug(object): 9 | 10 | def __init__(self, transforms, img_scale, flip=False): 11 | self.transforms = Compose(transforms) 12 | self.img_scale = img_scale if isinstance(img_scale, 13 | list) else [img_scale] 14 | assert mmcv.is_list_of(self.img_scale, tuple) 15 | self.flip = flip 16 | 17 | def __call__(self, results): 18 | aug_data = [] 19 | flip_aug = [False, True] if self.flip else [False] 20 | for scale in self.img_scale: 21 | for flip in flip_aug: 22 | _results = results.copy() 23 | _results['scale'] = scale 24 | _results['flip'] = flip 25 | data = self.transforms(_results) 26 | aug_data.append(data) 27 | # list of dict to dict of list 28 | aug_data_dict = {key: [] for key in aug_data[0]} 29 | for data in aug_data: 30 | for key, val in data.items(): 31 | aug_data_dict[key].append(val) 32 | return aug_data_dict 33 | 34 | def __repr__(self): 35 | repr_str = self.__class__.__name__ 36 | repr_str += '(transforms={}, img_scale={}, flip={})'.format( 37 | self.transforms, self.img_scale, self.flip) 38 | return repr_str 39 | -------------------------------------------------------------------------------- /mmdet/datasets/wider_face.py: -------------------------------------------------------------------------------- 1 | import os.path as osp 2 | import xml.etree.ElementTree as ET 3 | 4 | import mmcv 5 | 6 | from .registry import DATASETS 7 | from .xml_style import XMLDataset 8 | 9 | 10 | @DATASETS.register_module 11 | class WIDERFaceDataset(XMLDataset): 12 | """ 13 | Reader for the WIDER Face dataset in PASCAL VOC format. 14 | Conversion scripts can be found in 15 | https://github.com/sovrasov/wider-face-pascal-voc-annotations 16 | """ 17 | CLASSES = ('face', ) 18 | 19 | def __init__(self, **kwargs): 20 | super(WIDERFaceDataset, self).__init__(**kwargs) 21 | 22 | def load_annotations(self, ann_file): 23 | img_infos = [] 24 | img_ids = mmcv.list_from_file(ann_file) 25 | for img_id in img_ids: 26 | filename = '{}.jpg'.format(img_id) 27 | xml_path = osp.join(self.img_prefix, 'Annotations', 28 | '{}.xml'.format(img_id)) 29 | tree = ET.parse(xml_path) 30 | root = tree.getroot() 31 | size = root.find('size') 32 | width = int(size.find('width').text) 33 | height = int(size.find('height').text) 34 | folder = root.find('folder').text 35 | img_infos.append( 36 | dict( 37 | id=img_id, 38 | filename=osp.join(folder, filename), 39 | width=width, 40 | height=height)) 41 | 42 | return img_infos 43 | -------------------------------------------------------------------------------- /mmdet/models/losses/smooth_l1_loss.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import torch.nn as nn 3 | 4 | from ..registry import LOSSES 5 | from .utils import weighted_loss 6 | 7 | 8 | @weighted_loss 9 | def smooth_l1_loss(pred, target, beta=1.0): 10 | assert beta > 0 11 | assert pred.size() == target.size() and target.numel() > 0 12 | diff = torch.abs(pred - target) 13 | loss = torch.where(diff < beta, 0.5 * diff * diff / beta, 14 | diff - 0.5 * beta) 15 | return loss 16 | 17 | 18 | @LOSSES.register_module 19 | class SmoothL1Loss(nn.Module): 20 | 21 | def __init__(self, beta=1.0, reduction='mean', loss_weight=1.0): 22 | super(SmoothL1Loss, self).__init__() 23 | self.beta = beta 24 | self.reduction = reduction 25 | self.loss_weight = loss_weight 26 | 27 | def forward(self, 28 | pred, 29 | target, 30 | weight=None, 31 | avg_factor=None, 32 | reduction_override=None, 33 | **kwargs): 34 | assert reduction_override in (None, 'none', 'mean', 'sum') 35 | reduction = ( 36 | reduction_override if reduction_override else self.reduction) 37 | loss_bbox = self.loss_weight * smooth_l1_loss( 38 | pred, 39 | target, 40 | weight, 41 | beta=self.beta, 42 | reduction=reduction, 43 | avg_factor=avg_factor, 44 | **kwargs) 45 | return loss_bbox 46 | -------------------------------------------------------------------------------- /tools/upgrade_model_version.py: -------------------------------------------------------------------------------- 1 | import argparse 2 | import re 3 | from collections import OrderedDict 4 | 5 | import torch 6 | 7 | 8 | def convert(in_file, out_file): 9 | """Convert keys in checkpoints. 10 | 11 | There can be some breaking changes during the development of mmdetection, 12 | and this tool is used for upgrading checkpoints trained with old versions 13 | to the latest one. 14 | """ 15 | checkpoint = torch.load(in_file) 16 | in_state_dict = checkpoint.pop('state_dict') 17 | out_state_dict = OrderedDict() 18 | for key, val in in_state_dict.items(): 19 | # Use ConvModule instead of nn.Conv2d in RetinaNet 20 | # cls_convs.0.weight -> cls_convs.0.conv.weight 21 | m = re.search(r'(cls_convs|reg_convs).\d.(weight|bias)', key) 22 | if m is not None: 23 | param = m.groups()[1] 24 | new_key = key.replace(param, 'conv.{}'.format(param)) 25 | out_state_dict[new_key] = val 26 | continue 27 | 28 | out_state_dict[key] = val 29 | checkpoint['state_dict'] = out_state_dict 30 | torch.save(checkpoint, out_file) 31 | 32 | 33 | def main(): 34 | parser = argparse.ArgumentParser(description='Upgrade model version') 35 | parser.add_argument('in_file', help='input checkpoint file') 36 | parser.add_argument('out_file', help='output checkpoint file') 37 | args = parser.parse_args() 38 | convert(args.in_file, args.out_file) 39 | 40 | 41 | if __name__ == '__main__': 42 | main() 43 | -------------------------------------------------------------------------------- /mmdet/models/utils/conv_ws.py: -------------------------------------------------------------------------------- 1 | import torch.nn as nn 2 | import torch.nn.functional as F 3 | 4 | 5 | def conv_ws_2d(input, 6 | weight, 7 | bias=None, 8 | stride=1, 9 | padding=0, 10 | dilation=1, 11 | groups=1, 12 | eps=1e-5): 13 | c_in = weight.size(0) 14 | weight_flat = weight.view(c_in, -1) 15 | mean = weight_flat.mean(dim=1, keepdim=True).view(c_in, 1, 1, 1) 16 | std = weight_flat.std(dim=1, keepdim=True).view(c_in, 1, 1, 1) 17 | weight = (weight - mean) / (std + eps) 18 | return F.conv2d(input, weight, bias, stride, padding, dilation, groups) 19 | 20 | 21 | class ConvWS2d(nn.Conv2d): 22 | 23 | def __init__(self, 24 | in_channels, 25 | out_channels, 26 | kernel_size, 27 | stride=1, 28 | padding=0, 29 | dilation=1, 30 | groups=1, 31 | bias=True, 32 | eps=1e-5): 33 | super(ConvWS2d, self).__init__( 34 | in_channels, 35 | out_channels, 36 | kernel_size, 37 | stride=stride, 38 | padding=padding, 39 | dilation=dilation, 40 | groups=groups, 41 | bias=bias) 42 | self.eps = eps 43 | 44 | def forward(self, x): 45 | return conv_ws_2d(x, self.weight, self.bias, self.stride, self.padding, 46 | self.dilation, self.groups, self.eps) 47 | -------------------------------------------------------------------------------- /mmdet/datasets/builder.py: -------------------------------------------------------------------------------- 1 | import copy 2 | 3 | from mmdet.utils import build_from_cfg 4 | from .dataset_wrappers import ConcatDataset, RepeatDataset 5 | from .registry import DATASETS 6 | 7 | 8 | def _concat_dataset(cfg, default_args=None): 9 | ann_files = cfg['ann_file'] 10 | img_prefixes = cfg.get('img_prefix', None) 11 | seg_prefixes = cfg.get('seg_prefixes', None) 12 | proposal_files = cfg.get('proposal_file', None) 13 | 14 | datasets = [] 15 | num_dset = len(ann_files) 16 | for i in range(num_dset): 17 | data_cfg = copy.deepcopy(cfg) 18 | data_cfg['ann_file'] = ann_files[i] 19 | if isinstance(img_prefixes, (list, tuple)): 20 | data_cfg['img_prefix'] = img_prefixes[i] 21 | if isinstance(seg_prefixes, (list, tuple)): 22 | data_cfg['seg_prefix'] = seg_prefixes[i] 23 | if isinstance(proposal_files, (list, tuple)): 24 | data_cfg['proposal_file'] = proposal_files[i] 25 | datasets.append(build_dataset(data_cfg, default_args)) 26 | 27 | return ConcatDataset(datasets) 28 | 29 | 30 | def build_dataset(cfg, default_args=None): 31 | if isinstance(cfg, (list, tuple)): 32 | dataset = ConcatDataset([build_dataset(c, default_args) for c in cfg]) 33 | elif cfg['type'] == 'RepeatDataset': 34 | dataset = RepeatDataset( 35 | build_dataset(cfg['dataset'], default_args), cfg['times']) 36 | elif isinstance(cfg['ann_file'], (list, tuple)): 37 | dataset = _concat_dataset(cfg, default_args) 38 | else: 39 | dataset = build_from_cfg(cfg, DATASETS, default_args) 40 | 41 | return dataset 42 | -------------------------------------------------------------------------------- /mmdet/core/mask/mask_target.py: -------------------------------------------------------------------------------- 1 | import mmcv 2 | import numpy as np 3 | import torch 4 | from torch.nn.modules.utils import _pair 5 | 6 | 7 | def mask_target(pos_proposals_list, pos_assigned_gt_inds_list, gt_masks_list, 8 | cfg): 9 | cfg_list = [cfg for _ in range(len(pos_proposals_list))] 10 | mask_targets = map(mask_target_single, pos_proposals_list, 11 | pos_assigned_gt_inds_list, gt_masks_list, cfg_list) 12 | mask_targets = torch.cat(list(mask_targets)) 13 | return mask_targets 14 | 15 | 16 | def mask_target_single(pos_proposals, pos_assigned_gt_inds, gt_masks, cfg): 17 | mask_size = _pair(cfg.mask_size) 18 | num_pos = pos_proposals.size(0) 19 | mask_targets = [] 20 | if num_pos > 0: 21 | proposals_np = pos_proposals.cpu().numpy() 22 | pos_assigned_gt_inds = pos_assigned_gt_inds.cpu().numpy() 23 | for i in range(num_pos): 24 | gt_mask = gt_masks[pos_assigned_gt_inds[i]] 25 | bbox = proposals_np[i, :].astype(np.int32) 26 | x1, y1, x2, y2 = bbox 27 | w = np.maximum(x2 - x1 + 1, 1) 28 | h = np.maximum(y2 - y1 + 1, 1) 29 | # mask is uint8 both before and after resizing 30 | # mask_size (h, w) to (w, h) 31 | target = mmcv.imresize(gt_mask[y1:y1 + h, x1:x1 + w], 32 | mask_size[::-1]) 33 | mask_targets.append(target) 34 | mask_targets = torch.from_numpy(np.stack(mask_targets)).float().to( 35 | pos_proposals.device) 36 | else: 37 | mask_targets = pos_proposals.new_zeros((0, ) + mask_size) 38 | return mask_targets 39 | -------------------------------------------------------------------------------- /mmdet/models/utils/weight_init.py: -------------------------------------------------------------------------------- 1 | import numpy as np 2 | import torch.nn as nn 3 | 4 | 5 | def xavier_init(module, gain=1, bias=0, distribution='normal'): 6 | assert distribution in ['uniform', 'normal'] 7 | if distribution == 'uniform': 8 | nn.init.xavier_uniform_(module.weight, gain=gain) 9 | else: 10 | nn.init.xavier_normal_(module.weight, gain=gain) 11 | if hasattr(module, 'bias'): 12 | nn.init.constant_(module.bias, bias) 13 | 14 | 15 | def normal_init(module, mean=0, std=1, bias=0): 16 | nn.init.normal_(module.weight, mean, std) 17 | if hasattr(module, 'bias'): 18 | nn.init.constant_(module.bias, bias) 19 | 20 | 21 | def uniform_init(module, a=0, b=1, bias=0): 22 | nn.init.uniform_(module.weight, a, b) 23 | if hasattr(module, 'bias'): 24 | nn.init.constant_(module.bias, bias) 25 | 26 | 27 | def kaiming_init(module, 28 | mode='fan_out', 29 | nonlinearity='relu', 30 | bias=0, 31 | distribution='normal'): 32 | assert distribution in ['uniform', 'normal'] 33 | if distribution == 'uniform': 34 | nn.init.kaiming_uniform_( 35 | module.weight, mode=mode, nonlinearity=nonlinearity) 36 | else: 37 | nn.init.kaiming_normal_( 38 | module.weight, mode=mode, nonlinearity=nonlinearity) 39 | if hasattr(module, 'bias'): 40 | nn.init.constant_(module.bias, bias) 41 | 42 | 43 | def bias_init_with_prob(prior_prob): 44 | """ initialize conv/fc bias value according to giving probablity""" 45 | bias_init = float(-np.log((1 - prior_prob) / prior_prob)) 46 | return bias_init 47 | -------------------------------------------------------------------------------- /mmdet/datasets/loader/build_loader.py: -------------------------------------------------------------------------------- 1 | import platform 2 | from functools import partial 3 | 4 | from mmcv.parallel import collate 5 | from mmcv.runner import get_dist_info 6 | from torch.utils.data import DataLoader 7 | 8 | from .sampler import DistributedGroupSampler, DistributedSampler, GroupSampler 9 | 10 | if platform.system() != 'Windows': 11 | # https://github.com/pytorch/pytorch/issues/973 12 | import resource 13 | rlimit = resource.getrlimit(resource.RLIMIT_NOFILE) 14 | resource.setrlimit(resource.RLIMIT_NOFILE, (4096, rlimit[1])) 15 | 16 | 17 | def build_dataloader(dataset, 18 | imgs_per_gpu, 19 | workers_per_gpu, 20 | num_gpus=1, 21 | dist=True, 22 | **kwargs): 23 | shuffle = kwargs.get('shuffle', True) 24 | if dist: 25 | rank, world_size = get_dist_info() 26 | if shuffle: 27 | sampler = DistributedGroupSampler(dataset, imgs_per_gpu, 28 | world_size, rank) 29 | else: 30 | sampler = DistributedSampler( 31 | dataset, world_size, rank, shuffle=False) 32 | batch_size = imgs_per_gpu 33 | num_workers = workers_per_gpu 34 | else: 35 | sampler = GroupSampler(dataset, imgs_per_gpu) if shuffle else None 36 | batch_size = num_gpus * imgs_per_gpu 37 | num_workers = num_gpus * workers_per_gpu 38 | 39 | 40 | data_loader = DataLoader( 41 | dataset, 42 | batch_size=batch_size, 43 | sampler=sampler, 44 | num_workers=num_workers, 45 | collate_fn=partial(collate, samples_per_gpu=imgs_per_gpu), 46 | pin_memory=False, 47 | **kwargs) 48 | 49 | 50 | 51 | 52 | 53 | return data_loader 54 | -------------------------------------------------------------------------------- /tools/get_flops.py: -------------------------------------------------------------------------------- 1 | import argparse 2 | 3 | from mmcv import Config 4 | 5 | from mmdet.models import build_detector 6 | 7 | import os 8 | import sys 9 | sys.path.insert(0, os.path.abspath(os.path.dirname(__file__)) + "/../") 10 | from mmdet_extra import get_model_complexity_info 11 | 12 | def parse_args(): 13 | parser = argparse.ArgumentParser(description='Train a detector') 14 | parser.add_argument('config', help='train config file path') 15 | parser.add_argument( 16 | '--shape', 17 | type=int, 18 | nargs='+', 19 | help='input image size') 20 | args = parser.parse_args() 21 | return args 22 | 23 | 24 | def main(): 25 | 26 | args = parse_args() 27 | 28 | if len(args.shape) == 1: 29 | input_shape = (3, args.shape[0], args.shape[0]) 30 | elif len(args.shape) == 2: 31 | input_shape = (3, ) + tuple(args.shape) 32 | else: 33 | raise ValueError('invalid input shape') 34 | 35 | cfg = Config.fromfile(args.config) 36 | 37 | 38 | if hasattr(cfg.model, 'norm_cfg') and cfg.model.norm_cfg['type'] == 'SyncBN': 39 | cfg.model.norm_cfg['type'] = 'BN' 40 | 41 | model = build_detector( 42 | cfg.model, train_cfg=cfg.train_cfg, test_cfg=cfg.test_cfg).cuda() 43 | model.eval() 44 | 45 | if hasattr(model, 'forward_dummy'): 46 | model.forward = model.forward_dummy 47 | else: 48 | raise NotImplementedError( 49 | 'FLOPs counter is currently not currently supported with {}'. 50 | format(model.__class__.__name__)) 51 | 52 | flops, params = get_model_complexity_info(model, input_shape) 53 | split_line = '=' * 30 54 | print('{0}\nInput shape: {1}\nFlops: {2}\nParams: {3}\n{0}'.format( 55 | split_line, input_shape, flops, params)) 56 | 57 | 58 | if __name__ == '__main__': 59 | main() 60 | -------------------------------------------------------------------------------- /mmdet/core/evaluation/bbox_overlaps.py: -------------------------------------------------------------------------------- 1 | import numpy as np 2 | 3 | 4 | def bbox_overlaps(bboxes1, bboxes2, mode='iou'): 5 | """Calculate the ious between each bbox of bboxes1 and bboxes2. 6 | 7 | Args: 8 | bboxes1(ndarray): shape (n, 4) 9 | bboxes2(ndarray): shape (k, 4) 10 | mode(str): iou (intersection over union) or iof (intersection 11 | over foreground) 12 | 13 | Returns: 14 | ious(ndarray): shape (n, k) 15 | """ 16 | 17 | assert mode in ['iou', 'iof'] 18 | 19 | bboxes1 = bboxes1.astype(np.float32) 20 | bboxes2 = bboxes2.astype(np.float32) 21 | rows = bboxes1.shape[0] 22 | cols = bboxes2.shape[0] 23 | ious = np.zeros((rows, cols), dtype=np.float32) 24 | if rows * cols == 0: 25 | return ious 26 | exchange = False 27 | if bboxes1.shape[0] > bboxes2.shape[0]: 28 | bboxes1, bboxes2 = bboxes2, bboxes1 29 | ious = np.zeros((cols, rows), dtype=np.float32) 30 | exchange = True 31 | area1 = (bboxes1[:, 2] - bboxes1[:, 0] + 1) * ( 32 | bboxes1[:, 3] - bboxes1[:, 1] + 1) 33 | area2 = (bboxes2[:, 2] - bboxes2[:, 0] + 1) * ( 34 | bboxes2[:, 3] - bboxes2[:, 1] + 1) 35 | for i in range(bboxes1.shape[0]): 36 | x_start = np.maximum(bboxes1[i, 0], bboxes2[:, 0]) 37 | y_start = np.maximum(bboxes1[i, 1], bboxes2[:, 1]) 38 | x_end = np.minimum(bboxes1[i, 2], bboxes2[:, 2]) 39 | y_end = np.minimum(bboxes1[i, 3], bboxes2[:, 3]) 40 | overlap = np.maximum(x_end - x_start + 1, 0) * np.maximum( 41 | y_end - y_start + 1, 0) 42 | if mode == 'iou': 43 | union = area1[i] + area2 - overlap 44 | else: 45 | union = area1[i] if not exchange else area2 46 | ious[i, :] = overlap / union 47 | if exchange: 48 | ious = ious.T 49 | return ious 50 | -------------------------------------------------------------------------------- /mmdet/datasets/dataset_wrappers.py: -------------------------------------------------------------------------------- 1 | import numpy as np 2 | from torch.utils.data.dataset import ConcatDataset as _ConcatDataset 3 | 4 | from .registry import DATASETS 5 | 6 | 7 | @DATASETS.register_module 8 | class ConcatDataset(_ConcatDataset): 9 | """A wrapper of concatenated dataset. 10 | 11 | Same as :obj:`torch.utils.data.dataset.ConcatDataset`, but 12 | concat the group flag for image aspect ratio. 13 | 14 | Args: 15 | datasets (list[:obj:`Dataset`]): A list of datasets. 16 | """ 17 | 18 | def __init__(self, datasets): 19 | super(ConcatDataset, self).__init__(datasets) 20 | self.CLASSES = datasets[0].CLASSES 21 | if hasattr(datasets[0], 'flag'): 22 | flags = [] 23 | for i in range(0, len(datasets)): 24 | flags.append(datasets[i].flag) 25 | self.flag = np.concatenate(flags) 26 | 27 | 28 | @DATASETS.register_module 29 | class RepeatDataset(object): 30 | """A wrapper of repeated dataset. 31 | 32 | The length of repeated dataset will be `times` larger than the original 33 | dataset. This is useful when the data loading time is long but the dataset 34 | is small. Using RepeatDataset can reduce the data loading time between 35 | epochs. 36 | 37 | Args: 38 | dataset (:obj:`Dataset`): The dataset to be repeated. 39 | times (int): Repeat times. 