├── README.md ├── augseg ├── __init__.py ├── dataset │ ├── __init__.py │ ├── augs_ALIA.py │ ├── augs_TIBA.py │ ├── base.py │ ├── builder.py │ ├── cityscapes.py │ └── pascal_voc.py ├── models │ ├── __init__.py │ ├── decoder.py │ ├── model_helper.py │ └── resnet.py └── utils │ ├── __init__.py │ ├── dist_helper.py │ ├── loss_helper.py │ ├── lr_helper.py │ └── utils.py ├── data └── splitsall.tar.gz ├── docs ├── Augseg-diagram.png ├── augs-cutmix.png └── augs-intensity.png ├── exps ├── zrun_citys │ ├── citys_semi744 │ │ └── config_semi.yaml │ └── r50_citys_semi744 │ │ └── config_semi.yaml ├── zrun_vocs │ ├── r50_voc_semi662 │ │ └── config_semi.yaml │ └── voc_semi_fine92_r50 │ │ └── config_semi.yaml └── zrun_vocs_u2pl │ ├── r50_voc_semi662 │ └── config_semi.yaml │ └── voc_semi662 │ └── config_semi.yaml ├── requirements.txt ├── scripts ├── bashtorch ├── zsing_run_citys.sh └── zsing_run_voc.sh ├── single_run.sh ├── train_semi.py └── training-logs ├── 1 Sup-Voc ├── blender-u2pl │ ├── R101-1323 │ │ └── seg_2022-09-24_10_58_06.log │ ├── R101-2646 │ │ └── seg_2022-09-24_21_10_14.log │ ├── R101-662 │ │ └── seg_2022-09-24_18_50_02.log │ ├── R50-1323 │ │ └── seg_2022-09-24_10_51_42.log │ ├── R50-2646 │ │ └── seg_2022-09-24_13_13_23.log │ └── R50-662 │ │ └── seg_2022-09-23_22_55_28.log ├── blender │ ├── r101-1323 │ │ └── seg_2022-09-24_21_07_36.log │ ├── r101-2646 │ │ └── seg_2022-09-24_23_51_05.log │ ├── r101-662 │ │ └── seg_2022-09-24_10_56_07.log │ ├── r50-1323 │ │ └── seg_2022-09-24_12_20_19.log │ ├── r50-2646 │ │ └── seg_2022-09-24_14_22_47.log │ └── r50-662 │ │ └── seg_2022-09-24_10_45_00.log └── fine │ ├── R101-1464 │ └── seg_2022-09-25_16_59_09.log │ ├── R101-183 │ └── seg_2022-09-25_17_24_49.log │ ├── R101-366 │ └── seg_2022-09-25_18_46_02.log │ ├── R101-732 │ └── seg_2022-09-25_19_56_21.log │ ├── R101-92 │ └── seg_2022-09-25_15_53_41.log │ ├── R50-1464 │ └── seg_2022-09-25_00_02_47.log │ ├── R50-183 │ └── seg_2022-09-25_10_35_51.log │ ├── R50-366 │ └── seg_2022-09-25_12_25_53.log │ ├── R50-732 │ └── seg_2022-09-25_13_38_56.log │ └── R50-92 │ └── seg_2022-09-25_09_42_12.log ├── 2 Sup-Citys ├── r101-1488 │ └── seg_2022-09-26_12_47_19.log ├── r101-186 │ └── seg_2022-09-24_19_05_36.log ├── r101-372 │ └── seg_2022-09-25_09_50_10.log ├── r101-744 │ └── seg_2022-09-26_19_25_39.log ├── r50-1488 │ └── seg_2022-09-25_12_18_19.log ├── r50-186 │ └── seg_2022-09-23_22_52_31.log ├── r50-372 │ └── seg_2022-09-24_11_06_55.log └── r50-744 │ └── seg_2022-09-24_22_04_08.log ├── 3 Semi-Voc-fine ├── 92 │ ├── config_semi.yaml │ └── seg_2022-11-05_18:19:47.log ├── 183 │ ├── config_semi.yaml │ └── seg_2022-11-05_18:17:48.log ├── 366 │ ├── config_semi.yaml │ └── seg_2022-11-05_17_56_47.log ├── 732 │ ├── config_semi.yaml │ └── seg_2022-11-05_18:16:33.log └── 1464 │ ├── config_semi.yaml │ └── seg_2022-11-05_18:14:50.log ├── 4 Semi-Voc-blender ├── r101-1323 │ ├── config_semi.yaml │ └── seg_