40 | """ 41 | 42 | def __init__(self, dataset, times): 43 | self.dataset = dataset 44 | self.times = times 45 | self.CLASSES = dataset.CLASSES 46 | if hasattr(self.dataset, 'flag'): 47 | self.flag = np.tile(self.dataset.flag, times) 48 | 49 | self._ori_len = len(self.dataset) 50 | 51 | def __getitem__(self, idx): 52 | return self.dataset[idx % self._ori_len] 53 | 54 | def __len__(self): 55 | return self.times * self._ori_len 56 | -------------------------------------------------------------------------------- /mmdet/ops/sigmoid_focal_loss/sigmoid_focal_loss.py: -------------------------------------------------------------------------------- 1 | import torch.nn as nn 2 | from torch.autograd import Function 3 | from torch.autograd.function import once_differentiable 4 | 5 | from . import sigmoid_focal_loss_cuda 6 | 7 | 8 | class SigmoidFocalLossFunction(Function): 9 | 10 | @staticmethod 11 | def forward(ctx, input, target, gamma=2.0, alpha=0.25): 12 | ctx.save_for_backward(input, target) 13 | num_classes = input.shape[1] 14 | ctx.num_classes = num_classes 15 | ctx.gamma = gamma 16 | ctx.alpha = alpha 17 | 18 | loss = sigmoid_focal_loss_cuda.forward(input, target, num_classes, 19 | gamma, alpha) 20 | return loss 21 | 22 | @staticmethod 23 | @once_differentiable 24 | def backward(ctx, d_loss): 25 | input, target = ctx.saved_tensors 26 | num_classes = ctx.num_classes 27 | gamma = ctx.gamma 28 | alpha = ctx.alpha 29 | d_loss = d_loss.contiguous() 30 | d_input = sigmoid_focal_loss_cuda.backward(input, target, d_loss, 31 | num_classes, gamma, alpha) 32 | return d_input, None, None, None, None 33 | 34 | 35 | sigmoid_focal_loss = SigmoidFocalLossFunction.apply 36 | 37 | 38 | # TODO: remove this module 39 | class SigmoidFocalLoss(nn.Module): 40 | 41 | def __init__(self, gamma, alpha): 42 | super(SigmoidFocalLoss, self).__init__() 43 | self.gamma = gamma 44 | self.alpha = alpha 45 | 46 | def forward(self, logits, targets): 47 | assert logits.is_cuda 48 | loss = sigmoid_focal_loss(logits, targets, self.gamma, self.alpha) 49 | return loss.sum() 50 | 51 | def __repr__(self): 52 | tmpstr = self.__class__.__name__ + '(gamma={}, alpha={})'.format( 53 | self.gamma, self.alpha) 54 | return tmpstr 55 | -------------------------------------------------------------------------------- /mmdet/models/utils/norm.py: -------------------------------------------------------------------------------- 1 | import torch.nn as nn 2 | 3 | norm_cfg = { 4 | # format: layer_type: (abbreviation, module) 5 | 'BN': ('bn', nn.BatchNorm2d), 6 | 'SyncBN': ('bn', nn.SyncBatchNorm), 7 | 'GN': ('gn', nn.GroupNorm), 8 | # and potentially 'SN' 9 | } 10 | 11 | 12 | def build_norm_layer(cfg, num_features, postfix=''): 13 | """ Build normalization layer 14 | 15 | Args: 16 | cfg (dict): cfg should contain: 17 | type (str): identify norm layer type. 18 | layer args: args needed to instantiate a norm layer. 19 | requires_grad (bool): [optional] whether stop gradient updates 20 | num_features (int): number of channels from input. 21 | postfix (int, str): appended into norm abbreviation to 22 | create named layer. 23 | 24 | Returns: 25 | name (str): abbreviation + postfix 26 | layer (nn.Module): created norm layer 27 | """ 28 | assert isinstance(cfg, dict) and 'type' in cfg 29 | cfg_ = cfg.copy() 30 | 31 | layer_type = cfg_.pop('type') 32 | if layer_type not in norm_cfg: 33 | raise KeyError('Unrecognized norm type {}'.format(layer_type)) 34 | else: 35 | abbr, norm_layer = norm_cfg[layer_type] 36 | if norm_layer is None: 37 | raise NotImplementedError 38 | 39 | assert isinstance(postfix, (int, str)) 40 | name = abbr + str(postfix) 41 | 42 | requires_grad = cfg_.pop('requires_grad', True) 43 | cfg_.setdefault('eps', 1e-5) 44 | if layer_type != 'GN': 45 | layer = norm_layer(num_features, **cfg_) 46 | if layer_type == 'SyncBN': 47 | layer._specify_ddp_gpu_num(1) 48 | else: 49 | assert 'num_groups' in cfg_ 50 | layer = norm_layer(num_channels=num_features, **cfg_) 51 | 52 | for param in layer.parameters(): 53 | param.requires_grad = requires_grad 54 | 55 | return name, layer 56 | -------------------------------------------------------------------------------- /mmdet/core/bbox/samplers/instance_balanced_pos_sampler.py: -------------------------------------------------------------------------------- 1 | import numpy as np 2 | import torch 3 | 4 | from .random_sampler import RandomSampler 5 | 6 | 7 | class InstanceBalancedPosSampler(RandomSampler): 8 | 9 | def _sample_pos(self, assign_result, num_expected, **kwargs): 10 | pos_inds = torch.nonzero(assign_result.gt_inds > 0) 11 | if pos_inds.numel() != 0: 12 | pos_inds = pos_inds.squeeze(1) 13 | if pos_inds.numel() <= num_expected: 14 | return pos_inds 15 | else: 16 | unique_gt_inds = assign_result.gt_inds[pos_inds].unique() 17 | num_gts = len(unique_gt_inds) 18 | num_per_gt = int(round(num_expected / float(num_gts)) + 1) 19 | sampled_inds = [] 20 | for i in unique_gt_inds: 21 | inds = torch.nonzero(assign_result.gt_inds == i.item()) 22 | if inds.numel() != 0: 23 | inds = inds.squeeze(1) 24 | else: 25 | continue 26 | if len(inds) > num_per_gt: 27 | inds = self.random_choice(inds, num_per_gt) 28 | sampled_inds.append(inds) 29 | sampled_inds = torch.cat(sampled_inds) 30 | if len(sampled_inds) < num_expected: 31 | num_extra = num_expected - len(sampled_inds) 32 | extra_inds = np.array( 33 | list(set(pos_inds.cpu()) - set(sampled_inds.cpu()))) 34 | if len(extra_inds) > num_extra: 35 | extra_inds = self.random_choice(extra_inds, num_extra) 36 | extra_inds = torch.from_numpy(extra_inds).to( 37 | assign_result.gt_inds.device).long() 38 | sampled_inds = torch.cat([sampled_inds, extra_inds]) 39 | elif len(sampled_inds) > num_expected: 40 | sampled_inds = self.random_choice(sampled_inds, num_expected) 41 | return sampled_inds 42 | -------------------------------------------------------------------------------- /mmdet/models/detectors/fast_rcnn.py: -------------------------------------------------------------------------------- 1 | from ..registry import DETECTORS 2 | from .two_stage import TwoStageDetector 3 | 4 | 5 | @DETECTORS.register_module 6 | class FastRCNN(TwoStageDetector): 7 | 8 | def __init__(self, 9 | backbone, 10 | bbox_roi_extractor, 11 | bbox_head, 12 | train_cfg, 13 | test_cfg, 14 | neck=None, 15 | shared_head=None, 16 | mask_roi_extractor=None, 17 | mask_head=None, 18 | pretrained=None): 19 | super(FastRCNN, self).__init__( 20 | backbone=backbone, 21 | neck=neck, 22 | shared_head=shared_head, 23 | bbox_roi_extractor=bbox_roi_extractor, 24 | bbox_head=bbox_head, 25 | train_cfg=train_cfg, 26 | test_cfg=test_cfg, 27 | mask_roi_extractor=mask_roi_extractor, 28 | mask_head=mask_head, 29 | pretrained=pretrained) 30 | 31 | def forward_test(self, imgs, img_metas, proposals, **kwargs): 32 | for var, name in [(imgs, 'imgs'), (img_metas, 'img_metas')]: 33 | if not isinstance(var, list): 34 | raise TypeError('{} must be a list, but got {}'.format( 35 | name, type(var))) 36 | 37 | num_augs = len(imgs) 38 | if num_augs != len(img_metas): 39 | raise ValueError( 40 | 'num of augmentations ({}) != num of image meta ({})'.format( 41 | len(imgs), len(img_metas))) 42 | # TODO: remove the restriction of imgs_per_gpu == 1 when prepared 43 | imgs_per_gpu = imgs[0].size(0) 44 | assert imgs_per_gpu == 1 45 | 46 | if num_augs == 1: 47 | return self.simple_test(imgs[0], img_metas[0], proposals[0], 48 | **kwargs) 49 | else: 50 | return self.aug_test(imgs, img_metas, proposals, **kwargs) 51 | -------------------------------------------------------------------------------- /tools/voc_eval.py: -------------------------------------------------------------------------------- 1 | from argparse import ArgumentParser 2 | 3 | import mmcv 4 | import numpy as np 5 | 6 | from mmdet import datasets 7 | from mmdet.core import eval_map 8 | 9 | 10 | def voc_eval(result_file, dataset, iou_thr=0.5): 11 | det_results = mmcv.load(result_file) 12 | gt_bboxes = [] 13 | gt_labels = [] 14 | gt_ignore = [] 15 | for i in range(len(dataset)): 16 | ann = dataset.get_ann_info(i) 17 | bboxes = ann['bboxes'] 18 | labels = ann['labels'] 19 | if 'bboxes_ignore' in ann: 20 | ignore = np.concatenate([ 21 | np.zeros(bboxes.shape[0], dtype=np.bool), 22 | np.ones(ann['bboxes_ignore'].shape[0], dtype=np.bool) 23 | ]) 24 | gt_ignore.append(ignore) 25 | bboxes = np.vstack([bboxes, ann['bboxes_ignore']]) 26 | labels = np.concatenate([labels, ann['labels_ignore']]) 27 | gt_bboxes.append(bboxes) 28 | gt_labels.append(labels) 29 | if not gt_ignore: 30 | gt_ignore = None 31 | if hasattr(dataset, 'year') and dataset.year == 2007: 32 | dataset_name = 'voc07' 33 | else: 34 | dataset_name = dataset.CLASSES 35 | eval_map( 36 | det_results, 37 | gt_bboxes, 38 | gt_labels, 39 | gt_ignore=gt_ignore, 40 | scale_ranges=None, 41 | iou_thr=iou_thr, 42 | dataset=dataset_name, 43 | print_summary=True) 44 | 45 | 46 | def main(): 47 | parser = ArgumentParser(description='VOC Evaluation') 48 | parser.add_argument('result', help='result file path') 49 | parser.add_argument('config', help='config file path') 50 | parser.add_argument( 51 | '--iou-thr', 52 | type=float, 53 | default=0.5, 54 | help='IoU threshold for evaluation') 55 | args = parser.parse_args() 56 | cfg = mmcv.Config.fromfile(args.config) 57 | test_dataset = mmcv.runner.obj_from_dict(cfg.data.test, datasets) 58 | voc_eval(args.result, test_dataset, args.iou_thr) 59 | 60 | 61 | if __name__ == '__main__': 62 | main() 63 | -------------------------------------------------------------------------------- /mmdet_extra/ps_roi_align_ori/setup.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved. 2 | #!/usr/bin/env python 3 | 4 | import glob 5 | import os 6 | 7 | import torch 8 | from setuptools import find_packages 9 | from setuptools import setup 10 | from torch.utils.cpp_extension import CUDA_HOME 11 | from torch.utils.cpp_extension import CppExtension 12 | from torch.utils.cpp_extension import CUDAExtension 13 | 14 | requirements = ["torch", "torchvision"] 15 | 16 | 17 | def get_extensions(): 18 | this_dir = os.path.dirname(os.path.abspath(__file__)) 19 | extensions_dir = this_dir 20 | 21 | main_file = glob.glob(os.path.join(extensions_dir, "*.cpp")) 22 | source_cuda = glob.glob(os.path.join(extensions_dir, "*.cu")) 23 | 24 | sources = main_file 25 | extension = CppExtension 26 | 27 | extra_compile_args = {"cxx": []} 28 | define_macros = [] 29 | 30 | if torch.cuda.is_available() and CUDA_HOME is not None: 31 | extension = CUDAExtension 32 | sources += source_cuda 33 | define_macros += [("WITH_CUDA", None)] 34 | extra_compile_args["nvcc"] = [ 35 | "-DCUDA_HAS_FP16=1", 36 | "-D__CUDA_NO_HALF_OPERATORS__", 37 | "-D__CUDA_NO_HALF_CONVERSIONS__", 38 | "-D__CUDA_NO_HALF2_OPERATORS__", 39 | ] 40 | 41 | sources = [os.path.join(extensions_dir, s) for s in sources] 42 | 43 | include_dirs = [extensions_dir] 44 | 45 | ext_modules = [ 46 | extension( 47 | "_C", 48 | sources, 49 | include_dirs=include_dirs, 50 | define_macros=define_macros, 51 | extra_compile_args=extra_compile_args, 52 | ) 53 | ] 54 | 55 | return ext_modules 56 | 57 | 58 | setup( 59 | name="faster_rcnn", 60 | version="0.1", 61 | description="object detection in pytorch", 62 | packages=find_packages(exclude=("configs", "tests",)), 63 | # install_requires=requirements, 64 | ext_modules=get_extensions(), 65 | cmdclass={"build_ext": torch.utils.cpp_extension.BuildExtension}, 66 | ) 67 | -------------------------------------------------------------------------------- /mmdet/core/bbox/samplers/random_sampler.py: -------------------------------------------------------------------------------- 1 | import numpy as np 2 | import torch 3 | 4 | from .base_sampler import BaseSampler 5 | 6 | 7 | class RandomSampler(BaseSampler): 8 | 9 | def __init__(self, 10 | num, 11 | pos_fraction, 12 | neg_pos_ub=-1, 13 | add_gt_as_proposals=True, 14 | **kwargs): 15 | super(RandomSampler, self).__init__(num, pos_fraction, neg_pos_ub, 16 | add_gt_as_proposals) 17 | 18 | 19 | @staticmethod 20 | def random_choice(gallery, num): 21 | """Random select some elements from the gallery. 22 | 23 | It seems that Pytorch's implementation is slower than numpy so we use 24 | numpy to randperm the indices. 25 | """ 26 | assert len(gallery) >= num 27 | if isinstance(gallery, list): 28 | gallery = np.array(gallery) 29 | cands = np.arange(len(gallery)) 30 | np.random.shuffle(cands) 31 | rand_inds = cands[:num] 32 | if not isinstance(gallery, np.ndarray): 33 | rand_inds = torch.from_numpy(rand_inds).long().to(gallery.device) 34 | return gallery[rand_inds] 35 | 36 | def _sample_pos(self, assign_result, num_expected, **kwargs): 37 | """Randomly sample some positive samples.""" 38 | pos_inds = torch.nonzero(assign_result.gt_inds > 0) 39 | if pos_inds.numel() != 0: 40 | pos_inds = pos_inds.squeeze(1) 41 | if pos_inds.numel() <= num_expected: 42 | return pos_inds 43 | else: 44 | return self.random_choice(pos_inds, num_expected) 45 | 46 | def _sample_neg(self, assign_result, num_expected, **kwargs): 47 | """Randomly sample some negative samples.""" 48 | neg_inds = torch.nonzero(assign_result.gt_inds == 0) 49 | if neg_inds.numel() != 0: 50 | neg_inds = neg_inds.squeeze(1) 51 | if len(neg_inds) <= num_expected: 52 | return neg_inds 53 | else: 54 | return self.random_choice(neg_inds, num_expected) 55 | -------------------------------------------------------------------------------- /mmdet/models/losses/balanced_l1_loss.py: -------------------------------------------------------------------------------- 1 | import numpy as np 2 | import torch 3 | import torch.nn as nn 4 | 5 | from ..registry import LOSSES 6 | from .utils import weighted_loss 7 | 8 | 9 | @weighted_loss 10 | def balanced_l1_loss(pred, 11 | target, 12 | beta=1.0, 13 | alpha=0.5, 14 | gamma=1.5, 15 | reduction='mean'): 16 | assert beta > 0 17 | assert pred.size() == target.size() and target.numel() > 0 18 | 19 | diff = torch.abs(pred - target) 20 | b = np.e**(gamma / alpha) - 1 21 | loss = torch.where( 22 | diff < beta, alpha / b * 23 | (b * diff + 1) * torch.log(b * diff / beta + 1) - alpha * diff, 24 | gamma * diff + gamma / b - alpha * beta) 25 | 26 | return loss 27 | 28 | 29 | @LOSSES.register_module 30 | class BalancedL1Loss(nn.Module): 31 | """Balanced L1 Loss 32 | 33 | arXiv: https://arxiv.org/pdf/1904.02701.pdf (CVPR 2019) 34 | """ 35 | 36 | def __init__(self, 37 | alpha=0.5, 38 | gamma=1.5, 39 | beta=1.0, 40 | reduction='mean', 41 | loss_weight=1.0): 42 | super(BalancedL1Loss, self).__init__() 43 | self.alpha = alpha 44 | self.gamma = gamma 45 | self.beta = beta 46 | self.reduction = reduction 47 | self.loss_weight = loss_weight 48 | 49 | def forward(self, 50 | pred, 51 | target, 52 | weight=None, 53 | avg_factor=None, 54 | reduction_override=None, 55 | **kwargs): 56 | assert reduction_override in (None, 'none', 'mean', 'sum') 57 | reduction = ( 58 | reduction_override if reduction_override else self.reduction) 59 | loss_bbox = self.loss_weight * balanced_l1_loss( 60 | pred, 61 | target, 62 | weight, 63 | alpha=self.alpha, 64 | gamma=self.gamma, 65 | beta=self.beta, 66 | reduction=reduction, 67 | avg_factor=avg_factor, 68 | **kwargs) 69 | return loss_bbox 70 | -------------------------------------------------------------------------------- /mmdet/core/utils/dist_utils.py: -------------------------------------------------------------------------------- 1 | from collections import OrderedDict 2 | 3 | import torch.distributed as dist 4 | from mmcv.runner import OptimizerHook 5 | from torch._utils import (_flatten_dense_tensors, _take_tensors, 6 | _unflatten_dense_tensors) 7 | 8 | 9 | def _allreduce_coalesced(tensors, world_size, bucket_size_mb=-1): 10 | if bucket_size_mb > 0: 11 | bucket_size_bytes = bucket_size_mb * 1024 * 1024 12 | buckets = _take_tensors(tensors, bucket_size_bytes) 13 | else: 14 | buckets = OrderedDict() 15 | for tensor in tensors: 16 | tp = tensor.type() 17 | if tp not in buckets: 18 | buckets[tp] = [] 19 | buckets[tp].append(tensor) 20 | buckets = buckets.values() 21 | 22 | for bucket in buckets: 23 | flat_tensors = _flatten_dense_tensors(bucket) 24 | dist.all_reduce(flat_tensors) 25 | flat_tensors.div_(world_size) 26 | for tensor, synced in zip( 27 | bucket, _unflatten_dense_tensors(flat_tensors, bucket)): 28 | tensor.copy_(synced) 29 | 30 | 31 | def allreduce_grads(params, coalesce=True, bucket_size_mb=-1): 32 | 33 | grads = [ 34 | param.grad.data for param in params 35 | if param.requires_grad and param.grad is not None 36 | ] 37 | world_size = dist.get_world_size() 38 | if coalesce: 39 | _allreduce_coalesced(grads, world_size, bucket_size_mb) 40 | else: 41 | for tensor in grads: 42 | dist.all_reduce(tensor.div_(world_size)) 43 | 44 | 45 | class DistOptimizerHook(OptimizerHook): 46 | 47 | def __init__(self, grad_clip=None, coalesce=True, bucket_size_mb=-1): 48 | self.grad_clip = grad_clip 49 | self.coalesce = coalesce 50 | self.bucket_size_mb = bucket_size_mb 51 | 52 | def after_train_iter(self, runner): 53 | runner.optimizer.zero_grad() 54 | runner.outputs['loss'].backward() 55 | allreduce_grads(runner.model.parameters(), self.coalesce, 56 | self.bucket_size_mb) 57 | if self.grad_clip is not None: 58 | self.clip_grads(runner.model.parameters()) 59 | runner.optimizer.step() 60 | -------------------------------------------------------------------------------- /mmdet/ops/sigmoid_focal_loss/src/sigmoid_focal_loss.cpp: -------------------------------------------------------------------------------- 1 | // modify from 2 | // https://github.com/facebookresearch/maskrcnn-benchmark/blob/master/maskrcnn_benchmark/csrc/SigmoidFocalLoss.h 3 | #include 4 | 5 | at::Tensor SigmoidFocalLoss_forward_cuda(const at::Tensor &logits, 6 | const at::Tensor &targets, 7 | const int num_classes, 8 | const float gamma, const float alpha); 9 | 10 | at::Tensor SigmoidFocalLoss_backward_cuda(const at::Tensor &logits, 11 | const at::Tensor &targets, 12 | const at::Tensor &d_losses, 13 | const int num_classes, 14 | const float gamma, const float alpha); 15 | 16 | // Interface for Python 17 | at::Tensor SigmoidFocalLoss_forward(const at::Tensor &logits, 18 | const at::Tensor &targets, 19 | const int num_classes, const float gamma, 20 | const float alpha) { 21 | if (logits.type().is_cuda()) { 22 | return SigmoidFocalLoss_forward_cuda(logits, targets, num_classes, gamma, 23 | alpha); 24 | } 25 | AT_ERROR("SigmoidFocalLoss is not implemented on the CPU"); 26 | } 27 | 28 | at::Tensor SigmoidFocalLoss_backward(const at::Tensor &logits, 29 | const at::Tensor &targets, 30 | const at::Tensor &d_losses, 31 | const int num_classes, const float gamma, 32 | const float alpha) { 33 | if (logits.type().is_cuda()) { 34 | return SigmoidFocalLoss_backward_cuda(logits, targets, d_losses, 35 | num_classes, gamma, alpha); 36 | } 37 | AT_ERROR("SigmoidFocalLoss is not implemented on the CPU"); 38 | } 39 | 40 | PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) { 41 | m.def("forward", &SigmoidFocalLoss_forward, 42 | "SigmoidFocalLoss forward (CUDA)"); 43 | m.def("backward", &SigmoidFocalLoss_backward, 44 | "SigmoidFocalLoss backward (CUDA)"); 45 | } 46 | -------------------------------------------------------------------------------- /mmdet/apis/env.py: -------------------------------------------------------------------------------- 1 | import logging 2 | import os 3 | import random 4 | import subprocess 5 | 6 | import numpy as np 7 | import torch 8 | import torch.distributed as dist 9 | import torch.multiprocessing as mp 10 | from mmcv.runner import get_dist_info 11 | 12 | 13 | def init_dist(launcher, backend='nccl', **kwargs): 14 | if mp.get_start_method(allow_none=True) is None: 15 | mp.set_start_method('spawn') 16 | if launcher == 'pytorch': 17 | _init_dist_pytorch(backend, **kwargs) 18 | elif launcher == 'mpi': 19 | _init_dist_mpi(backend, **kwargs) 20 | elif launcher == 'slurm': 21 | _init_dist_slurm(backend, **kwargs) 22 | else: 23 | raise ValueError('Invalid launcher type: {}'.format(launcher)) 24 | 25 | 26 | def _init_dist_pytorch(backend, **kwargs): 27 | # TODO: use local_rank instead of rank % num_gpus 28 | rank = int(os.environ['RANK']) 29 | num_gpus = torch.cuda.device_count() 30 | torch.cuda.set_device(rank % num_gpus) 31 | dist.init_process_group(backend=backend, **kwargs) 32 | 33 | 34 | def _init_dist_mpi(backend, **kwargs): 35 | raise NotImplementedError 36 | 37 | 38 | def _init_dist_slurm(backend, port=29500, **kwargs): 39 | proc_id = int(os.environ['SLURM_PROCID']) 40 | ntasks = int(os.environ['SLURM_NTASKS']) 41 | node_list = os.environ['SLURM_NODELIST'] 42 | num_gpus = torch.cuda.device_count() 43 | torch.cuda.set_device(proc_id % num_gpus) 44 | addr = subprocess.getoutput( 45 | 'scontrol show hostname {} | head -n1'.format(node_list)) 46 | os.environ['MASTER_PORT'] = str(port) 47 | os.environ['MASTER_ADDR'] = addr 48 | os.environ['WORLD_SIZE'] = str(ntasks) 49 | os.environ['RANK'] = str(proc_id) 50 | dist.init_process_group(backend=backend) 51 | 52 | 53 | def set_random_seed(seed): 54 | random.seed(seed) 55 | np.random.seed(seed) 56 | torch.manual_seed(seed) 57 | torch.cuda.manual_seed_all(seed) 58 | 59 | 60 | def get_root_logger(log_level=logging.INFO): 61 | logger = logging.getLogger() 62 | if not logger.hasHandlers(): 63 | logging.basicConfig( 64 | format='%(asctime)s - %(levelname)s - %(message)s', 65 | level=log_level) 66 | rank, _ = get_dist_info() 67 | if rank != 0: 68 | logger.setLevel('ERROR') 69 | return logger 70 | -------------------------------------------------------------------------------- /mmdet/core/bbox/geometry.py: -------------------------------------------------------------------------------- 1 | import torch 2 | 3 | 4 | def bbox_overlaps(bboxes1, bboxes2, mode='iou', is_aligned=False): 5 | """Calculate overlap between two set of bboxes. 6 | 7 | If ``is_aligned`` is ``False``, then calculate the ious between each bbox 8 | of bboxes1 and bboxes2, otherwise the ious between each aligned pair of 9 | bboxes1 and bboxes2. 10 | 11 | Args: 12 | bboxes1 (Tensor): shape (m, 4) 13 | bboxes2 (Tensor): shape (n, 4), if is_aligned is ``True``, then m and n 14 | must be equal. 15 | mode (str): "iou" (intersection over union) or iof (intersection over 16 | foreground). 17 | 18 | Returns: 19 | ious(Tensor): shape (m, n) if is_aligned == False else shape (m, 1) 20 | """ 21 | 22 | assert mode in ['iou', 'iof'] 23 | 24 | rows = bboxes1.size(0) 25 | cols = bboxes2.size(0) 26 | if is_aligned: 27 | assert rows == cols 28 | 29 | if rows * cols == 0: 30 | return bboxes1.new(rows, 1) if is_aligned else bboxes1.new(rows, cols) 31 | 32 | if is_aligned: 33 | lt = torch.max(bboxes1[:, :2], bboxes2[:, :2]) # [rows, 2] 34 | rb = torch.min(bboxes1[:, 2:], bboxes2[:, 2:]) # [rows, 2] 35 | 36 | wh = (rb - lt + 1).clamp(min=0) # [rows, 2] 37 | overlap = wh[:, 0] * wh[:, 1] 38 | area1 = (bboxes1[:, 2] - bboxes1[:, 0] + 1) * ( 39 | bboxes1[:, 3] - bboxes1[:, 1] + 1) 40 | 41 | if mode == 'iou': 42 | area2 = (bboxes2[:, 2] - bboxes2[:, 0] + 1) * ( 43 | bboxes2[:, 3] - bboxes2[:, 1] + 1) 44 | ious = overlap / (area1 + area2 - overlap) 45 | else: 46 | ious = overlap / area1 47 | else: 48 | lt = torch.max(bboxes1[:, None, :2], bboxes2[:, :2]) # [rows, cols, 2] 49 | rb = torch.min(bboxes1[:, None, 2:], bboxes2[:, 2:]) # [rows, cols, 2] 50 | 51 | wh = (rb - lt + 1).clamp(min=0) # [rows, cols, 2] 52 | overlap = wh[:, :, 0] * wh[:, :, 1] 53 | area1 = (bboxes1[:, 2] - bboxes1[:, 0] + 1) * ( 54 | bboxes1[:, 3] - bboxes1[:, 1] + 1) 55 | 56 | if mode == 'iou': 57 | area2 = (bboxes2[:, 2] - bboxes2[:, 0] + 1) * ( 58 | bboxes2[:, 3] - bboxes2[:, 1] + 1) 59 | ious = overlap / (area1[:, None] + area2 - overlap) 60 | else: 61 | ious = overlap / (area1[:, None]) 62 | 63 | return ious 64 | -------------------------------------------------------------------------------- /mmdet/models/shared_heads/res_layer.py: -------------------------------------------------------------------------------- 1 | import logging 2 | 3 | import torch.nn as nn 4 | from mmcv.cnn import constant_init, kaiming_init 5 | from mmcv.runner import load_checkpoint 6 | 7 | from mmdet.core import auto_fp16 8 | from ..backbones import ResNet, make_res_layer 9 | from ..registry import SHARED_HEADS 10 | 11 | 12 | @SHARED_HEADS.register_module 13 | class ResLayer(nn.Module): 14 | 15 | def __init__(self, 16 | depth, 17 | stage=3, 18 | stride=2, 19 | dilation=1, 20 | style='pytorch', 21 | norm_cfg=dict(type='BN', requires_grad=True), 22 | norm_eval=True, 23 | with_cp=False, 24 | dcn=None): 25 | super(ResLayer, self).__init__() 26 | self.norm_eval = norm_eval 27 | self.norm_cfg = norm_cfg 28 | self.stage = stage 29 | self.fp16_enabled = False 30 | block, stage_blocks = ResNet.arch_settings[depth] 31 | stage_block = stage_blocks[stage] 32 | planes = 64 * 2**stage 33 | inplanes = 64 * 2**(stage - 1) * block.expansion 34 | 35 | res_layer = make_res_layer( 36 | block, 37 | inplanes, 38 | planes, 39 | stage_block, 40 | stride=stride, 41 | dilation=dilation, 42 | style=style, 43 | with_cp=with_cp, 44 | norm_cfg=self.norm_cfg, 45 | dcn=dcn) 46 | self.add_module('layer{}'.format(stage + 1), res_layer) 47 | 48 | def init_weights(self, pretrained=None): 49 | if isinstance(pretrained, str): 50 | logger = logging.getLogger() 51 | load_checkpoint(self, pretrained, strict=False, logger=logger) 52 | elif pretrained is None: 53 | for m in self.modules(): 54 | if isinstance(m, nn.Conv2d): 55 | kaiming_init(m) 56 | elif isinstance(m, nn.BatchNorm2d): 57 | constant_init(m, 1) 58 | else: 59 | raise TypeError('pretrained must be a str or None') 60 | 61 | @auto_fp16() 62 | def forward(self, x): 63 | res_layer = getattr(self, 'layer{}'.format(self.stage + 1)) 64 | out = res_layer(x) 65 | return out 66 | 67 | def train(self, mode=True): 68 | super(ResLayer, self).train(mode) 69 | if self.norm_eval: 70 | for m in self.modules(): 71 | if isinstance(m, nn.BatchNorm2d): 72 | m.eval() 73 | -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 |
2 | TinyDet: Accurately Detecting Small Objects within 3 | 1 GFLOPs 4 |
5 |
6 | 7 | ### Requirements 8 | - Python 3.5 or higher 9 | - PyTorch 1.2 or higher 10 | - CUDA 9.2 or higher 11 | - GCC(G++) 4.9 or higher 12 | 13 | ### Installation 14 | 15 | a. Create a conda virtual environment and activate it (Optional but recommended). 16 | 17 | ```shell 18 | conda create --name tinydet python=3.7 19 | conda activate tinydet 20 | ``` 21 | 22 | b. Install pytorch and torchvision. 23 | 24 | ```shell 25 | conda install pytorch=1.2.0 torchvision=0.4.0 -c pytorch 26 | ``` 27 | 28 | 29 | c. Install mmdet (other dependencies wil be installed automatically). 30 | 31 | ```shell 32 | pip install -r requirements.txt 33 | python setup.py build develop 34 | ``` 35 | 36 | d. Install PSRoI align. 37 | 38 | ```shell 39 | cd mmdet_extra/ps_roi_align_ori 40 | python setup.py build_ext --inplace 41 | cd ../.. 42 | ``` 43 | 44 | e. Prepare dataset and checkpoint file. 45 | 46 | Download [coco dataset](http://cocodataset.org/#download) and [checkpoint file](https://drive.google.com/drive/folders/1GC3AlgTRo9xgH_ydVYatLS8BGW0hV_w-?usp=sharing) 47 | 48 | Folder structure: 49 | ``` 50 | TinyDet 51 | ├── mmdet 52 | ├── mmdet_extra 53 | ├── tools 54 | ├── scripts 55 | ├── configs 56 | ├── data 57 | │ ├── coco 58 | │ │ ├── annotations 59 | │ │ ├── train2017 60 | │ │ ├── val2017 61 | │ │ ├── test2017 62 | ├── pth_file 63 | │ ├── mobilenetv3_bc.pt 64 | │ ├── mobilenetv3_d.pt 65 | │ ├── tinydet_L.pth 66 | │ ├── tinydet_M.pth 67 | │ ├── tinydet_S.pth 68 | ``` 69 | 70 | ### Inference 71 | ``` 72 | python -m torch.distributed.launch --nproc_per_node=1 ./tools/test.py \ 73 | ./configs/tinydet_M.py \ 74 | ./pth_file/tinydet_M.pth \ 75 | --launcher pytorch 76 | ``` 77 | 78 | 79 | ### Citation 80 | 81 | If you find our paper and code useful in your research, please consider giving a star ⭐ and citation 📝 : 82 | 83 | 84 | ``` 85 | @article{chen2023tinydet, 86 | title={TinyDet: accurately detecting small objects within 1 GFLOPs}, 87 | author={Chen, Shaoyu and Cheng, Tianheng and Fang, Jiemin and Zhang, Qian and Li, Yuan and Liu, Wenyu and Wang, Xinggang}, 88 | journal={Science China Information Sciences}, 89 | volume={66}, 90 | number={1}, 91 | pages={1--2}, 92 | year={2023}, 93 | publisher={Springer} 94 | } 95 | ``` 96 | 97 | -------------------------------------------------------------------------------- /mmdet/ops/nms/src/nms_cpu.cpp: -------------------------------------------------------------------------------- 1 | // Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved. 2 | #include 3 | 4 | template 5 | at::Tensor nms_cpu_kernel(const at::Tensor& dets, const float threshold) { 6 | AT_ASSERTM(!dets.type().is_cuda(), "dets must be a CPU tensor"); 7 | 8 | if (dets.numel() == 0) { 9 | return at::empty({0}, dets.options().dtype(at::kLong).device(at::kCPU)); 10 | } 11 | 12 | auto x1_t = dets.select(1, 0).contiguous(); 13 | auto y1_t = dets.select(1, 1).contiguous(); 14 | auto x2_t = dets.select(1, 2).contiguous(); 15 | auto y2_t = dets.select(1, 3).contiguous(); 16 | auto scores = dets.select(1, 4).contiguous(); 17 | 18 | at::Tensor areas_t = (x2_t - x1_t + 1) * (y2_t - y1_t + 1); 19 | 20 | auto order_t = std::get<1>(scores.sort(0, /* descending=*/true)); 21 | 22 | auto ndets = dets.size(0); 23 | at::Tensor suppressed_t = 24 | at::zeros({ndets}, dets.options().dtype(at::kByte).device(at::kCPU)); 25 | 26 | auto suppressed = suppressed_t.data(); 27 | auto order = order_t.data(); 28 | auto x1 = x1_t.data(); 29 | auto y1 = y1_t.data(); 30 | auto x2 = x2_t.data(); 31 | auto y2 = y2_t.data(); 32 | auto areas = areas_t.data(); 33 | 34 | for (int64_t _i = 0; _i < ndets; _i++) { 35 | auto i = order[_i]; 36 | if (suppressed[i] == 1) continue; 37 | auto ix1 = x1[i]; 38 | auto iy1 = y1[i]; 39 | auto ix2 = x2[i]; 40 | auto iy2 = y2[i]; 41 | auto iarea = areas[i]; 42 | 43 | for (int64_t _j = _i + 1; _j < ndets; _j++) { 44 | auto j = order[_j]; 45 | if (suppressed[j] == 1) continue; 46 | auto xx1 = std::max(ix1, x1[j]); 47 | auto yy1 = std::max(iy1, y1[j]); 48 | auto xx2 = std::min(ix2, x2[j]); 49 | auto yy2 = std::min(iy2, y2[j]); 50 | 51 | auto w = std::max(static_cast(0), xx2 - xx1 + 1); 52 | auto h = std::max(static_cast(0), yy2 - yy1 + 1); 53 | auto inter = w * h; 54 | auto ovr = inter / (iarea + areas[j] - inter); 55 | if (ovr >= threshold) suppressed[j] = 1; 56 | } 57 | } 58 | return at::nonzero(suppressed_t == 0).squeeze(1); 59 | } 60 | 61 | at::Tensor nms(const at::Tensor& dets, const float threshold) { 62 | at::Tensor result; 63 | AT_DISPATCH_FLOATING_TYPES(dets.scalar_type(), "nms", [&] { 64 | result = nms_cpu_kernel(dets, threshold); 65 | }); 66 | return result; 67 | } 68 | 69 | PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) { 70 | m.def("nms", &nms, "non-maximum suppression"); 71 | } -------------------------------------------------------------------------------- /mmdet_extra/ps_roi_align_ori/PSROIAlign.h: -------------------------------------------------------------------------------- 1 | #pragma once 2 | #include 3 | 4 | #include "vision.h" 5 | 6 | #ifdef WITH_CUDA 7 | 8 | #define CHECK_CUDA(x) AT_CHECK(x.type().is_cuda(), #x, " must be a CUDAtensor ") 9 | #define CHECK_CONTIGUOUS(x) \ 10 | AT_CHECK(x.is_contiguous(), #x, " must be contiguous ") 11 | #define CHECK_INPUT(x) \ 12 | CHECK_CUDA(x); \ 13 | CHECK_CONTIGUOUS(x) 14 | 15 | extern THCState* state; 16 | 17 | #endif 18 | 19 | 20 | int PSROIAlign_forward( 21 | at::Tensor bottom_data, 22 | at::Tensor bottom_rois, 23 | at::Tensor top_data, 24 | at::Tensor argmax_data, 25 | float spatial_scale, 26 | int group_size, 27 | int sampling_ratio) { 28 | 29 | #ifdef WITH_CUDA 30 | CHECK_INPUT(bottom_data); 31 | CHECK_INPUT(bottom_rois); 32 | CHECK_INPUT(top_data); 33 | CHECK_INPUT(argmax_data); 34 | 35 | int size_rois = bottom_rois.size(1); 36 | 37 | if (size_rois != 5) { 38 | printf("wrong roi size. (roi size should be 5)\n"); 39 | return 0; 40 | } 41 | 42 | cudaStream_t stream = THCState_getCurrentStream(state); 43 | 44 | PSROIAlignForwardLaucher(bottom_data, 45 | bottom_rois, 46 | top_data, 47 | argmax_data, 48 | spatial_scale, 49 | group_size, 50 | sampling_ratio, 51 | stream); 52 | #endif 53 | return 1; 54 | } 55 | 56 | int PSROIAlign_backward( 57 | at::Tensor top_diff, 58 | at::Tensor argmax_data, 59 | at::Tensor bottom_rois, 60 | at::Tensor bottom_diff, 61 | float spatial_scale, 62 | int group_size, 63 | int sampling_ratio) { 64 | 65 | #ifdef WITH_CUDA 66 | CHECK_INPUT(top_diff); 67 | CHECK_INPUT(bottom_rois); 68 | CHECK_INPUT(bottom_diff); 69 | CHECK_INPUT(argmax_data); 70 | 71 | int size_rois = bottom_rois.size(1); 72 | 73 | if (size_rois != 5) { 74 | printf("wrong roi size. (roi size should be 5)\n"); 75 | return 0; 76 | } 77 | 78 | cudaStream_t stream = THCState_getCurrentStream(state); 79 | 80 | PSROIAlignBackwardLaucher(top_diff, 81 | argmax_data, 82 | bottom_rois, 83 | bottom_diff, 84 | spatial_scale, 85 | group_size, 86 | sampling_ratio, 87 | stream); 88 | #endif 89 | return 1; 90 | } 91 | -------------------------------------------------------------------------------- /mmdet/utils/registry.py: -------------------------------------------------------------------------------- 1 | import inspect 2 | 3 | import mmcv 4 | 5 | 6 | class Registry(object): 7 | 8 | def __init__(self, name): 9 | self._name = name 10 | self._module_dict = dict() 11 | 12 | def __repr__(self): 13 | format_str = self.__class__.__name__ + '(name={}, items={})'.format( 14 | self._name, list(self._module_dict.keys())) 15 | return format_str 16 | 17 | @property 18 | def name(self): 19 | return self._name 20 | 21 | @property 22 | def module_dict(self): 23 | return self._module_dict 24 | 25 | def get(self, key): 26 | return self._module_dict.get(key, None) 27 | 28 | def _register_module(self, module_class): 29 | """Register a module. 30 | 31 | Args: 32 | module (:obj:`nn.Module`): Module to be registered. 33 | """ 34 | if not inspect.isclass(module_class): 35 | raise TypeError('module must be a class, but got {}'.format( 36 | type(module_class))) 37 | module_name = module_class.__name__ 38 | if module_name in self._module_dict: 39 | raise KeyError('{} is already registered in {}'.format( 40 | module_name, self.name)) 41 | self._module_dict[module_name] = module_class 42 | 43 | def register_module(self, cls): 44 | self._register_module(cls) 45 | return cls 46 | 47 | 48 | def build_from_cfg(cfg, registry, default_args=None): 49 | """Build a module from config dict. 50 | 51 | Args: 52 | cfg (dict): Config dict. It should at least contain the key "type". 53 | registry (:obj:`Registry`): The registry to search the type from. 54 | default_args (dict, optional): Default initialization arguments. 55 | 56 | Returns: 57 | obj: The constructed object. 58 | """ 59 | assert isinstance(cfg, dict) and 'type' in cfg 60 | assert isinstance(default_args, dict) or default_args is None 61 | args = cfg.copy() 62 | obj_type = args.pop('type') 63 | if mmcv.is_str(obj_type): 64 | obj_cls = registry.get(obj_type) 65 | if obj_cls is None: 66 | raise KeyError('{} is not in the {} registry'.format( 67 | obj_type, registry.name)) 68 | elif inspect.isclass(obj_type): 69 | obj_cls = obj_type 70 | else: 71 | raise TypeError('type must be a str or valid type, but got {}'.format( 72 | type(obj_type))) 73 | if default_args is not None: 74 | for name, value in default_args.items(): 75 | args.setdefault(name, value) 76 | return obj_cls(**args) 77 | -------------------------------------------------------------------------------- /mmdet_extra/shared_fc_bbox_head_modified.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import torch.nn.functional as F 3 | from mmdet.models import HEADS 4 | from mmdet.models import ConvFCBBoxHead 5 | from mmdet.core import (auto_fp16, bbox_target, delta2bbox, force_fp32, 6 | multiclass_nms) 7 | 8 | @HEADS.register_module 9 | class SharedFCBBoxHeadModified(ConvFCBBoxHead): 10 | 11 | def __init__(self, num_fcs=2, fc_out_channels=1024, *args, **kwargs): 12 | assert num_fcs >= 1 13 | super(SharedFCBBoxHeadModified, self).__init__( 14 | num_shared_convs=0, 15 | num_shared_fcs=num_fcs, 16 | num_cls_convs=0, 17 | num_cls_fcs=0, 18 | num_reg_convs=0, 19 | num_reg_fcs=0, 20 | fc_out_channels=fc_out_channels, 21 | *args, 22 | **kwargs) 23 | 24 | 25 | @force_fp32(apply_to=('cls_score', 'bbox_pred')) 26 | def get_det_bboxes(self, 27 | rois, 28 | cls_score, 29 | bbox_pred, 30 | img_shape, 31 | scale_factor, 32 | rescale=False, 33 | cfg=None): 34 | 35 | 36 | if isinstance(cls_score, list): 37 | cls_score = sum(cls_score) / float(len(cls_score)) 38 | scores = F.softmax(cls_score, dim=1) if cls_score is not None else None 39 | 40 | if bbox_pred is not None: 41 | bboxes = delta2bbox(rois[:, 1:], bbox_pred, self.target_means, 42 | self.target_stds, img_shape) 43 | else: 44 | bboxes = rois[:, 1:].clone() 45 | if img_shape is not None: 46 | bboxes[:, [0, 2]].clamp_(min=0, max=img_shape[1] - 1) 47 | bboxes[:, [1, 3]].clamp_(min=0, max=img_shape[0] - 1) 48 | 49 | if rescale: 50 | if isinstance(scale_factor, float): 51 | bboxes /= scale_factor 52 | else: 53 | bboxes = ( bboxes.reshape([-1,4]) / torch.from_numpy(scale_factor).to(bboxes.device) ).reshape(-1, bboxes.shape[-1])#### 54 | 55 | 56 | 57 | if cfg is None: 58 | return bboxes, scores 59 | else: 60 | 61 | det_bboxes, det_labels = multiclass_nms(bboxes, scores, 62 | cfg.score_thr, cfg.nms, 63 | cfg.max_per_img) 64 | 65 | return det_bboxes, det_labels 66 | 67 | 68 | 69 | -------------------------------------------------------------------------------- /mmdet/core/post_processing/bbox_nms.py: -------------------------------------------------------------------------------- 1 | import torch 2 | 3 | from mmdet.ops.nms import nms_wrapper 4 | 5 | 6 | def multiclass_nms(multi_bboxes, 7 | multi_scores, 8 | score_thr, 9 | nms_cfg, 10 | max_num=-1, 11 | score_factors=None): 12 | """NMS for multi-class bboxes. 13 | 14 | Args: 15 | multi_bboxes (Tensor): shape (n, #class*4) or (n, 4) 16 | multi_scores (Tensor): shape (n, #class) 17 | score_thr (float): bbox threshold, bboxes with scores lower than it 18 | will not be considered. 19 | nms_thr (float): NMS IoU threshold 20 | max_num (int): if there are more than max_num bboxes after NMS, 21 | only top max_num will be kept. 22 | score_factors (Tensor): The factors multiplied to scores before 23 | applying NMS 24 | 25 | Returns: 26 | tuple: (bboxes, labels), tensors of shape (k, 5) and (k, 1). Labels 27 | are 0-based. 28 | """ 29 | num_classes = multi_scores.shape[1] 30 | bboxes, labels = [], [] 31 | nms_cfg_ = nms_cfg.copy() 32 | nms_type = nms_cfg_.pop('type', 'nms') 33 | nms_op = getattr(nms_wrapper, nms_type) 34 | 35 | 36 | for i in range(1, num_classes): 37 | cls_inds = multi_scores[:, i] > score_thr 38 | if not cls_inds.any(): 39 | continue 40 | # get bboxes and scores of this class 41 | if multi_bboxes.shape[1] == 4: 42 | _bboxes = multi_bboxes[cls_inds, :] 43 | else: 44 | _bboxes = multi_bboxes[cls_inds, i * 4:(i + 1) * 4] 45 | _scores = multi_scores[cls_inds, i] 46 | if score_factors is not None: 47 | _scores *= score_factors[cls_inds] 48 | cls_dets = torch.cat([_bboxes, _scores[:, None]], dim=1) 49 | cls_dets, _ = nms_op(cls_dets, **nms_cfg_) 50 | cls_labels = multi_bboxes.new_full((cls_dets.shape[0], ), 51 | i - 1, 52 | dtype=torch.long) 53 | bboxes.append(cls_dets) 54 | labels.append(cls_labels) 55 | 56 | 57 | if bboxes: 58 | bboxes = torch.cat(bboxes) 59 | labels = torch.cat(labels) 60 | if bboxes.shape[0] > max_num: 61 | 62 | _, inds = bboxes[:, -1].sort(descending=True) 63 | inds = inds[:max_num] 64 | bboxes = bboxes[inds] 65 | labels = labels[inds] 66 | else: 67 | bboxes = multi_bboxes.new_zeros((0, 5)) 68 | labels = multi_bboxes.new_zeros((0, ), dtype=torch.long) 69 | 70 | 71 | return bboxes, labels 72 | -------------------------------------------------------------------------------- /mmdet/models/detectors/single_stage.py: -------------------------------------------------------------------------------- 1 | import torch.nn as nn 2 | 3 | from mmdet.core import bbox2result 4 | from .. import builder 5 | from ..registry import DETECTORS 6 | from .base import BaseDetector 7 | 8 | 9 | @DETECTORS.register_module 10 | class SingleStageDetector(BaseDetector): 11 | 12 | def __init__(self, 13 | backbone, 14 | neck=None, 15 | bbox_head=None, 16 | train_cfg=None, 17 | test_cfg=None, 18 | pretrained=None): 19 | super(SingleStageDetector, self).__init__() 20 | self.backbone = builder.build_backbone(backbone) 21 | if neck is not None: 22 | self.neck = builder.build_neck(neck) 23 | self.bbox_head = builder.build_head(bbox_head) 24 | self.train_cfg = train_cfg 25 | self.test_cfg = test_cfg 26 | self.init_weights(pretrained=pretrained) 27 | 28 | def init_weights(self, pretrained=None): 29 | super(SingleStageDetector, self).init_weights(pretrained) 30 | self.backbone.init_weights(pretrained=pretrained) 31 | if self.with_neck: 32 | if isinstance(self.neck, nn.Sequential): 33 | for m in self.neck: 34 | m.init_weights() 35 | else: 36 | self.neck.init_weights() 37 | self.bbox_head.init_weights() 38 | 39 | def extract_feat(self, img): 40 | x = self.backbone(img) 41 | if self.with_neck: 42 | x = self.neck(x) 43 | return x 44 | 45 | def forward_dummy(self, img): 46 | x = self.extract_feat(img) 47 | outs = self.bbox_head(x) 48 | return outs 49 | 50 | def forward_train(self, 51 | img, 52 | img_metas, 53 | gt_bboxes, 54 | gt_labels, 55 | gt_bboxes_ignore=None): 56 | x = self.extract_feat(img) 57 | outs = self.bbox_head(x) 58 | loss_inputs = outs + (gt_bboxes, gt_labels, img_metas, self.train_cfg) 59 | losses = self.bbox_head.loss( 60 | *loss_inputs, gt_bboxes_ignore=gt_bboxes_ignore) 61 | return losses 62 | 63 | def simple_test(self, img, img_meta, rescale=False): 64 | x = self.extract_feat(img) 65 | outs = self.bbox_head(x) 66 | bbox_inputs = outs + (img_meta, self.test_cfg, rescale) 67 | bbox_list = self.bbox_head.get_bboxes(*bbox_inputs) 68 | bbox_results = [ 69 | bbox2result(det_bboxes, det_labels, self.bbox_head.num_classes) 70 | for det_bboxes, det_labels in bbox_list 71 | ] 72 | return bbox_results[0] 73 | 74 | def aug_test(self, imgs, img_metas, rescale=False): 75 | raise NotImplementedError 76 | -------------------------------------------------------------------------------- /mmdet/ops/roi_pool/roi_pool.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import torch.nn as nn 3 | from torch.autograd import Function 4 | from torch.autograd.function import once_differentiable 5 | from torch.nn.modules.utils import _pair 6 | 7 | from . import roi_pool_cuda 8 | 9 | 10 | class RoIPoolFunction(Function): 11 | 12 | @staticmethod 13 | def forward(ctx, features, rois, out_size, spatial_scale): 14 | assert features.is_cuda 15 | out_h, out_w = _pair(out_size) 16 | assert isinstance(out_h, int) and isinstance(out_w, int) 17 | ctx.save_for_backward(rois) 18 | num_channels = features.size(1) 19 | num_rois = rois.size(0) 20 | out_size = (num_rois, num_channels, out_h, out_w) 21 | output = features.new_zeros(out_size) 22 | argmax = features.new_zeros(out_size, dtype=torch.int) 23 | roi_pool_cuda.forward(features, rois, out_h, out_w, spatial_scale, 24 | output, argmax) 25 | ctx.spatial_scale = spatial_scale 26 | ctx.feature_size = features.size() 27 | ctx.argmax = argmax 28 | 29 | return output 30 | 31 | @staticmethod 32 | @once_differentiable 33 | def backward(ctx, grad_output): 34 | assert grad_output.is_cuda 35 | spatial_scale = ctx.spatial_scale 36 | feature_size = ctx.feature_size 37 | argmax = ctx.argmax 38 | rois = ctx.saved_tensors[0] 39 | assert feature_size is not None 40 | 41 | grad_input = grad_rois = None 42 | if ctx.needs_input_grad[0]: 43 | grad_input = grad_output.new_zeros(feature_size) 44 | roi_pool_cuda.backward(grad_output.contiguous(), rois, argmax, 45 | spatial_scale, grad_input) 46 | 47 | return grad_input, grad_rois, None, None 48 | 49 | 50 | roi_pool = RoIPoolFunction.apply 51 | 52 | 53 | class RoIPool(nn.Module): 54 | 55 | def __init__(self, out_size, spatial_scale, use_torchvision=False): 56 | super(RoIPool, self).__init__() 57 | 58 | self.out_size = _pair(out_size) 59 | self.spatial_scale = float(spatial_scale) 60 | self.use_torchvision = use_torchvision 61 | 62 | def forward(self, features, rois): 63 | if self.use_torchvision: 64 | from torchvision.ops import roi_pool as tv_roi_pool 65 | return tv_roi_pool(features, rois, self.out_size, 66 | self.spatial_scale) 67 | else: 68 | return roi_pool(features, rois, self.out_size, self.spatial_scale) 69 | 70 | def __repr__(self): 71 | format_str = self.__class__.__name__ 72 | format_str += '(out_size={}, spatial_scale={}'.format( 73 | self.out_size, self.spatial_scale) 74 | format_str += ', use_torchvision={})'.format(self.use_torchvision) 75 | return format_str 76 | -------------------------------------------------------------------------------- /mmdet/ops/nms/nms_wrapper.py: -------------------------------------------------------------------------------- 1 | import numpy as np 2 | import torch 3 | 4 | from . import nms_cpu, nms_cuda 5 | from .soft_nms_cpu import soft_nms_cpu 6 | 7 | 8 | def nms(dets, iou_thr, device_id=None): 9 | """Dispatch to either CPU or GPU NMS implementations. 10 | 11 | The input can be either a torch tensor or numpy array. GPU NMS will be used 12 | if the input is a gpu tensor or device_id is specified, otherwise CPU NMS 13 | will be used. The returned type will always be the same as inputs. 14 | 15 | Arguments: 16 | dets (torch.Tensor or np.ndarray): bboxes with scores. 17 | iou_thr (float): IoU threshold for NMS. 18 | device_id (int, optional): when `dets` is a numpy array, if `device_id` 19 | is None, then cpu nms is used, otherwise gpu_nms will be used. 20 | 21 | Returns: 22 | tuple: kept bboxes and indice, which is always the same data type as 23 | the input. 24 | """ 25 | # convert dets (tensor or numpy array) to tensor 26 | if isinstance(dets, torch.Tensor): 27 | is_numpy = False 28 | dets_th = dets 29 | elif isinstance(dets, np.ndarray): 30 | is_numpy = True 31 | device = 'cpu' if device_id is None else 'cuda:{}'.format(device_id) 32 | dets_th = torch.from_numpy(dets).to(device) 33 | else: 34 | raise TypeError( 35 | 'dets must be either a Tensor or numpy array, but got {}'.format( 36 | type(dets))) 37 | 38 | # execute cpu or cuda nms 39 | if dets_th.shape[0] == 0: 40 | inds = dets_th.new_zeros(0, dtype=torch.long) 41 | else: 42 | if dets_th.is_cuda: 43 | inds = nms_cuda.nms(dets_th, iou_thr) 44 | else: 45 | inds = nms_cpu.nms(dets_th, iou_thr) 46 | 47 | if is_numpy: 48 | inds = inds.cpu().numpy() 49 | return dets[inds, :], inds 50 | 51 | 52 | def soft_nms(dets, iou_thr, method='linear', sigma=0.5, min_score=1e-3): 53 | if isinstance(dets, torch.Tensor): 54 | is_tensor = True 55 | dets_np = dets.detach().cpu().numpy() 56 | elif isinstance(dets, np.ndarray): 57 | is_tensor = False 58 | dets_np = dets 59 | else: 60 | raise TypeError( 61 | 'dets must be either a Tensor or numpy array, but got {}'.format( 62 | type(dets))) 63 | 64 | method_codes = {'linear': 1, 'gaussian': 2} 65 | if method not in method_codes: 66 | raise ValueError('Invalid method for SoftNMS: {}'.format(method)) 67 | 68 | new_dets, inds = soft_nms_cpu( 69 | dets_np, 70 | iou_thr, 71 | method=method_codes[method], 72 | sigma=sigma, 73 | min_score=min_score) 74 | 75 | 76 | if is_tensor: 77 | return dets.new_tensor(new_dets), dets.new_tensor( 78 | inds, dtype=torch.long) 79 | else: 80 | return new_dets.astype(np.float32), inds.astype(np.int64) 81 | -------------------------------------------------------------------------------- /mmdet/core/bbox/bbox_target.py: -------------------------------------------------------------------------------- 1 | import torch 2 | 3 | from ..utils import multi_apply 4 | from .transforms import bbox2delta 5 | 6 | def bbox_target(pos_bboxes_list, 7 | neg_bboxes_list, 8 | pos_gt_bboxes_list, 9 | pos_gt_labels_list, 10 | cfg, 11 | reg_classes=1, 12 | target_means=[.0, .0, .0, .0], 13 | target_stds=[1.0, 1.0, 1.0, 1.0], 14 | concat=True): 15 | labels, label_weights, bbox_targets, bbox_weights = multi_apply( 16 | bbox_target_single, 17 | pos_bboxes_list, 18 | neg_bboxes_list, 19 | pos_gt_bboxes_list, 20 | pos_gt_labels_list, 21 | cfg=cfg, 22 | reg_classes=reg_classes, 23 | target_means=target_means, 24 | target_stds=target_stds) 25 | 26 | if concat: 27 | labels = torch.cat(labels, 0) 28 | label_weights = torch.cat(label_weights, 0) 29 | bbox_targets = torch.cat(bbox_targets, 0) 30 | bbox_weights = torch.cat(bbox_weights, 0) 31 | return labels, label_weights, bbox_targets, bbox_weights 32 | 33 | def bbox_target_single(pos_bboxes, 34 | neg_bboxes, 35 | pos_gt_bboxes, 36 | pos_gt_labels, 37 | cfg, 38 | reg_classes=1, 39 | target_means=[.0, .0, .0, .0], 40 | target_stds=[1.0, 1.0, 1.0, 1.0]): 41 | num_pos = pos_bboxes.size(0) 42 | num_neg = neg_bboxes.size(0) 43 | num_samples = num_pos + num_neg 44 | labels = pos_bboxes.new_zeros(num_samples, dtype=torch.long) 45 | label_weights = pos_bboxes.new_zeros(num_samples) 46 | bbox_targets = pos_bboxes.new_zeros(num_samples, 4) 47 | bbox_weights = pos_bboxes.new_zeros(num_samples, 4) 48 | if num_pos > 0: 49 | labels[:num_pos] = pos_gt_labels 50 | pos_weight = 1.0 if cfg.pos_weight <= 0 else cfg.pos_weight 51 | label_weights[:num_pos] = pos_weight 52 | pos_bbox_targets = bbox2delta(pos_bboxes, pos_gt_bboxes, target_means, 53 | target_stds) 54 | bbox_targets[:num_pos, :] = pos_bbox_targets 55 | bbox_weights[:num_pos, :] = 1 56 | if num_neg > 0: 57 | label_weights[-num_neg:] = 1.0 58 | 59 | return labels, label_weights, bbox_targets, bbox_weights 60 | 61 | 62 | def expand_target(bbox_targets, bbox_weights, labels, num_classes): 63 | bbox_targets_expand = bbox_targets.new_zeros( 64 | (bbox_targets.size(0), 4 * num_classes)) 65 | bbox_weights_expand = bbox_weights.new_zeros( 66 | (bbox_weights.size(0), 4 * num_classes)) 67 | for i in torch.nonzero(labels > 0).squeeze(-1): 68 | start, end = labels[i] * 4, (labels[i] + 1) * 4 69 | bbox_targets_expand[i, start:end] = bbox_targets[i, :] 70 | bbox_weights_expand[i, start:end] = bbox_weights[i, :] 71 | return bbox_targets_expand, bbox_weights_expand 72 | -------------------------------------------------------------------------------- /mmdet/core/bbox/samplers/ohem_sampler.py: -------------------------------------------------------------------------------- 1 | import torch 2 | 3 | from ..transforms import bbox2roi 4 | from .base_sampler import BaseSampler 5 | 6 | 7 | class OHEMSampler(BaseSampler): 8 | 9 | def __init__(self, 10 | num, 11 | pos_fraction, 12 | context, 13 | neg_pos_ub=-1, 14 | add_gt_as_proposals=True, 15 | **kwargs): 16 | super(OHEMSampler, self).__init__(num, pos_fraction, neg_pos_ub, 17 | add_gt_as_proposals) 18 | if not hasattr(context, 'num_stages'): 19 | self.bbox_roi_extractor = context.bbox_roi_extractor 20 | self.bbox_head = context.bbox_head 21 | else: 22 | self.bbox_roi_extractor = context.bbox_roi_extractor[ 23 | context.current_stage] 24 | self.bbox_head = context.bbox_head[context.current_stage] 25 | 26 | def hard_mining(self, inds, num_expected, bboxes, labels, feats): 27 | with torch.no_grad(): 28 | rois = bbox2roi([bboxes]) 29 | bbox_feats = self.bbox_roi_extractor( 30 | feats[:self.bbox_roi_extractor.num_inputs], rois) 31 | cls_score, _ = self.bbox_head(bbox_feats) 32 | loss = self.bbox_head.loss( 33 | cls_score=cls_score, 34 | bbox_pred=None, 35 | labels=labels, 36 | label_weights=cls_score.new_ones(cls_score.size(0)), 37 | bbox_targets=None, 38 | bbox_weights=None, 39 | reduction_override='none')['loss_cls'] 40 | _, topk_loss_inds = loss.topk(num_expected) 41 | return inds[topk_loss_inds] 42 | 43 | def _sample_pos(self, 44 | assign_result, 45 | num_expected, 46 | bboxes=None, 47 | feats=None, 48 | **kwargs): 49 | # Sample some hard positive samples 50 | pos_inds = torch.nonzero(assign_result.gt_inds > 0) 51 | if pos_inds.numel() != 0: 52 | pos_inds = pos_inds.squeeze(1) 53 | if pos_inds.numel() <= num_expected: 54 | return pos_inds 55 | else: 56 | return self.hard_mining(pos_inds, num_expected, bboxes[pos_inds], 57 | assign_result.labels[pos_inds], feats) 58 | 59 | def _sample_neg(self, 60 | assign_result, 61 | num_expected, 62 | bboxes=None, 63 | feats=None, 64 | **kwargs): 65 | # Sample some hard negative samples 66 | neg_inds = torch.nonzero(assign_result.gt_inds == 0) 67 | if neg_inds.numel() != 0: 68 | neg_inds = neg_inds.squeeze(1) 69 | if len(neg_inds) <= num_expected: 70 | return neg_inds 71 | else: 72 | return self.hard_mining(neg_inds, num_expected, bboxes[neg_inds], 73 | assign_result.labels[neg_inds], feats) 74 | -------------------------------------------------------------------------------- /mmdet/core/bbox/samplers/base_sampler.py: -------------------------------------------------------------------------------- 1 | from abc import ABCMeta, abstractmethod 2 | 3 | import torch 4 | 5 | from .sampling_result import SamplingResult 6 | 7 | 8 | class BaseSampler(metaclass=ABCMeta): 9 | 10 | def __init__(self, 11 | num, 12 | pos_fraction, 13 | neg_pos_ub=-1, 14 | add_gt_as_proposals=True, 15 | **kwargs): 16 | self.num = num 17 | self.pos_fraction = pos_fraction 18 | self.neg_pos_ub = neg_pos_ub 19 | self.add_gt_as_proposals = add_gt_as_proposals 20 | self.pos_sampler = self 21 | self.neg_sampler = self 22 | 23 | @abstractmethod 24 | def _sample_pos(self, assign_result, num_expected, **kwargs): 25 | pass 26 | 27 | @abstractmethod 28 | def _sample_neg(self, assign_result, num_expected, **kwargs): 29 | pass 30 | 31 | def sample(self, 32 | assign_result, 33 | bboxes, 34 | gt_bboxes, 35 | gt_labels=None, 36 | **kwargs): 37 | """Sample positive and negative bboxes. 38 | 39 | This is a simple implementation of bbox sampling given candidates, 40 | assigning results and ground truth bboxes. 41 | 42 | Args: 43 | assign_result (:obj:`AssignResult`): Bbox assigning results. 44 | bboxes (Tensor): Boxes to be sampled from. 45 | gt_bboxes (Tensor): Ground truth bboxes. 46 | gt_labels (Tensor, optional): Class labels of ground truth bboxes. 47 | 48 | Returns: 49 | :obj:`SamplingResult`: Sampling result. 50 | """ 51 | bboxes = bboxes[:, :4] 52 | 53 | 54 | gt_flags = bboxes.new_zeros((bboxes.shape[0], ), dtype=torch.uint8) 55 | if self.add_gt_as_proposals: 56 | bboxes = torch.cat([gt_bboxes, bboxes], dim=0) 57 | assign_result.add_gt_(gt_labels) 58 | gt_ones = bboxes.new_ones(gt_bboxes.shape[0], dtype=torch.uint8) 59 | gt_flags = torch.cat([gt_ones, gt_flags]) 60 | 61 | num_expected_pos = int(self.num * self.pos_fraction) 62 | pos_inds = self.pos_sampler._sample_pos( 63 | assign_result, num_expected_pos, bboxes=bboxes, **kwargs) 64 | # We found that sampled indices have duplicated items occasionally. 65 | # (may be a bug of PyTorch) 66 | pos_inds = pos_inds.unique() 67 | num_sampled_pos = pos_inds.numel() 68 | num_expected_neg = self.num - num_sampled_pos 69 | if self.neg_pos_ub >= 0: 70 | _pos = max(1, num_sampled_pos) 71 | neg_upper_bound = int(self.neg_pos_ub * _pos) 72 | if num_expected_neg > neg_upper_bound: 73 | num_expected_neg = neg_upper_bound 74 | neg_inds = self.neg_sampler._sample_neg( 75 | assign_result, num_expected_neg, bboxes=bboxes, **kwargs) 76 | neg_inds = neg_inds.unique() 77 | 78 | 79 | return SamplingResult(pos_inds, neg_inds, bboxes, gt_bboxes, 80 | assign_result, gt_flags) 81 | -------------------------------------------------------------------------------- /mmdet/models/losses/focal_loss.py: -------------------------------------------------------------------------------- 1 | import torch.nn as nn 2 | import torch.nn.functional as F 3 | 4 | from mmdet.ops import sigmoid_focal_loss as _sigmoid_focal_loss 5 | from ..registry import LOSSES 6 | from .utils import weight_reduce_loss 7 | 8 | 9 | # This method is only for debugging 10 | def py_sigmoid_focal_loss(pred, 11 | target, 12 | weight=None, 13 | gamma=2.0, 14 | alpha=0.25, 15 | reduction='mean', 16 | avg_factor=None): 17 | pred_sigmoid = pred.sigmoid() 18 | target = target.type_as(pred) 19 | pt = (1 - pred_sigmoid) * target + pred_sigmoid * (1 - target) 20 | focal_weight = (alpha * target + (1 - alpha) * 21 | (1 - target)) * pt.pow(gamma) 22 | loss = F.binary_cross_entropy_with_logits( 23 | pred, target, reduction='none') * focal_weight 24 | loss = weight_reduce_loss(loss, weight, reduction, avg_factor) 25 | return loss 26 | 27 | 28 | def sigmoid_focal_loss(pred, 29 | target, 30 | weight=None, 31 | gamma=2.0, 32 | alpha=0.25, 33 | reduction='mean', 34 | avg_factor=None): 35 | # Function.apply does not accept keyword arguments, so the decorator 36 | # "weighted_loss" is not applicable 37 | loss = _sigmoid_focal_loss(pred, target, gamma, alpha) 38 | # TODO: find a proper way to handle the shape of weight 39 | if weight is not None: 40 | weight = weight.view(-1, 1) 41 | loss = weight_reduce_loss(loss, weight, reduction, avg_factor) 42 | return loss 43 | 44 | 45 | @LOSSES.register_module 46 | class FocalLoss(nn.Module): 47 | 48 | def __init__(self, 49 | use_sigmoid=True, 50 | gamma=2.0, 51 | alpha=0.25, 52 | reduction='mean', 53 | loss_weight=1.0): 54 | super(FocalLoss, self).__init__() 55 | assert use_sigmoid is True, 'Only sigmoid focal loss supported now.' 56 | self.use_sigmoid = use_sigmoid 57 | self.gamma = gamma 58 | self.alpha = alpha 59 | self.reduction = reduction 60 | self.loss_weight = loss_weight 61 | 62 | def forward(self, 63 | pred, 64 | target, 65 | weight=None, 66 | avg_factor=None, 67 | reduction_override=None): 68 | assert reduction_override in (None, 'none', 'mean', 'sum') 69 | reduction = ( 70 | reduction_override if reduction_override else self.reduction) 71 | if self.use_sigmoid: 72 | loss_cls = self.loss_weight * sigmoid_focal_loss( 73 | pred, 74 | target, 75 | weight, 76 | gamma=self.gamma, 77 | alpha=self.alpha, 78 | reduction=reduction, 79 | avg_factor=avg_factor) 80 | else: 81 | raise NotImplementedError 82 | return loss_cls 83 | -------------------------------------------------------------------------------- /mmdet/ops/masked_conv/src/masked_conv2d_cuda.cpp: -------------------------------------------------------------------------------- 1 | #include 2 | 3 | #include 4 | #include 5 | 6 | int MaskedIm2colForwardLaucher(const at::Tensor im, const int height, 7 | const int width, const int channels, 8 | const int kernel_h, const int kernel_w, 9 | const int pad_h, const int pad_w, 10 | const at::Tensor mask_h_idx, 11 | const at::Tensor mask_w_idx, const int mask_cnt, 12 | at::Tensor col); 13 | 14 | int MaskedCol2imForwardLaucher(const at::Tensor col, const int height, 15 | const int width, const int channels, 16 | const at::Tensor mask_h_idx, 17 | const at::Tensor mask_w_idx, const int mask_cnt, 18 | at::Tensor im); 19 | 20 | #define CHECK_CUDA(x) AT_CHECK(x.type().is_cuda(), #x, " must be a CUDAtensor ") 21 | #define CHECK_CONTIGUOUS(x) \ 22 | AT_CHECK(x.is_contiguous(), #x, " must be contiguous ") 23 | #define CHECK_INPUT(x) \ 24 | CHECK_CUDA(x); \ 25 | CHECK_CONTIGUOUS(x) 26 | 27 | int masked_im2col_forward_cuda(const at::Tensor im, const at::Tensor mask_h_idx, 28 | const at::Tensor mask_w_idx, const int kernel_h, 29 | const int kernel_w, const int pad_h, 30 | const int pad_w, at::Tensor col) { 31 | CHECK_INPUT(im); 32 | CHECK_INPUT(mask_h_idx); 33 | CHECK_INPUT(mask_w_idx); 34 | CHECK_INPUT(col); 35 | // im: (n, ic, h, w), kernel size (kh, kw) 36 | // kernel: (oc, ic * kh * kw), col: (kh * kw * ic, ow * oh) 37 | 38 | int channels = im.size(1); 39 | int height = im.size(2); 40 | int width = im.size(3); 41 | int mask_cnt = mask_h_idx.size(0); 42 | 43 | MaskedIm2colForwardLaucher(im, height, width, channels, kernel_h, kernel_w, 44 | pad_h, pad_w, mask_h_idx, mask_w_idx, mask_cnt, 45 | col); 46 | 47 | return 1; 48 | } 49 | 50 | int masked_col2im_forward_cuda(const at::Tensor col, 51 | const at::Tensor mask_h_idx, 52 | const at::Tensor mask_w_idx, int height, 53 | int width, int channels, at::Tensor im) { 54 | CHECK_INPUT(col); 55 | CHECK_INPUT(mask_h_idx); 56 | CHECK_INPUT(mask_w_idx); 57 | CHECK_INPUT(im); 58 | // im: (n, ic, h, w), kernel size (kh, kw) 59 | // kernel: (oc, ic * kh * kh), col: (kh * kw * ic, ow * oh) 60 | 61 | int mask_cnt = mask_h_idx.size(0); 62 | 63 | MaskedCol2imForwardLaucher(col, height, width, channels, mask_h_idx, 64 | mask_w_idx, mask_cnt, im); 65 | 66 | return 1; 67 | } 68 | 69 | PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) { 70 | m.def("masked_im2col_forward", &masked_im2col_forward_cuda, 71 | "masked_im2col forward (CUDA)"); 72 | m.def("masked_col2im_forward", &masked_col2im_forward_cuda, 73 | "masked_col2im forward (CUDA)"); 74 | } -------------------------------------------------------------------------------- /mmdet_extra/utils_extra.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import torch.nn as nn 3 | from tools.multadds_count import comp_multadds_fw 4 | from tools.utils import count_parameters_in_MB 5 | 6 | def convert_sync_batchnorm(module, process_group=None): 7 | module_output = module 8 | if isinstance(module, torch.nn.modules.batchnorm._BatchNorm): 9 | module_output = torch.nn.SyncBatchNorm(module.num_features, 10 | module.eps, module.momentum, 11 | module.affine, 12 | module.track_running_stats, 13 | process_group) 14 | if module.affine: 15 | module_output.weight.data = module.weight.data.clone().detach() 16 | module_output.bias.data = module.bias.data.clone().detach() 17 | # keep reuqires_grad unchanged 18 | module_output.weight.requires_grad = module.weight.requires_grad 19 | module_output.bias.requires_grad = module.bias.requires_grad 20 | module_output.running_mean = module.running_mean 21 | module_output.running_var = module.running_var 22 | module_output.num_batches_tracked = module.num_batches_tracked 23 | module_output._specify_ddp_gpu_num(1) 24 | for name, child in module.named_children(): 25 | module_output.add_module(name, convert_sync_batchnorm(child, process_group)) 26 | del module 27 | return module_output 28 | 29 | 30 | def replace_relu6_with_relu(module, process_group=None): 31 | module_output = module 32 | if isinstance(module, torch.nn.ReLU6): 33 | module_output = torch.nn.ReLU() 34 | 35 | 36 | for name, child in module.named_children(): 37 | module_output.add_module(name, replace_relu6_with_relu(child, process_group)) 38 | del module 39 | return module_output 40 | 41 | def remove_relu_series(module, process_group=None): 42 | module_output = module 43 | relu_series = (nn.ReLU, nn.PReLU, nn.ELU, nn.LeakyReLU, nn.ReLU6) 44 | if isinstance(module, relu_series): 45 | module_output = torch.nn.Sequential() 46 | 47 | 48 | for name, child in module.named_children(): 49 | module_output.add_module(name, remove_relu_series(child, process_group)) 50 | del module 51 | return module_output 52 | 53 | def get_backbone_madds(backbone, input_size, logger): 54 | 55 | input_data = torch.randn((1,3,)+input_size).cuda() 56 | backbone_madds, backbone_data = comp_multadds_fw(backbone, input_data) 57 | backbone_params = count_parameters_in_MB(backbone) 58 | 59 | 60 | logger.info("Derived Mult-Adds: [Backbone] %.2fGB", backbone_madds/1e3) 61 | logger.info("Derived Num Params: [Backbone] %.2fMB", backbone_params) 62 | 63 | 64 | 65 | class AttentionModule(nn.Module): 66 | def __init__(self, in_channels, groups=1): 67 | super(AttentionModule, self).__init__() 68 | self.attention = nn.Sequential( 69 | nn.Conv2d(in_channels, in_channels, 1, groups=groups, bias=False), 70 | nn.BatchNorm2d(in_channels), 71 | nn.Sigmoid() 72 | ) 73 | def forward(self, x): 74 | return self.attention(x).mul(x) 75 | 76 | -------------------------------------------------------------------------------- /mmdet/models/anchor_heads/retina_head.py: -------------------------------------------------------------------------------- 1 | import numpy as np 2 | import torch.nn as nn 3 | from mmcv.cnn import normal_init 4 | 5 | from ..registry import HEADS 6 | from ..utils import ConvModule, bias_init_with_prob 7 | from .anchor_head import AnchorHead 8 | 9 | 10 | @HEADS.register_module 11 | class RetinaHead(AnchorHead): 12 | 13 | def __init__(self, 14 | num_classes, 15 | in_channels, 16 | stacked_convs=4, 17 | octave_base_scale=4, 18 | scales_per_octave=3, 19 | conv_cfg=None, 20 | norm_cfg=None, 21 | **kwargs): 22 | self.stacked_convs = stacked_convs 23 | self.octave_base_scale = octave_base_scale 24 | self.scales_per_octave = scales_per_octave 25 | self.conv_cfg = conv_cfg 26 | self.norm_cfg = norm_cfg 27 | octave_scales = np.array( 28 | [2**(i / scales_per_octave) for i in range(scales_per_octave)]) 29 | anchor_scales = octave_scales * octave_base_scale 30 | super(RetinaHead, self).__init__( 31 | num_classes, in_channels, anchor_scales=anchor_scales, **kwargs) 32 | 33 | def _init_layers(self): 34 | self.relu = nn.ReLU(inplace=True) 35 | self.cls_convs = nn.ModuleList() 36 | self.reg_convs = nn.ModuleList() 37 | for i in range(self.stacked_convs): 38 | chn = self.in_channels if i == 0 else self.feat_channels 39 | self.cls_convs.append( 40 | ConvModule( 41 | chn, 42 | self.feat_channels, 43 | 3, 44 | stride=1, 45 | padding=1, 46 | conv_cfg=self.conv_cfg, 47 | norm_cfg=self.norm_cfg)) 48 | self.reg_convs.append( 49 | ConvModule( 50 | chn, 51 | self.feat_channels, 52 | 3, 53 | stride=1, 54 | padding=1, 55 | conv_cfg=self.conv_cfg, 56 | norm_cfg=self.norm_cfg)) 57 | self.retina_cls = nn.Conv2d( 58 | self.feat_channels, 59 | self.num_anchors * self.cls_out_channels, 60 | 3, 61 | padding=1) 62 | self.retina_reg = nn.Conv2d( 63 | self.feat_channels, self.num_anchors * 4, 3, padding=1) 64 | 65 | def init_weights(self): 66 | for m in self.cls_convs: 67 | normal_init(m.conv, std=0.01) 68 | for m in self.reg_convs: 69 | normal_init(m.conv, std=0.01) 70 | bias_cls = bias_init_with_prob(0.01) 71 | normal_init(self.retina_cls, std=0.01, bias=bias_cls) 72 | normal_init(self.retina_reg, std=0.01) 73 | 74 | def forward_single(self, x): 75 | cls_feat = x 76 | reg_feat = x 77 | for cls_conv in self.cls_convs: 78 | cls_feat = cls_conv(cls_feat) 79 | for reg_conv in self.reg_convs: 80 | reg_feat = reg_conv(reg_feat) 81 | cls_score = self.retina_cls(cls_feat) 82 | bbox_pred = self.retina_reg(reg_feat) 83 | return cls_score, bbox_pred 84 | -------------------------------------------------------------------------------- /mmdet/ops/roi_pool/src/roi_pool_cuda.cpp: -------------------------------------------------------------------------------- 1 | #include 2 | 3 | #include 4 | #include 5 | 6 | int ROIPoolForwardLaucher(const at::Tensor features, const at::Tensor rois, 7 | const float spatial_scale, const int channels, 8 | const int height, const int width, const int num_rois, 9 | const int pooled_h, const int pooled_w, 10 | at::Tensor output, at::Tensor argmax); 11 | 12 | int ROIPoolBackwardLaucher(const at::Tensor top_grad, const at::Tensor rois, 13 | const at::Tensor argmax, const float spatial_scale, 14 | const int batch_size, const int channels, 15 | const int height, const int width, 16 | const int num_rois, const int pooled_h, 17 | const int pooled_w, at::Tensor bottom_grad); 18 | 19 | #define CHECK_CUDA(x) AT_CHECK(x.type().is_cuda(), #x, " must be a CUDAtensor ") 20 | #define CHECK_CONTIGUOUS(x) \ 21 | AT_CHECK(x.is_contiguous(), #x, " must be contiguous ") 22 | #define CHECK_INPUT(x) \ 23 | CHECK_CUDA(x); \ 24 | CHECK_CONTIGUOUS(x) 25 | 26 | int roi_pooling_forward_cuda(at::Tensor features, at::Tensor rois, 27 | int pooled_height, int pooled_width, 28 | float spatial_scale, at::Tensor output, 29 | at::Tensor argmax) { 30 | CHECK_INPUT(features); 31 | CHECK_INPUT(rois); 32 | CHECK_INPUT(output); 33 | CHECK_INPUT(argmax); 34 | 35 | // Number of ROIs 36 | int num_rois = rois.size(0); 37 | int size_rois = rois.size(1); 38 | 39 | if (size_rois != 5) { 40 | printf("wrong roi size\n"); 41 | return 0; 42 | } 43 | 44 | int channels = features.size(1); 45 | int height = features.size(2); 46 | int width = features.size(3); 47 | 48 | ROIPoolForwardLaucher(features, rois, spatial_scale, channels, height, width, 49 | num_rois, pooled_height, pooled_width, output, argmax); 50 | 51 | return 1; 52 | } 53 | 54 | int roi_pooling_backward_cuda(at::Tensor top_grad, at::Tensor rois, 55 | at::Tensor argmax, float spatial_scale, 56 | at::Tensor bottom_grad) { 57 | CHECK_INPUT(top_grad); 58 | CHECK_INPUT(rois); 59 | CHECK_INPUT(argmax); 60 | CHECK_INPUT(bottom_grad); 61 | 62 | int pooled_height = top_grad.size(2); 63 | int pooled_width = top_grad.size(3); 64 | int num_rois = rois.size(0); 65 | int size_rois = rois.size(1); 66 | 67 | if (size_rois != 5) { 68 | printf("wrong roi size\n"); 69 | return 0; 70 | } 71 | int batch_size = bottom_grad.size(0); 72 | int channels = bottom_grad.size(1); 73 | int height = bottom_grad.size(2); 74 | int width = bottom_grad.size(3); 75 | 76 | ROIPoolBackwardLaucher(top_grad, rois, argmax, spatial_scale, batch_size, 77 | channels, height, width, num_rois, pooled_height, 78 | pooled_width, bottom_grad); 79 | 80 | return 1; 81 | } 82 | 83 | PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) { 84 | m.def("forward", &roi_pooling_forward_cuda, "Roi_Pooling forward (CUDA)"); 85 | m.def("backward", &roi_pooling_backward_cuda, "Roi_Pooling backward (CUDA)"); 86 | } 87 | -------------------------------------------------------------------------------- /mmdet/ops/roi_align/src/roi_align_cuda.cpp: -------------------------------------------------------------------------------- 1 | #include 2 | 3 | #include 4 | #include 5 | 6 | int ROIAlignForwardLaucher(const at::Tensor features, const at::Tensor rois, 7 | const float spatial_scale, const int sample_num, 8 | const int channels, const int height, 9 | const int width, const int num_rois, 10 | const int pooled_height, const int pooled_width, 11 | at::Tensor output); 12 | 13 | int ROIAlignBackwardLaucher(const at::Tensor top_grad, const at::Tensor rois, 14 | const float spatial_scale, const int sample_num, 15 | const int channels, const int height, 16 | const int width, const int num_rois, 17 | const int pooled_height, const int pooled_width, 18 | at::Tensor bottom_grad); 19 | 20 | #define CHECK_CUDA(x) AT_CHECK(x.type().is_cuda(), #x, " must be a CUDAtensor ") 21 | #define CHECK_CONTIGUOUS(x) \ 22 | AT_CHECK(x.is_contiguous(), #x, " must be contiguous ") 23 | #define CHECK_INPUT(x) \ 24 | CHECK_CUDA(x); \ 25 | CHECK_CONTIGUOUS(x) 26 | 27 | int roi_align_forward_cuda(at::Tensor features, at::Tensor rois, 28 | int pooled_height, int pooled_width, 29 | float spatial_scale, int sample_num, 30 | at::Tensor output) { 31 | CHECK_INPUT(features); 32 | CHECK_INPUT(rois); 33 | CHECK_INPUT(output); 34 | 35 | // Number of ROIs 36 | int num_rois = rois.size(0); 37 | int size_rois = rois.size(1); 38 | 39 | if (size_rois != 5) { 40 | printf("wrong roi size\n"); 41 | return 0; 42 | } 43 | 44 | int num_channels = features.size(1); 45 | int data_height = features.size(2); 46 | int data_width = features.size(3); 47 | 48 | ROIAlignForwardLaucher(features, rois, spatial_scale, sample_num, 49 | num_channels, data_height, data_width, num_rois, 50 | pooled_height, pooled_width, output); 51 | 52 | return 1; 53 | } 54 | 55 | int roi_align_backward_cuda(at::Tensor top_grad, at::Tensor rois, 56 | int pooled_height, int pooled_width, 57 | float spatial_scale, int sample_num, 58 | at::Tensor bottom_grad) { 59 | CHECK_INPUT(top_grad); 60 | CHECK_INPUT(rois); 61 | CHECK_INPUT(bottom_grad); 62 | 63 | // Number of ROIs 64 | int num_rois = rois.size(0); 65 | int size_rois = rois.size(1); 66 | if (size_rois != 5) { 67 | printf("wrong roi size\n"); 68 | return 0; 69 | } 70 | 71 | int num_channels = bottom_grad.size(1); 72 | int data_height = bottom_grad.size(2); 73 | int data_width = bottom_grad.size(3); 74 | 75 | ROIAlignBackwardLaucher(top_grad, rois, spatial_scale, sample_num, 76 | num_channels, data_height, data_width, num_rois, 77 | pooled_height, pooled_width, bottom_grad); 78 | 79 | return 1; 80 | } 81 | 82 | PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) { 83 | m.def("forward", &roi_align_forward_cuda, "Roi_Align forward (CUDA)"); 84 | m.def("backward", &roi_align_backward_cuda, "Roi_Align backward (CUDA)"); 85 | } 86 | -------------------------------------------------------------------------------- /mmdet/models/losses/utils.py: -------------------------------------------------------------------------------- 1 | import functools 2 | 3 | import torch.nn.functional as F 4 | 5 | 6 | def reduce_loss(loss, reduction): 7 | """Reduce loss as specified. 8 | 9 | Args: 10 | loss (Tensor): Elementwise loss tensor. 11 | reduction (str): Options are "none", "mean" and "sum". 12 | 13 | Return: 14 | Tensor: Reduced loss tensor. 15 | """ 16 | reduction_enum = F._Reduction.get_enum(reduction) 17 | # none: 0, elementwise_mean:1, sum: 2 18 | if reduction_enum == 0: 19 | return loss 20 | elif reduction_enum == 1: 21 | return loss.mean() 22 | elif reduction_enum == 2: 23 | return loss.sum() 24 | 25 | 26 | def weight_reduce_loss(loss, weight=None, reduction='mean', avg_factor=None): 27 | """Apply element-wise weight and reduce loss. 28 | 29 | Args: 30 | loss (Tensor): Element-wise loss. 31 | weight (Tensor): Element-wise weights. 32 | reduction (str): Same as built-in losses of PyTorch. 33 | avg_factor (float): Avarage factor when computing the mean of losses. 34 | 35 | Returns: 36 | Tensor: Processed loss values. 37 | """ 38 | # if weight is specified, apply element-wise weight 39 | if weight is not None: 40 | loss = loss * weight 41 | 42 | # if avg_factor is not specified, just reduce the loss 43 | if avg_factor is None: 44 | loss = reduce_loss(loss, reduction) 45 | else: 46 | # if reduction is mean, then average the loss by avg_factor 47 | if reduction == 'mean': 48 | loss = loss.sum() / avg_factor 49 | # if reduction is 'none', then do nothing, otherwise raise an error 50 | elif reduction != 'none': 51 | raise ValueError('avg_factor can not be used with reduction="sum"') 52 | return loss 53 | 54 | 55 | def weighted_loss(loss_func): 56 | """Create a weighted version of a given loss function. 57 | 58 | To use this decorator, the loss function must have the signature like 59 | `loss_func(pred, target, **kwargs)`. The function only needs to compute 60 | element-wise loss without any reduction. This decorator will add weight 61 | and reduction arguments to the function. The decorated function will have 62 | the signature like `loss_func(pred, target, weight=None, reduction='mean', 63 | avg_factor=None, **kwargs)`. 64 | 65 | :Example: 66 | 67 | >>> @weighted_loss 68 | >>> def l1_loss(pred, target): 69 | >>> return (pred - target).abs() 70 | 71 | >>> pred = torch.Tensor([0, 2, 3]) 72 | >>> target = torch.Tensor([1, 1, 1]) 73 | >>> weight = torch.Tensor([1, 0, 1]) 74 | 75 | >>> l1_loss(pred, target) 76 | tensor(1.3333) 77 | >>> l1_loss(pred, target, weight) 78 | tensor(1.) 79 | >>> l1_loss(pred, target, reduction='none') 80 | tensor([1., 1., 2.]) 81 | >>> l1_loss(pred, target, weight, avg_factor=2) 82 | tensor(1.5000) 83 | """ 84 | 85 | @functools.wraps(loss_func) 86 | def wrapper(pred, 87 | target, 88 | weight=None, 89 | reduction='mean', 90 | avg_factor=None, 91 | **kwargs): 92 | # get element-wise loss 93 | loss = loss_func(pred, target, **kwargs) 94 | loss = weight_reduce_loss(loss, weight, reduction, avg_factor) 95 | return loss 96 | 97 | return wrapper 98 | -------------------------------------------------------------------------------- /mmdet/datasets/xml_style.py: -------------------------------------------------------------------------------- 1 | import os.path as osp 2 | import xml.etree.ElementTree as ET 3 | 4 | import mmcv 5 | import numpy as np 6 | 7 | from .custom import CustomDataset 8 | from .registry import DATASETS 9 | 10 | 11 | @DATASETS.register_module 12 | class XMLDataset(CustomDataset): 13 | 14 | def __init__(self, min_size=None, **kwargs): 15 | super(XMLDataset, self).__init__(**kwargs) 16 | self.cat2label = {cat: i + 1 for i, cat in enumerate(self.CLASSES)} 17 | self.min_size = min_size 18 | 19 | def load_annotations(self, ann_file): 20 | img_infos = [] 21 | img_ids = mmcv.list_from_file(ann_file) 22 | for img_id in img_ids: 23 | filename = 'JPEGImages/{}.jpg'.format(img_id) 24 | xml_path = osp.join(self.img_prefix, 'Annotations', 25 | '{}.xml'.format(img_id)) 26 | tree = ET.parse(xml_path) 27 | root = tree.getroot() 28 | size = root.find('size') 29 | width = int(size.find('width').text) 30 | height = int(size.find('height').text) 31 | img_infos.append( 32 | dict(id=img_id, filename=filename, width=width, height=height)) 33 | return img_infos 34 | 35 | def get_ann_info(self, idx): 36 | img_id = self.img_infos[idx]['id'] 37 | xml_path = osp.join(self.img_prefix, 'Annotations', 38 | '{}.xml'.format(img_id)) 39 | tree = ET.parse(xml_path) 40 | root = tree.getroot() 41 | bboxes = [] 42 | labels = [] 43 | bboxes_ignore = [] 44 | labels_ignore = [] 45 | for obj in root.findall('object'): 46 | name = obj.find('name').text 47 | label = self.cat2label[name] 48 | difficult = int(obj.find('difficult').text) 49 | bnd_box = obj.find('bndbox') 50 | bbox = [ 51 | int(bnd_box.find('xmin').text), 52 | int(bnd_box.find('ymin').text), 53 | int(bnd_box.find('xmax').text), 54 | int(bnd_box.find('ymax').text) 55 | ] 56 | ignore = False 57 | if self.min_size: 58 | assert not self.test_mode 59 | w = bbox[2] - bbox[0] 60 | h = bbox[3] - bbox[1] 61 | if w < self.min_size or h < self.min_size: 62 | ignore = True 63 | if difficult or ignore: 64 | bboxes_ignore.append(bbox) 65 | labels_ignore.append(label) 66 | else: 67 | bboxes.append(bbox) 68 | labels.append(label) 69 | if not bboxes: 70 | bboxes = np.zeros((0, 4)) 71 | labels = np.zeros((0, )) 72 | else: 73 | bboxes = np.array(bboxes, ndmin=2) - 1 74 | labels = np.array(labels) 75 | if not bboxes_ignore: 76 | bboxes_ignore = np.zeros((0, 4)) 77 | labels_ignore = np.zeros((0, )) 78 | else: 79 | bboxes_ignore = np.array(bboxes_ignore, ndmin=2) - 1 80 | labels_ignore = np.array(labels_ignore) 81 | ann = dict( 82 | bboxes=bboxes.astype(np.float32), 83 | labels=labels.astype(np.int64), 84 | bboxes_ignore=bboxes_ignore.astype(np.float32), 85 | labels_ignore=labels_ignore.astype(np.int64)) 86 | return ann 87 | -------------------------------------------------------------------------------- /mmdet/ops/roi_align/roi_align.py: -------------------------------------------------------------------------------- 1 | import torch.nn as nn 2 | from torch.autograd import Function 3 | from torch.autograd.function import once_differentiable 4 | from torch.nn.modules.utils import _pair 5 | 6 | from . import roi_align_cuda 7 | 8 | 9 | class RoIAlignFunction(Function): 10 | 11 | @staticmethod 12 | def forward(ctx, features, rois, out_size, spatial_scale, sample_num=0): 13 | out_h, out_w = _pair(out_size) 14 | assert isinstance(out_h, int) and isinstance(out_w, int) 15 | ctx.spatial_scale = spatial_scale 16 | ctx.sample_num = sample_num 17 | ctx.save_for_backward(rois) 18 | ctx.feature_size = features.size() 19 | 20 | batch_size, num_channels, data_height, data_width = features.size() 21 | num_rois = rois.size(0) 22 | 23 | output = features.new_zeros(num_rois, num_channels, out_h, out_w) 24 | if features.is_cuda: 25 | roi_align_cuda.forward(features, rois, out_h, out_w, spatial_scale, 26 | sample_num, output) 27 | else: 28 | raise NotImplementedError 29 | 30 | return output 31 | 32 | @staticmethod 33 | @once_differentiable 34 | def backward(ctx, grad_output): 35 | feature_size = ctx.feature_size 36 | spatial_scale = ctx.spatial_scale 37 | sample_num = ctx.sample_num 38 | rois = ctx.saved_tensors[0] 39 | assert (feature_size is not None and grad_output.is_cuda) 40 | 41 | batch_size, num_channels, data_height, data_width = feature_size 42 | out_w = grad_output.size(3) 43 | out_h = grad_output.size(2) 44 | 45 | grad_input = grad_rois = None 46 | if ctx.needs_input_grad[0]: 47 | grad_input = rois.new_zeros(batch_size, num_channels, data_height, 48 | data_width) 49 | roi_align_cuda.backward(grad_output.contiguous(), rois, out_h, 50 | out_w, spatial_scale, sample_num, 51 | grad_input) 52 | 53 | return grad_input, grad_rois, None, None, None 54 | 55 | 56 | roi_align = RoIAlignFunction.apply 57 | 58 | 59 | class RoIAlign(nn.Module): 60 | 61 | def __init__(self, 62 | out_size, 63 | spatial_scale, 64 | sample_num=0, 65 | use_torchvision=False): 66 | super(RoIAlign, self).__init__() 67 | 68 | self.out_size = _pair(out_size) 69 | self.spatial_scale = float(spatial_scale) 70 | self.sample_num = int(sample_num) 71 | self.use_torchvision = use_torchvision 72 | 73 | def forward(self, features, rois): 74 | if self.use_torchvision: 75 | from torchvision.ops import roi_align as tv_roi_align 76 | return tv_roi_align(features, rois, self.out_size, 77 | self.spatial_scale, self.sample_num) 78 | else: 79 | return roi_align(features, rois, self.out_size, self.spatial_scale, 80 | self.sample_num) 81 | 82 | def __repr__(self): 83 | format_str = self.__class__.__name__ 84 | format_str += '(out_size={}, spatial_scale={}, sample_num={}'.format( 85 | self.out_size, self.spatial_scale, self.sample_num) 86 | format_str += ', use_torchvision={})'.format(self.use_torchvision) 87 | return format_str 88 | -------------------------------------------------------------------------------- /mmdet/models/detectors/reppoints_detector.py: -------------------------------------------------------------------------------- 1 | import torch 2 | 3 | from mmdet.core import bbox2result, bbox_mapping_back, multiclass_nms 4 | from ..registry import DETECTORS 5 | from .single_stage import SingleStageDetector 6 | 7 | 8 | @DETECTORS.register_module 9 | class RepPointsDetector(SingleStageDetector): 10 | """RepPoints: Point Set Representation for Object Detection. 11 | 12 | This detector is the implementation of: 13 | - RepPoints detector (https://arxiv.org/pdf/1904.11490) 14 | """ 15 | 16 | def __init__(self, 17 | backbone, 18 | neck, 19 | bbox_head, 20 | train_cfg=None, 21 | test_cfg=None, 22 | pretrained=None): 23 | super(RepPointsDetector, 24 | self).__init__(backbone, neck, bbox_head, train_cfg, test_cfg, 25 | pretrained) 26 | 27 | def merge_aug_results(self, aug_bboxes, aug_scores, img_metas): 28 | """Merge augmented detection bboxes and scores. 29 | 30 | Args: 31 | aug_bboxes (list[Tensor]): shape (n, 4*#class) 32 | aug_scores (list[Tensor] or None): shape (n, #class) 33 | img_shapes (list[Tensor]): shape (3, ). 34 | 35 | Returns: 36 | tuple: (bboxes, scores) 37 | """ 38 | recovered_bboxes = [] 39 | for bboxes, img_info in zip(aug_bboxes, img_metas): 40 | img_shape = img_info[0]['img_shape'] 41 | scale_factor = img_info[0]['scale_factor'] 42 | flip = img_info[0]['flip'] 43 | bboxes = bbox_mapping_back(bboxes, img_shape, scale_factor, flip) 44 | recovered_bboxes.append(bboxes) 45 | bboxes = torch.cat(recovered_bboxes, dim=0) 46 | if aug_scores is None: 47 | return bboxes 48 | else: 49 | scores = torch.cat(aug_scores, dim=0) 50 | return bboxes, scores 51 | 52 | def aug_test(self, imgs, img_metas, rescale=False): 53 | # recompute feats to save memory 54 | feats = self.extract_feats(imgs) 55 | 56 | aug_bboxes = [] 57 | aug_scores = [] 58 | for x, img_meta in zip(feats, img_metas): 59 | # only one image in the batch 60 | outs = self.bbox_head(x) 61 | bbox_inputs = outs + (img_meta, self.test_cfg, False, False) 62 | det_bboxes, det_scores = self.bbox_head.get_bboxes(*bbox_inputs)[0] 63 | aug_bboxes.append(det_bboxes) 64 | aug_scores.append(det_scores) 65 | 66 | # after merging, bboxes will be rescaled to the original image size 67 | merged_bboxes, merged_scores = self.merge_aug_results( 68 | aug_bboxes, aug_scores, img_metas) 69 | det_bboxes, det_labels = multiclass_nms(merged_bboxes, merged_scores, 70 | self.test_cfg.score_thr, 71 | self.test_cfg.nms, 72 | self.test_cfg.max_per_img) 73 | 74 | if rescale: 75 | _det_bboxes = det_bboxes 76 | else: 77 | _det_bboxes = det_bboxes.clone() 78 | _det_bboxes[:, :4] *= img_metas[0][0]['scale_factor'] 79 | bbox_results = bbox2result(_det_bboxes, det_labels, 80 | self.bbox_head.num_classes) 81 | return bbox_results 82 | -------------------------------------------------------------------------------- /mmdet/core/anchor/anchor_generator.py: -------------------------------------------------------------------------------- 1 | import torch 2 | 3 | 4 | class AnchorGenerator(object): 5 | 6 | def __init__(self, base_size, scales, ratios, scale_major=True, ctr=None): 7 | self.base_size = base_size 8 | self.scales = torch.Tensor(scales) 9 | self.ratios = torch.Tensor(ratios) 10 | self.scale_major = scale_major 11 | self.ctr = ctr 12 | self.base_anchors = self.gen_base_anchors() 13 | 14 | @property 15 | def num_base_anchors(self): 16 | return self.base_anchors.size(0) 17 | 18 | def gen_base_anchors(self): 19 | w = self.base_size 20 | h = self.base_size 21 | if self.ctr is None: 22 | x_ctr = 0.5 * (w - 1) 23 | y_ctr = 0.5 * (h - 1) 24 | else: 25 | x_ctr, y_ctr = self.ctr 26 | 27 | h_ratios = torch.sqrt(self.ratios) 28 | w_ratios = 1 / h_ratios 29 | 30 | if self.scale_major: 31 | ws = (w * w_ratios[:, None] * self.scales[None, :]).view(-1) 32 | hs = (h * h_ratios[:, None] * self.scales[None, :]).view(-1) 33 | else: 34 | ws = (w * self.scales[:, None] * w_ratios[None, :]).view(-1) 35 | hs = (h * self.scales[:, None] * h_ratios[None, :]).view(-1) 36 | 37 | # yapf: disable 38 | base_anchors = torch.stack( 39 | [ 40 | x_ctr - 0.5 * (ws - 1), y_ctr - 0.5 * (hs - 1), 41 | x_ctr + 0.5 * (ws - 1), y_ctr + 0.5 * (hs - 1) 42 | ], 43 | dim=-1).round() 44 | # yapf: enable 45 | 46 | return base_anchors 47 | 48 | def _meshgrid(self, x, y, row_major=True): 49 | xx = x.repeat(len(y)) 50 | yy = y.view(-1, 1).repeat(1, len(x)).view(-1) 51 | if row_major: 52 | return xx, yy 53 | else: 54 | return yy, xx 55 | 56 | def grid_anchors(self, featmap_size, stride=16, device='cuda'): 57 | base_anchors = self.base_anchors.to(device) 58 | 59 | feat_h, feat_w = featmap_size 60 | shift_x = torch.arange(0, feat_w, device=device) * stride 61 | shift_y = torch.arange(0, feat_h, device=device) * stride 62 | shift_xx, shift_yy = self._meshgrid(shift_x, shift_y) 63 | shifts = torch.stack([shift_xx, shift_yy, shift_xx, shift_yy], dim=-1) 64 | shifts = shifts.type_as(base_anchors) 65 | # first feat_w elements correspond to the first row of shifts 66 | # add A anchors (1, A, 4) to K shifts (K, 1, 4) to get 67 | # shifted anchors (K, A, 4), reshape to (K*A, 4) 68 | 69 | all_anchors = base_anchors[None, :, :] + shifts[:, None, :] 70 | all_anchors = all_anchors.view(-1, 4) 71 | # first A rows correspond to A anchors of (0, 0) in feature map, 72 | # then (0, 1), (0, 2), ... 73 | return all_anchors 74 | 75 | def valid_flags(self, featmap_size, valid_size, device='cuda'): 76 | feat_h, feat_w = featmap_size 77 | valid_h, valid_w = valid_size 78 | assert valid_h <= feat_h and valid_w <= feat_w 79 | valid_x = torch.zeros(feat_w, dtype=torch.uint8, device=device) 80 | valid_y = torch.zeros(feat_h, dtype=torch.uint8, device=device) 81 | valid_x[:valid_w] = 1 82 | valid_y[:valid_h] = 1 83 | valid_xx, valid_yy = self._meshgrid(valid_x, valid_y) 84 | valid = valid_xx & valid_yy 85 | valid = valid[:, None].expand( 86 | valid.size(0), self.num_base_anchors).contiguous().view(-1) 87 | return valid 88 | -------------------------------------------------------------------------------- /mmdet/core/post_processing/merge_augs.py: -------------------------------------------------------------------------------- 1 | import numpy as np 2 | import torch 3 | 4 | from mmdet.ops import nms 5 | from ..bbox import bbox_mapping_back 6 | 7 | 8 | def merge_aug_proposals(aug_proposals, img_metas, rpn_test_cfg): 9 | """Merge augmented proposals (multiscale, flip, etc.) 10 | 11 | Args: 12 | aug_proposals (list[Tensor]): proposals from different testing 13 | schemes, shape (n, 5). Note that they are not rescaled to the 14 | original image size. 15 | img_metas (list[dict]): image info including "shape_scale" and "flip". 16 | rpn_test_cfg (dict): rpn test config. 17 | 18 | Returns: 19 | Tensor: shape (n, 4), proposals corresponding to original image scale. 20 | """ 21 | recovered_proposals = [] 22 | for proposals, img_info in zip(aug_proposals, img_metas): 23 | img_shape = img_info['img_shape'] 24 | scale_factor = img_info['scale_factor'] 25 | flip = img_info['flip'] 26 | _proposals = proposals.clone() 27 | _proposals[:, :4] = bbox_mapping_back(_proposals[:, :4], img_shape, 28 | scale_factor, flip) 29 | recovered_proposals.append(_proposals) 30 | aug_proposals = torch.cat(recovered_proposals, dim=0) 31 | merged_proposals, _ = nms(aug_proposals, rpn_test_cfg.nms_thr) 32 | scores = merged_proposals[:, 4] 33 | _, order = scores.sort(0, descending=True) 34 | num = min(rpn_test_cfg.max_num, merged_proposals.shape[0]) 35 | order = order[:num] 36 | merged_proposals = merged_proposals[order, :] 37 | return merged_proposals 38 | 39 | 40 | def merge_aug_bboxes(aug_bboxes, aug_scores, img_metas, rcnn_test_cfg): 41 | """Merge augmented detection bboxes and scores. 42 | 43 | Args: 44 | aug_bboxes (list[Tensor]): shape (n, 4*#class) 45 | aug_scores (list[Tensor] or None): shape (n, #class) 46 | img_shapes (list[Tensor]): shape (3, ). 47 | rcnn_test_cfg (dict): rcnn test config. 48 | 49 | Returns: 50 | tuple: (bboxes, scores) 51 | """ 52 | recovered_bboxes = [] 53 | for bboxes, img_info in zip(aug_bboxes, img_metas): 54 | img_shape = img_info[0]['img_shape'] 55 | scale_factor = img_info[0]['scale_factor'] 56 | flip = img_info[0]['flip'] 57 | bboxes = bbox_mapping_back(bboxes, img_shape, scale_factor, flip) 58 | recovered_bboxes.append(bboxes) 59 | bboxes = torch.stack(recovered_bboxes).mean(dim=0) 60 | if aug_scores is None: 61 | return bboxes 62 | else: 63 | scores = torch.stack(aug_scores).mean(dim=0) 64 | return bboxes, scores 65 | 66 | 67 | def merge_aug_scores(aug_scores): 68 | """Merge augmented bbox scores.""" 69 | if isinstance(aug_scores[0], torch.Tensor): 70 | return torch.mean(torch.stack(aug_scores), dim=0) 71 | else: 72 | return np.mean(aug_scores, axis=0) 73 | 74 | 75 | def merge_aug_masks(aug_masks, img_metas, rcnn_test_cfg, weights=None): 76 | """Merge augmented mask prediction. 77 | 78 | Args: 79 | aug_masks (list[ndarray]): shape (n, #class, h, w) 80 | img_shapes (list[ndarray]): shape (3, ). 81 | rcnn_test_cfg (dict): rcnn test config. 82 | 83 | Returns: 84 | tuple: (bboxes, scores) 85 | """ 86 | recovered_masks = [ 87 | mask if not img_info[0]['flip'] else mask[..., ::-1] 88 | for mask, img_info in zip(aug_masks, img_metas) 89 | ] 90 | if weights is None: 91 | merged_masks = np.mean(recovered_masks, axis=0) 92 | else: 93 | merged_masks = np.average( 94 | np.array(recovered_masks), axis=0, weights=np.array(weights)) 95 | return merged_masks 96 | -------------------------------------------------------------------------------- /mmdet/ops/masked_conv/masked_conv.py: -------------------------------------------------------------------------------- 1 | import math 2 | 3 | import torch 4 | import torch.nn as nn 5 | from torch.autograd import Function 6 | from torch.autograd.function import once_differentiable 7 | from torch.nn.modules.utils import _pair 8 | 9 | from . import masked_conv2d_cuda 10 | 11 | 12 | class MaskedConv2dFunction(Function): 13 | 14 | @staticmethod 15 | def forward(ctx, features, mask, weight, bias, padding=0, stride=1): 16 | assert mask.dim() == 3 and mask.size(0) == 1 17 | assert features.dim() == 4 and features.size(0) == 1 18 | assert features.size()[2:] == mask.size()[1:] 19 | pad_h, pad_w = _pair(padding) 20 | stride_h, stride_w = _pair(stride) 21 | if stride_h != 1 or stride_w != 1: 22 | raise ValueError( 23 | 'Stride could not only be 1 in masked_conv2d currently.') 24 | if not features.is_cuda: 25 | raise NotImplementedError 26 | 27 | out_channel, in_channel, kernel_h, kernel_w = weight.size() 28 | 29 | batch_size = features.size(0) 30 | out_h = int( 31 | math.floor((features.size(2) + 2 * pad_h - 32 | (kernel_h - 1) - 1) / stride_h + 1)) 33 | out_w = int( 34 | math.floor((features.size(3) + 2 * pad_w - 35 | (kernel_h - 1) - 1) / stride_w + 1)) 36 | mask_inds = torch.nonzero(mask[0] > 0) 37 | output = features.new_zeros(batch_size, out_channel, out_h, out_w) 38 | if mask_inds.numel() > 0: 39 | mask_h_idx = mask_inds[:, 0].contiguous() 40 | mask_w_idx = mask_inds[:, 1].contiguous() 41 | data_col = features.new_zeros(in_channel * kernel_h * kernel_w, 42 | mask_inds.size(0)) 43 | masked_conv2d_cuda.masked_im2col_forward(features, mask_h_idx, 44 | mask_w_idx, kernel_h, 45 | kernel_w, pad_h, pad_w, 46 | data_col) 47 | 48 | masked_output = torch.addmm(1, bias[:, None], 1, 49 | weight.view(out_channel, -1), data_col) 50 | masked_conv2d_cuda.masked_col2im_forward(masked_output, mask_h_idx, 51 | mask_w_idx, out_h, out_w, 52 | out_channel, output) 53 | return output 54 | 55 | @staticmethod 56 | @once_differentiable 57 | def backward(ctx, grad_output): 58 | return (None, ) * 5 59 | 60 | 61 | masked_conv2d = MaskedConv2dFunction.apply 62 | 63 | 64 | class MaskedConv2d(nn.Conv2d): 65 | """A MaskedConv2d which inherits the official Conv2d. 66 | 67 | The masked forward doesn't implement the backward function and only 68 | supports the stride parameter to be 1 currently. 69 | """ 70 | 71 | def __init__(self, 72 | in_channels, 73 | out_channels, 74 | kernel_size, 75 | stride=1, 76 | padding=0, 77 | dilation=1, 78 | groups=1, 79 | bias=True): 80 | super(MaskedConv2d, 81 | self).__init__(in_channels, out_channels, kernel_size, stride, 82 | padding, dilation, groups, bias) 83 | 84 | def forward(self, input, mask=None): 85 | if mask is None: # fallback to the normal Conv2d 86 | return super(MaskedConv2d, self).forward(input) 87 | else: 88 | return masked_conv2d(input, mask, self.weight, self.bias, 89 | self.padding) 90 | -------------------------------------------------------------------------------- /mmdet/models/necks/hrfpn.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import torch.nn as nn 3 | import torch.nn.functional as F 4 | from mmcv.cnn.weight_init import caffe2_xavier_init 5 | from torch.utils.checkpoint import checkpoint 6 | 7 | from ..registry import NECKS 8 | from ..utils import ConvModule 9 | 10 | 11 | @NECKS.register_module 12 | class HRFPN(nn.Module): 13 | """HRFPN (High Resolution Feature Pyrmamids) 14 | 15 | arXiv: https://arxiv.org/abs/1904.04514 16 | 17 | Args: 18 | in_channels (list): number of channels for each branch. 19 | out_channels (int): output channels of feature pyramids. 20 | num_outs (int): number of output stages. 21 | pooling_type (str): pooling for generating feature pyramids 22 | from {MAX, AVG}. 23 | conv_cfg (dict): dictionary to construct and config conv layer. 24 | norm_cfg (dict): dictionary to construct and config norm layer. 25 | with_cp (bool): Use checkpoint or not. Using checkpoint will save some 26 | memory while slowing down the training speed. 27 | stride (int): stride of 3x3 convolutional layers 28 | """ 29 | 30 | def __init__(self, 31 | in_channels, 32 | out_channels, 33 | num_outs=5, 34 | pooling_type='AVG', 35 | conv_cfg=None, 36 | norm_cfg=None, 37 | with_cp=False, 38 | stride=1): 39 | super(HRFPN, self).__init__() 40 | assert isinstance(in_channels, list) 41 | self.in_channels = in_channels 42 | self.out_channels = out_channels 43 | self.num_ins = len(in_channels) 44 | self.num_outs = num_outs 45 | self.with_cp = with_cp 46 | self.conv_cfg = conv_cfg 47 | self.norm_cfg = norm_cfg 48 | 49 | self.reduction_conv = ConvModule( 50 | sum(in_channels), 51 | out_channels, 52 | kernel_size=1, 53 | conv_cfg=self.conv_cfg, 54 | activation=None) 55 | 56 | self.fpn_convs = nn.ModuleList() 57 | for i in range(self.num_outs): 58 | self.fpn_convs.append( 59 | ConvModule( 60 | out_channels, 61 | out_channels, 62 | kernel_size=3, 63 | padding=1, 64 | stride=stride, 65 | conv_cfg=self.conv_cfg, 66 | activation=None)) 67 | 68 | if pooling_type == 'MAX': 69 | self.pooling = F.max_pool2d 70 | else: 71 | self.pooling = F.avg_pool2d 72 | 73 | def init_weights(self): 74 | for m in self.modules(): 75 | if isinstance(m, nn.Conv2d): 76 | caffe2_xavier_init(m) 77 | 78 | def forward(self, inputs): 79 | assert len(inputs) == self.num_ins 80 | outs = [inputs[0]] 81 | for i in range(1, self.num_ins): 82 | outs.append( 83 | F.interpolate(inputs[i], scale_factor=2**i, mode='bilinear')) 84 | out = torch.cat(outs, dim=1) 85 | if out.requires_grad and self.with_cp: 86 | out = checkpoint(self.reduction_conv, out) 87 | else: 88 | out = self.reduction_conv(out) 89 | outs = [out] 90 | for i in range(1, self.num_outs): 91 | outs.append(self.pooling(out, kernel_size=2**i, stride=2**i)) 92 | outputs = [] 93 | 94 | for i in range(self.num_outs): 95 | if outs[i].requires_grad and self.with_cp: 96 | tmp_out = checkpoint(self.fpn_convs[i], outs[i]) 97 | else: 98 | tmp_out = self.fpn_convs[i](outs[i]) 99 | outputs.append(tmp_out) 100 | return tuple(outputs) 101 | -------------------------------------------------------------------------------- /mmdet/models/detectors/rpn.py: -------------------------------------------------------------------------------- 1 | import mmcv 2 | 3 | from mmdet.core import bbox_mapping, tensor2imgs 4 | from .. import builder 5 | from ..registry import DETECTORS 6 | from .base import BaseDetector 7 | from .test_mixins import RPNTestMixin 8 | 9 | 10 | @DETECTORS.register_module 11 | class RPN(BaseDetector, RPNTestMixin): 12 | 13 | def __init__(self, 14 | backbone, 15 | neck, 16 | rpn_head, 17 | train_cfg, 18 | test_cfg, 19 | pretrained=None): 20 | super(RPN, self).__init__() 21 | self.backbone = builder.build_backbone(backbone) 22 | self.neck = builder.build_neck(neck) if neck is not None else None 23 | self.rpn_head = builder.build_head(rpn_head) 24 | self.train_cfg = train_cfg 25 | self.test_cfg = test_cfg 26 | self.init_weights(pretrained=pretrained) 27 | 28 | def init_weights(self, pretrained=None): 29 | super(RPN, self).init_weights(pretrained) 30 | self.backbone.init_weights(pretrained=pretrained) 31 | if self.with_neck: 32 | self.neck.init_weights() 33 | self.rpn_head.init_weights() 34 | 35 | def extract_feat(self, img): 36 | x = self.backbone(img) 37 | if self.with_neck: 38 | x = self.neck(x) 39 | return x 40 | 41 | def forward_dummy(self, img): 42 | x = self.extract_feat(img) 43 | rpn_outs = self.rpn_head(x) 44 | return rpn_outs 45 | 46 | def forward_train(self, 47 | img, 48 | img_meta, 49 | gt_bboxes=None, 50 | gt_bboxes_ignore=None): 51 | if self.train_cfg.rpn.get('debug', False): 52 | self.rpn_head.debug_imgs = tensor2imgs(img) 53 | 54 | x = self.extract_feat(img) 55 | rpn_outs = self.rpn_head(x) 56 | 57 | rpn_loss_inputs = rpn_outs + (gt_bboxes, img_meta, self.train_cfg.rpn) 58 | losses = self.rpn_head.loss( 59 | *rpn_loss_inputs, gt_bboxes_ignore=gt_bboxes_ignore) 60 | return losses 61 | 62 | def simple_test(self, img, img_meta, rescale=False): 63 | x = self.extract_feat(img) 64 | proposal_list = self.simple_test_rpn(x, img_meta, self.test_cfg.rpn) 65 | if rescale: 66 | for proposals, meta in zip(proposal_list, img_meta): 67 | proposals[:, :4] /= meta['scale_factor'] 68 | # TODO: remove this restriction 69 | return proposal_list[0].cpu().numpy() 70 | 71 | def aug_test(self, imgs, img_metas, rescale=False): 72 | proposal_list = self.aug_test_rpn( 73 | self.extract_feats(imgs), img_metas, self.test_cfg.rpn) 74 | if not rescale: 75 | for proposals, img_meta in zip(proposal_list, img_metas[0]): 76 | img_shape = img_meta['img_shape'] 77 | scale_factor = img_meta['scale_factor'] 78 | flip = img_meta['flip'] 79 | proposals[:, :4] = bbox_mapping(proposals[:, :4], img_shape, 80 | scale_factor, flip) 81 | # TODO: remove this restriction 82 | return proposal_list[0].cpu().numpy() 83 | 84 | def show_result(self, data, result, dataset=None, top_k=20): 85 | """Show RPN proposals on the image. 86 | 87 | Although we assume batch size is 1, this method supports arbitrary 88 | batch size. 89 | """ 90 | img_tensor = data['img'][0] 91 | img_metas = data['img_meta'][0].data[0] 92 | imgs = tensor2imgs(img_tensor, **img_metas[0]['img_norm_cfg']) 93 | assert len(imgs) == len(img_metas) 94 | for img, img_meta in zip(imgs, img_metas): 95 | h, w, _ = img_meta['img_shape'] 96 | img_show = img[:h, :w, :] 97 | mmcv.imshow_bboxes(img_show, result, top_k=top_k) 98 | -------------------------------------------------------------------------------- /mmdet/models/losses/cross_entropy_loss.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import torch.nn as nn 3 | import torch.nn.functional as F 4 | 5 | from ..registry import LOSSES 6 | from .utils import weight_reduce_loss 7 | 8 | 9 | def cross_entropy(pred, label, weight=None, reduction='mean', avg_factor=None): 10 | # element-wise losses 11 | loss = F.cross_entropy(pred, label, reduction='none') 12 | 13 | # apply weights and do the reduction 14 | if weight is not None: 15 | weight = weight.float() 16 | loss = weight_reduce_loss( 17 | loss, weight=weight, reduction=reduction, avg_factor=avg_factor) 18 | 19 | return loss 20 | 21 | 22 | def _expand_binary_labels(labels, label_weights, label_channels): 23 | bin_labels = labels.new_full((labels.size(0), label_channels), 0) 24 | inds = torch.nonzero(labels >= 1).squeeze() 25 | if inds.numel() > 0: 26 | bin_labels[inds, labels[inds] - 1] = 1 27 | if label_weights is None: 28 | bin_label_weights = None 29 | else: 30 | bin_label_weights = label_weights.view(-1, 1).expand( 31 | label_weights.size(0), label_channels) 32 | return bin_labels, bin_label_weights 33 | 34 | 35 | def binary_cross_entropy(pred, 36 | label, 37 | weight=None, 38 | reduction='mean', 39 | avg_factor=None): 40 | if pred.dim() != label.dim(): 41 | label, weight = _expand_binary_labels(label, weight, pred.size(-1)) 42 | 43 | # weighted element-wise losses 44 | if weight is not None: 45 | weight = weight.float() 46 | loss = F.binary_cross_entropy_with_logits( 47 | pred, label.float(), weight, reduction='none') 48 | # do the reduction for the weighted loss 49 | loss = weight_reduce_loss(loss, reduction=reduction, avg_factor=avg_factor) 50 | 51 | return loss 52 | 53 | 54 | def mask_cross_entropy(pred, target, label, reduction='mean', avg_factor=None): 55 | # TODO: handle these two reserved arguments 56 | assert reduction == 'mean' and avg_factor is None 57 | num_rois = pred.size()[0] 58 | inds = torch.arange(0, num_rois, dtype=torch.long, device=pred.device) 59 | pred_slice = pred[inds, label].squeeze(1) 60 | return F.binary_cross_entropy_with_logits( 61 | pred_slice, target, reduction='mean')[None] 62 | 63 | 64 | @LOSSES.register_module 65 | class CrossEntropyLoss(nn.Module): 66 | 67 | def __init__(self, 68 | use_sigmoid=False, 69 | use_mask=False, 70 | reduction='mean', 71 | loss_weight=1.0): 72 | super(CrossEntropyLoss, self).__init__() 73 | assert (use_sigmoid is False) or (use_mask is False) 74 | self.use_sigmoid = use_sigmoid 75 | self.use_mask = use_mask 76 | self.reduction = reduction 77 | self.loss_weight = loss_weight 78 | 79 | if self.use_sigmoid: 80 | self.cls_criterion = binary_cross_entropy 81 | elif self.use_mask: 82 | self.cls_criterion = mask_cross_entropy 83 | else: 84 | self.cls_criterion = cross_entropy 85 | 86 | def forward(self, 87 | cls_score, 88 | label, 89 | weight=None, 90 | avg_factor=None, 91 | reduction_override=None, 92 | **kwargs): 93 | assert reduction_override in (None, 'none', 'mean', 'sum') 94 | reduction = ( 95 | reduction_override if reduction_override else self.reduction) 96 | loss_cls = self.loss_weight * self.cls_criterion( 97 | cls_score, 98 | label, 99 | weight, 100 | reduction=reduction, 101 | avg_factor=avg_factor, 102 | **kwargs) 103 | return loss_cls 104 | -------------------------------------------------------------------------------- /mmdet_extra/ps_roi_align_ori/ps_roi_align.py: -------------------------------------------------------------------------------- 1 | import math 2 | from torch import nn 3 | from torch.autograd import Function 4 | from torch.autograd.function import once_differentiable 5 | import torch 6 | from . import _C 7 | 8 | 9 | class _PSROIAlign(Function): 10 | @staticmethod 11 | def forward(ctx, bottom_data, bottom_rois, spatial_scale, roi_size, sampling_ratio, pooled_dim): 12 | ctx.spatial_scale = spatial_scale # 1./16. 13 | ctx.roi_size = roi_size # 7 14 | ctx.sampling_ratio = sampling_ratio # 2 15 | ctx.pooled_dim = pooled_dim # 10 16 | ctx.feature_size = bottom_data.size() # (B, 490, H, W) 17 | num_rois = bottom_rois.size(0) # B*K 18 | # (B*K, 10, 7, 7) 19 | top_data = torch.zeros([num_rois, pooled_dim, roi_size, roi_size], dtype=torch.float32).to(bottom_data.device) 20 | # (B*K, 10, 7, 7) 21 | argmax_data = torch.zeros([num_rois, pooled_dim, roi_size, roi_size], dtype=torch.int32).to(bottom_data.device) 22 | 23 | if bottom_data.is_cuda: 24 | _C.ps_roi_align_forward(bottom_data, # (B, 490, H, W) 25 | bottom_rois, # (B*K, 5), e.g. K = 128 26 | top_data, # (B*K, 10, 7, 7) 27 | argmax_data, # (B*K, 10, 7, 7) 28 | spatial_scale, # 1./16. 29 | roi_size, # 7 30 | sampling_ratio # 2 31 | ) 32 | ctx.save_for_backward(bottom_rois, argmax_data) 33 | else: 34 | raise NotImplementedError 35 | 36 | return top_data 37 | 38 | @staticmethod 39 | @once_differentiable 40 | def backward(ctx, top_diff): 41 | spatial_scale = ctx.spatial_scale # 1./16. 42 | roi_size = ctx.roi_size # 7 43 | sampling_ratio = ctx.sampling_ratio # 2 44 | batch_size, channels, height, width = ctx.feature_size 45 | [bottom_rois, argmax_data] = ctx.saved_tensors 46 | bottom_diff = None 47 | if ctx.needs_input_grad[0]: 48 | bottom_diff = torch.zeros([batch_size, channels, height, width], dtype=torch.float32).to(top_diff.device) 49 | _C.ps_roi_align_backward(top_diff, # (B*K, 10, 7, 7) 50 | argmax_data, # (B*K, 10, 7, 7) 51 | bottom_rois, # (B*K, 10, 7, 7) 52 | bottom_diff, # (B, 490, H, W) 53 | spatial_scale, # 1./16. 54 | roi_size, # 7 55 | sampling_ratio # 2 56 | ) 57 | 58 | return bottom_diff, None, None, None, None, None 59 | 60 | 61 | ps_roi_align = _PSROIAlign.apply 62 | 63 | 64 | class PSROIAlign_ori(nn.Module): 65 | def __init__(self, spatial_scale, roi_size, sampling_ratio, pooled_dim): 66 | super(PSROIAlign_ori, self).__init__() 67 | self.spatial_scale = spatial_scale 68 | self.roi_size = roi_size 69 | self.sampling_ratio = sampling_ratio 70 | self.pooled_dim = pooled_dim 71 | 72 | def forward(self, bottom_data, bottom_rois): 73 | 74 | return ps_roi_align(bottom_data, # (B, 490, H, W) 75 | bottom_rois, # (B*K, 5) 76 | self.spatial_scale, # 1./16. 77 | self.roi_size, # 7 78 | self.sampling_ratio, # 2 79 | self.pooled_dim # 10 80 | ) 81 | 82 | def __repr__(self): 83 | tmpstr = self.__class__.__name__ + "(" 84 | tmpstr += "spatial_scale=" + str(self.spatial_scale) 85 | tmpstr += ", roi_size=" + str(self.roi_size) 86 | tmpstr += ", sampling_ratio=" + str(self.sampling_ratio) 87 | tmpstr += ", pooled_dim=" + str(self.pooled_dim) 88 | tmpstr += ")" 89 | return tmpstr 90 | -------------------------------------------------------------------------------- /tools/divide_dataset.py: -------------------------------------------------------------------------------- 1 | import copy 2 | import random 3 | import logging 4 | from collections import defaultdict 5 | 6 | from pycocotools.coco import COCO 7 | 8 | import mmcv 9 | from mmdet.datasets import CocoDataset 10 | 11 | 12 | def build_divide_dataset(cfg, part_1_ratio=0.5, seed=520): 13 | """Need to change `coco.getImgIds()` output, i.e., 14 | `self.imgs` in a COCO instance. 15 | `self.imgs` is created by `self.dataset['images']`. 16 | Thus, hacking `self.dataset['images']` when initializing a 17 | COCO class. 18 | """ 19 | logging.info('Prepare datasets.') 20 | assert cfg.train.pop('type') == 'CocoDataset', 'Only support COCO.' 21 | annotations = mmcv.load(cfg.train.pop('ann_file')) 22 | images = annotations.pop('images') 23 | part_1_annotations = copy.copy(annotations) 24 | part_2_annotations = copy.copy(annotations) 25 | 26 | part_1_length = int(part_1_ratio * len(images)) 27 | if seed is not None: 28 | random.seed(seed) 29 | random.shuffle(images) 30 | part_1_images = images[:part_1_length] 31 | part_2_images = images[part_1_length:] 32 | 33 | part_1_annotations['images'] = part_1_images 34 | part_2_annotations['images'] = part_2_images 35 | 36 | part_1_coco = COCOFromDict(part_1_annotations) 37 | part_2_coco = COCOFromDict(part_2_annotations) 38 | # 39 | part_1_dataset = InitDatasetFromCOCOClass(**cfg.train, ann_file=part_1_coco) 40 | part_2_dataset = InitDatasetFromCOCOClass(**cfg.val, ann_file=part_2_coco) 41 | logging.info(f'Finished preparing datasets.') 42 | 43 | return part_1_dataset, part_2_dataset 44 | 45 | 46 | class InitDatasetFromCOCOClass(CocoDataset): 47 | 48 | def load_annotations(self, coco_class): 49 | """Passing ann_file the coco_class when initializing 50 | the class. 51 | """ 52 | # original loading 53 | self.coco = coco_class 54 | self.cat_ids = self.coco.getCatIds() 55 | self.cat2label = { 56 | cat_id: i + 1 57 | for i, cat_id in enumerate(self.cat_ids) 58 | } 59 | self.img_ids = self.coco.getImgIds() 60 | img_infos = [] 61 | for i in self.img_ids: 62 | info = self.coco.loadImgs([i])[0] 63 | info['filename'] = info['file_name'] 64 | img_infos.append(info) 65 | return img_infos 66 | 67 | 68 | class COCOFromDict(COCO): 69 | 70 | def __init__(self, anno_dict): 71 | self.dataset, self.anns, self.cats, self.imgs = dict(), dict(), dict(), dict() 72 | self.imgToAnns, self.catToImgs = defaultdict(list), defaultdict(list) 73 | assert type(anno_dict) == dict 74 | self.dataset = anno_dict 75 | self.createIndex() 76 | 77 | def createIndex(self): 78 | """Re-write this method because I do not wanna see the default printing... 79 | """ 80 | # Annoying! 81 | anns, cats, imgs = {}, {}, {} 82 | imgToAnns, catToImgs = defaultdict(list), defaultdict(list) 83 | if 'annotations' in self.dataset: 84 | for ann in self.dataset['annotations']: 85 | imgToAnns[ann['image_id']].append(ann) 86 | anns[ann['id']] = ann 87 | 88 | if 'images' in self.dataset: 89 | for img in self.dataset['images']: 90 | imgs[img['id']] = img 91 | 92 | if 'categories' in self.dataset: 93 | for cat in self.dataset['categories']: 94 | cats[cat['id']] = cat 95 | 96 | if 'annotations' in self.dataset and 'categories' in self.dataset: 97 | for ann in self.dataset['annotations']: 98 | catToImgs[ann['category_id']].append(ann['image_id']) 99 | 100 | self.anns = anns 101 | self.imgToAnns = imgToAnns 102 | self.catToImgs = catToImgs 103 | self.imgs = imgs 104 | self.cats = cats 105 | -------------------------------------------------------------------------------- /mmdet/models/mask_heads/fused_semantic_head.py: -------------------------------------------------------------------------------- 1 | import torch.nn as nn 2 | import torch.nn.functional as F 3 | from mmcv.cnn import kaiming_init 4 | 5 | from mmdet.core import auto_fp16, force_fp32 6 | from ..registry import HEADS 7 | from ..utils import ConvModule 8 | 9 | 10 | @HEADS.register_module 11 | class FusedSemanticHead(nn.Module): 12 | """Multi-level fused semantic segmentation head. 13 | 14 | in_1 -> 1x1 conv --- 15 | | 16 | in_2 -> 1x1 conv -- | 17 | || 18 | in_3 -> 1x1 conv - || 19 | ||| /-> 1x1 conv (mask prediction) 20 | in_4 -> 1x1 conv -----> 3x3 convs (*4) 21 | | \-> 1x1 conv (feature) 22 | in_5 -> 1x1 conv --- 23 | """ # noqa: W605 24 | 25 | def __init__(self, 26 | num_ins, 27 | fusion_level, 28 | num_convs=4, 29 | in_channels=256, 30 | conv_out_channels=256, 31 | num_classes=183, 32 | ignore_label=255, 33 | loss_weight=0.2, 34 | conv_cfg=None, 35 | norm_cfg=None): 36 | super(FusedSemanticHead, self).__init__() 37 | self.num_ins = num_ins 38 | self.fusion_level = fusion_level 39 | self.num_convs = num_convs 40 | self.in_channels = in_channels 41 | self.conv_out_channels = conv_out_channels 42 | self.num_classes = num_classes 43 | self.ignore_label = ignore_label 44 | self.loss_weight = loss_weight 45 | self.conv_cfg = conv_cfg 46 | self.norm_cfg = norm_cfg 47 | self.fp16_enabled = False 48 | 49 | self.lateral_convs = nn.ModuleList() 50 | for i in range(self.num_ins): 51 | self.lateral_convs.append( 52 | ConvModule( 53 | self.in_channels, 54 | self.in_channels, 55 | 1, 56 | conv_cfg=self.conv_cfg, 57 | norm_cfg=self.norm_cfg, 58 | inplace=False)) 59 | 60 | self.convs = nn.ModuleList() 61 | for i in range(self.num_convs): 62 | in_channels = self.in_channels if i == 0 else conv_out_channels 63 | self.convs.append( 64 | ConvModule( 65 | in_channels, 66 | conv_out_channels, 67 | 3, 68 | padding=1, 69 | conv_cfg=self.conv_cfg, 70 | norm_cfg=self.norm_cfg)) 71 | self.conv_embedding = ConvModule( 72 | conv_out_channels, 73 | conv_out_channels, 74 | 1, 75 | conv_cfg=self.conv_cfg, 76 | norm_cfg=self.norm_cfg) 77 | self.conv_logits = nn.Conv2d(conv_out_channels, self.num_classes, 1) 78 | 79 | self.criterion = nn.CrossEntropyLoss(ignore_index=ignore_label) 80 | 81 | def init_weights(self): 82 | kaiming_init(self.conv_logits) 83 | 84 | @auto_fp16() 85 | def forward(self, feats): 86 | x = self.lateral_convs[self.fusion_level](feats[self.fusion_level]) 87 | fused_size = tuple(x.shape[-2:]) 88 | for i, feat in enumerate(feats): 89 | if i != self.fusion_level: 90 | feat = F.interpolate( 91 | feat, size=fused_size, mode='bilinear', align_corners=True) 92 | x += self.lateral_convs[i](feat) 93 | 94 | for i in range(self.num_convs): 95 | x = self.convs[i](x) 96 | 97 | mask_pred = self.conv_logits(x) 98 | x = self.conv_embedding(x) 99 | return mask_pred, x 100 | 101 | @force_fp32(apply_to=('mask_pred', )) 102 | def loss(self, mask_pred, labels): 103 | labels = labels.squeeze(1).long() 104 | loss_semantic_seg = self.criterion(mask_pred, labels) 105 | loss_semantic_seg *= self.loss_weight 106 | return loss_semantic_seg 107 | -------------------------------------------------------------------------------- /mmdet/models/necks/bfp.py: -------------------------------------------------------------------------------- 1 | import torch.nn as nn 2 | import torch.nn.functional as F 3 | from mmcv.cnn import xavier_init 4 | 5 | from ..plugins import NonLocal2D 6 | from ..registry import NECKS 7 | from ..utils import ConvModule 8 | 9 | 10 | @NECKS.register_module 11 | class BFP(nn.Module): 12 | """BFP (Balanced Feature Pyrmamids) 13 | 14 | BFP takes multi-level features as inputs and gather them into a single one, 15 | then refine the gathered feature and scatter the refined results to 16 | multi-level features. This module is used in Libra R-CNN (CVPR 2019), see 17 | https://arxiv.org/pdf/1904.02701.pdf for details. 18 | 19 | Args: 20 | in_channels (int): Number of input channels (feature maps of all levels 21 | should have the same channels). 22 | num_levels (int): Number of input feature levels. 23 | conv_cfg (dict): The config dict for convolution layers. 24 | norm_cfg (dict): The config dict for normalization layers. 25 | refine_level (int): Index of integration and refine level of BSF in 26 | multi-level features from bottom to top. 27 | refine_type (str): Type of the refine op, currently support 28 | [None, 'conv', 'non_local']. 29 | """ 30 | 31 | def __init__(self, 32 | in_channels, 33 | num_levels, 34 | refine_level=2, 35 | refine_type=None, 36 | conv_cfg=None, 37 | norm_cfg=None): 38 | super(BFP, self).__init__() 39 | assert refine_type in [None, 'conv', 'non_local'] 40 | 41 | self.in_channels = in_channels 42 | self.num_levels = num_levels 43 | self.conv_cfg = conv_cfg 44 | self.norm_cfg = norm_cfg 45 | 46 | self.refine_level = refine_level 47 | self.refine_type = refine_type 48 | assert 0 <= self.refine_level < self.num_levels 49 | 50 | if self.refine_type == 'conv': 51 | self.refine = ConvModule( 52 | self.in_channels, 53 | self.in_channels, 54 | 3, 55 | padding=1, 56 | conv_cfg=self.conv_cfg, 57 | norm_cfg=self.norm_cfg) 58 | elif self.refine_type == 'non_local': 59 | self.refine = NonLocal2D( 60 | self.in_channels, 61 | reduction=1, 62 | use_scale=False, 63 | conv_cfg=self.conv_cfg, 64 | norm_cfg=self.norm_cfg) 65 | 66 | def init_weights(self): 67 | for m in self.modules(): 68 | if isinstance(m, nn.Conv2d): 69 | xavier_init(m, distribution='uniform') 70 | 71 | def forward(self, inputs): 72 | assert len(inputs) == self.num_levels 73 | 74 | # step 1: gather multi-level features by resize and average 75 | feats = [] 76 | gather_size = inputs[self.refine_level].size()[2:] 77 | for i in range(self.num_levels): 78 | if i < self.refine_level: 79 | gathered = F.adaptive_max_pool2d( 80 | inputs[i], output_size=gather_size) 81 | else: 82 | gathered = F.interpolate( 83 | inputs[i], size=gather_size, mode='nearest') 84 | feats.append(gathered) 85 | 86 | bsf = sum(feats) / len(feats) 87 | 88 | # step 2: refine gathered features 89 | if self.refine_type is not None: 90 | bsf = self.refine(bsf) 91 | 92 | # step 3: scatter refined features to multi-levels by a residual path 93 | outs = [] 94 | for i in range(self.num_levels): 95 | out_size = inputs[i].size()[2:] 96 | if i < self.refine_level: 97 | residual = F.interpolate(bsf, size=out_size, mode='nearest') 98 | else: 99 | residual = F.adaptive_max_pool2d(bsf, output_size=out_size) 100 | outs.append(residual + inputs[i]) 101 | 102 | return tuple(outs) 103 | -------------------------------------------------------------------------------- /tools/train.py: -------------------------------------------------------------------------------- 1 | from __future__ import division 2 | import argparse 3 | import os 4 | 5 | import torch 6 | from mmcv import Config 7 | 8 | from mmdet import __version__ 9 | from mmdet.apis import (get_root_logger, init_dist, set_random_seed, 10 | train_detector) 11 | from mmdet.datasets import build_dataset 12 | from mmdet.models import build_detector 13 | 14 | 15 | import sys 16 | sys.path.insert(0, os.path.abspath(os.path.dirname(__file__)) + "/../") 17 | from mmdet_extra import * 18 | 19 | 20 | 21 | def parse_args(): 22 | parser = argparse.ArgumentParser(description='Train a detector') 23 | parser.add_argument('config', help='train config file path') 24 | parser.add_argument('--work_dir', help='the dir to save logs and models') 25 | parser.add_argument( 26 | '--resume_from', help='the checkpoint file to resume from') 27 | parser.add_argument( 28 | '--validate', 29 | action='store_true', 30 | help='whether to evaluate the checkpoint during training') 31 | parser.add_argument( 32 | '--gpus', 33 | type=int, 34 | default=1, 35 | help='number of gpus to use ' 36 | '(only applicable to non-distributed training)') 37 | parser.add_argument('--seed', type=int, default=None, help='random seed') 38 | parser.add_argument( 39 | '--launcher', 40 | choices=['none', 'pytorch', 'slurm', 'mpi'], 41 | default='none', 42 | help='job launcher') 43 | parser.add_argument('--local_rank', type=int, default=0) 44 | parser.add_argument( 45 | '--autoscale-lr', 46 | action='store_true', 47 | help='automatically scale lr with the number of gpus') 48 | args = parser.parse_args() 49 | if 'LOCAL_RANK' not in os.environ: 50 | os.environ['LOCAL_RANK'] = str(args.local_rank) 51 | 52 | return args 53 | 54 | 55 | def main(): 56 | args = parse_args() 57 | 58 | cfg = Config.fromfile(args.config) 59 | # set cudnn_benchmark 60 | if cfg.get('cudnn_benchmark', False): 61 | torch.backends.cudnn.benchmark = True 62 | # update configs according to CLI args 63 | if args.work_dir is not None: 64 | cfg.work_dir = args.work_dir 65 | if args.resume_from is not None: 66 | cfg.resume_from = args.resume_from 67 | cfg.gpus = args.gpus 68 | 69 | if args.autoscale_lr: 70 | # apply the linear scaling rule (https://arxiv.org/abs/1706.02677) 71 | cfg.optimizer['lr'] = cfg.optimizer['lr'] * cfg.gpus / 8 72 | 73 | # init distributed env first, since logger depends on the dist info. 74 | if args.launcher == 'none': 75 | distributed = False 76 | else: 77 | distributed = True 78 | init_dist(args.launcher, **cfg.dist_params) 79 | 80 | # init logger before other steps 81 | logger = get_root_logger(cfg.log_level) 82 | logger.info('Distributed training: {}'.format(distributed)) 83 | 84 | # set random seeds 85 | if args.seed is not None: 86 | logger.info('Set random seed to {}'.format(args.seed)) 87 | set_random_seed(args.seed) 88 | 89 | model = build_detector( 90 | cfg.model, train_cfg=cfg.train_cfg, test_cfg=cfg.test_cfg) 91 | 92 | datasets = [build_dataset(cfg.data.train)] 93 | if len(cfg.workflow) == 2: 94 | datasets.append(build_dataset(cfg.data.val)) 95 | if cfg.checkpoint_config is not None: 96 | # save mmdet version, config file content and class names in 97 | # checkpoints as meta data 98 | cfg.checkpoint_config.meta = dict( 99 | mmdet_version=__version__, 100 | config=cfg.text, 101 | CLASSES=datasets[0].CLASSES) 102 | # add an attribute for visualization convenience 103 | model.CLASSES = datasets[0].CLASSES 104 | train_detector( 105 | model, 106 | datasets, 107 | cfg, 108 | distributed=distributed, 109 | validate=args.validate, 110 | logger=logger) 111 | 112 | 113 | if __name__ == '__main__': 114 | main() 115 | -------------------------------------------------------------------------------- /tools/detectron2pytorch.py: -------------------------------------------------------------------------------- 1 | import argparse 2 | from collections import OrderedDict 3 | 4 | import mmcv 5 | import torch 6 | 7 | arch_settings = {50: (3, 4, 6, 3), 101: (3, 4, 23, 3)} 8 | 9 | 10 | def convert_bn(blobs, state_dict, caffe_name, torch_name, converted_names): 11 | # detectron replace bn with affine channel layer 12 | state_dict[torch_name + '.bias'] = torch.from_numpy(blobs[caffe_name + 13 | '_b']) 14 | state_dict[torch_name + '.weight'] = torch.from_numpy(blobs[caffe_name + 15 | '_s']) 16 | bn_size = state_dict[torch_name + '.weight'].size() 17 | state_dict[torch_name + '.running_mean'] = torch.zeros(bn_size) 18 | state_dict[torch_name + '.running_var'] = torch.ones(bn_size) 19 | converted_names.add(caffe_name + '_b') 20 | converted_names.add(caffe_name + '_s') 21 | 22 | 23 | def convert_conv_fc(blobs, state_dict, caffe_name, torch_name, 24 | converted_names): 25 | state_dict[torch_name + '.weight'] = torch.from_numpy(blobs[caffe_name + 26 | '_w']) 27 | converted_names.add(caffe_name + '_w') 28 | if caffe_name + '_b' in blobs: 29 | state_dict[torch_name + '.bias'] = torch.from_numpy(blobs[caffe_name + 30 | '_b']) 31 | converted_names.add(caffe_name + '_b') 32 | 33 | 34 | def convert(src, dst, depth): 35 | """Convert keys in detectron pretrained ResNet models to pytorch style.""" 36 | # load arch_settings 37 | if depth not in arch_settings: 38 | raise ValueError('Only support ResNet-50 and ResNet-101 currently') 39 | block_nums = arch_settings[depth] 40 | # load caffe model 41 | caffe_model = mmcv.load(src, encoding='latin1') 42 | blobs = caffe_model['blobs'] if 'blobs' in caffe_model else caffe_model 43 | # convert to pytorch style 44 | state_dict = OrderedDict() 45 | converted_names = set() 46 | convert_conv_fc(blobs, state_dict, 'conv1', 'conv1', converted_names) 47 | convert_bn(blobs, state_dict, 'res_conv1_bn', 'bn1', converted_names) 48 | for i in range(1, len(block_nums) + 1): 49 | for j in range(block_nums[i - 1]): 50 | if j == 0: 51 | convert_conv_fc(blobs, state_dict, 52 | 'res{}_{}_branch1'.format(i + 1, j), 53 | 'layer{}.{}.downsample.0'.format(i, j), 54 | converted_names) 55 | convert_bn(blobs, state_dict, 56 | 'res{}_{}_branch1_bn'.format(i + 1, j), 57 | 'layer{}.{}.downsample.1'.format(i, j), 58 | converted_names) 59 | for k, letter in enumerate(['a', 'b', 'c']): 60 | convert_conv_fc(blobs, state_dict, 61 | 'res{}_{}_branch2{}'.format(i + 1, j, letter), 62 | 'layer{}.{}.conv{}'.format(i, j, k + 1), 63 | converted_names) 64 | convert_bn(blobs, state_dict, 65 | 'res{}_{}_branch2{}_bn'.format(i + 1, j, letter), 66 | 'layer{}.{}.bn{}'.format(i, j, 67 | k + 1), converted_names) 68 | # check if all layers are converted 69 | for key in blobs: 70 | if key not in converted_names: 71 | print('Not Convert: {}'.format(key)) 72 | # save checkpoint 73 | checkpoint = dict() 74 | checkpoint['state_dict'] = state_dict 75 | torch.save(checkpoint, dst) 76 | 77 | 78 | def main(): 79 | parser = argparse.ArgumentParser(description='Convert model keys') 80 | parser.add_argument('src', help='src detectron model path') 81 | parser.add_argument('dst', help='save path') 82 | parser.add_argument('depth', type=int, help='ResNet model depth') 83 | args = parser.parse_args() 84 | convert(args.src, args.dst, args.depth) 85 | 86 | 87 | if __name__ == '__main__': 88 | main() 89 | -------------------------------------------------------------------------------- /mmdet/ops/context_block.py: -------------------------------------------------------------------------------- 1 | import torch 2 | from mmcv.cnn import constant_init, kaiming_init 3 | from torch import nn 4 | 5 | 6 | def last_zero_init(m): 7 | if isinstance(m, nn.Sequential): 8 | constant_init(m[-1], val=0) 9 | else: 10 | constant_init(m, val=0) 11 | 12 | 13 | class ContextBlock(nn.Module): 14 | 15 | def __init__(self, 16 | inplanes, 17 | ratio, 18 | pooling_type='att', 19 | fusion_types=('channel_add', )): 20 | super(ContextBlock, self).__init__() 21 | assert pooling_type in ['avg', 'att'] 22 | assert isinstance(fusion_types, (list, tuple)) 23 | valid_fusion_types = ['channel_add', 'channel_mul'] 24 | assert all([f in valid_fusion_types for f in fusion_types]) 25 | assert len(fusion_types) > 0, 'at least one fusion should be used' 26 | self.inplanes = inplanes 27 | self.ratio = ratio 28 | self.planes = int(inplanes * ratio) 29 | self.pooling_type = pooling_type 30 | self.fusion_types = fusion_types 31 | if pooling_type == 'att': 32 | self.conv_mask = nn.Conv2d(inplanes, 1, kernel_size=1) 33 | self.softmax = nn.Softmax(dim=2) 34 | else: 35 | self.avg_pool = nn.AdaptiveAvgPool2d(1) 36 | if 'channel_add' in fusion_types: 37 | self.channel_add_conv = nn.Sequential( 38 | nn.Conv2d(self.inplanes, self.planes, kernel_size=1), 39 | nn.LayerNorm([self.planes, 1, 1]), 40 | nn.ReLU(inplace=True), # yapf: disable 41 | nn.Conv2d(self.planes, self.inplanes, kernel_size=1)) 42 | else: 43 | self.channel_add_conv = None 44 | if 'channel_mul' in fusion_types: 45 | self.channel_mul_conv = nn.Sequential( 46 | nn.Conv2d(self.inplanes, self.planes, kernel_size=1), 47 | nn.LayerNorm([self.planes, 1, 1]), 48 | nn.ReLU(inplace=True), # yapf: disable 49 | nn.Conv2d(self.planes, self.inplanes, kernel_size=1)) 50 | else: 51 | self.channel_mul_conv = None 52 | self.reset_parameters() 53 | 54 | def reset_parameters(self): 55 | if self.pooling_type == 'att': 56 | kaiming_init(self.conv_mask, mode='fan_in') 57 | self.conv_mask.inited = True 58 | 59 | if self.channel_add_conv is not None: 60 | last_zero_init(self.channel_add_conv) 61 | if self.channel_mul_conv is not None: 62 | last_zero_init(self.channel_mul_conv) 63 | 64 | def spatial_pool(self, x): 65 | batch, channel, height, width = x.size() 66 | if self.pooling_type == 'att': 67 | input_x = x 68 | # [N, C, H * W] 69 | input_x = input_x.view(batch, channel, height * width) 70 | # [N, 1, C, H * W] 71 | input_x = input_x.unsqueeze(1) 72 | # [N, 1, H, W] 73 | context_mask = self.conv_mask(x) 74 | # [N, 1, H * W] 75 | context_mask = context_mask.view(batch, 1, height * width) 76 | # [N, 1, H * W] 77 | context_mask = self.softmax(context_mask) 78 | # [N, 1, H * W, 1] 79 | context_mask = context_mask.unsqueeze(-1) 80 | # [N, 1, C, 1] 81 | context = torch.matmul(input_x, context_mask) 82 | # [N, C, 1, 1] 83 | context = context.view(batch, channel, 1, 1) 84 | else: 85 | # [N, C, 1, 1] 86 | context = self.avg_pool(x) 87 | 88 | return context 89 | 90 | def forward(self, x): 91 | # [N, C, 1, 1] 92 | context = self.spatial_pool(x) 93 | 94 | out = x 95 | if self.channel_mul_conv is not None: 96 | # [N, C, 1, 1] 97 | channel_mul_term = torch.sigmoid(self.channel_mul_conv(context)) 98 | out = out * channel_mul_term 99 | if self.channel_add_conv is not None: 100 | # [N, C, 1, 1] 101 | channel_add_term = self.channel_add_conv(context) 102 | out = out + channel_add_term 103 | 104 | return out 105 | -------------------------------------------------------------------------------- /mmdet/models/plugins/non_local.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import torch.nn as nn 3 | from mmcv.cnn import constant_init, normal_init 4 | 5 | from ..utils import ConvModule 6 | 7 | 8 | class NonLocal2D(nn.Module): 9 | """Non-local module. 10 | 11 | See https://arxiv.org/abs/1711.07971 for details. 12 | 13 | Args: 14 | in_channels (int): Channels of the input feature map. 15 | reduction (int): Channel reduction ratio. 16 | use_scale (bool): Whether to scale pairwise_weight by 1/inter_channels. 17 | conv_cfg (dict): The config dict for convolution layers. 18 | (only applicable to conv_out) 19 | norm_cfg (dict): The config dict for normalization layers. 20 | (only applicable to conv_out) 21 | mode (str): Options are `embedded_gaussian` and `dot_product`. 22 | """ 23 | 24 | def __init__(self, 25 | in_channels, 26 | reduction=2, 27 | use_scale=True, 28 | conv_cfg=None, 29 | norm_cfg=None, 30 | mode='embedded_gaussian'): 31 | super(NonLocal2D, self).__init__() 32 | self.in_channels = in_channels 33 | self.reduction = reduction 34 | self.use_scale = use_scale 35 | self.inter_channels = in_channels // reduction 36 | self.mode = mode 37 | assert mode in ['embedded_gaussian', 'dot_product'] 38 | 39 | # g, theta, phi are actually `nn.Conv2d`. Here we use ConvModule for 40 | # potential usage. 41 | self.g = ConvModule( 42 | self.in_channels, 43 | self.inter_channels, 44 | kernel_size=1, 45 | activation=None) 46 | self.theta = ConvModule( 47 | self.in_channels, 48 | self.inter_channels, 49 | kernel_size=1, 50 | activation=None) 51 | self.phi = ConvModule( 52 | self.in_channels, 53 | self.inter_channels, 54 | kernel_size=1, 55 | activation=None) 56 | self.conv_out = ConvModule( 57 | self.inter_channels, 58 | self.in_channels, 59 | kernel_size=1, 60 | conv_cfg=conv_cfg, 61 | norm_cfg=norm_cfg, 62 | activation=None) 63 | 64 | self.init_weights() 65 | 66 | def init_weights(self, std=0.01, zeros_init=True): 67 | for m in [self.g, self.theta, self.phi]: 68 | normal_init(m.conv, std=std) 69 | if zeros_init: 70 | constant_init(self.conv_out.conv, 0) 71 | else: 72 | normal_init(self.conv_out.conv, std=std) 73 | 74 | def embedded_gaussian(self, theta_x, phi_x): 75 | # pairwise_weight: [N, HxW, HxW] 76 | pairwise_weight = torch.matmul(theta_x, phi_x) 77 | if self.use_scale: 78 | # theta_x.shape[-1] is `self.inter_channels` 79 | pairwise_weight /= theta_x.shape[-1]**-0.5 80 | pairwise_weight = pairwise_weight.softmax(dim=-1) 81 | return pairwise_weight 82 | 83 | def dot_product(self, theta_x, phi_x): 84 | # pairwise_weight: [N, HxW, HxW] 85 | pairwise_weight = torch.matmul(theta_x, phi_x) 86 | pairwise_weight /= pairwise_weight.shape[-1] 87 | return pairwise_weight 88 | 89 | def forward(self, x): 90 | n, _, h, w = x.shape 91 | 92 | # g_x: [N, HxW, C] 93 | g_x = self.g(x).view(n, self.inter_channels, -1) 94 | g_x = g_x.permute(0, 2, 1) 95 | 96 | # theta_x: [N, HxW, C] 97 | theta_x = self.theta(x).view(n, self.inter_channels, -1) 98 | theta_x = theta_x.permute(0, 2, 1) 99 | 100 | # phi_x: [N, C, HxW] 101 | phi_x = self.phi(x).view(n, self.inter_channels, -1) 102 | 103 | pairwise_func = getattr(self, self.mode) 104 | # pairwise_weight: [N, HxW, HxW] 105 | pairwise_weight = pairwise_func(theta_x, phi_x) 106 | 107 | # y: [N, HxW, C] 108 | y = torch.matmul(pairwise_weight, g_x) 109 | # y: [N, C, H, W] 110 | y = y.permute(0, 2, 1).reshape(n, self.inter_channels, h, w) 111 | 112 | output = x + self.conv_out(y) 113 | 114 | return output 115 | --------------------------------------------------------------------------------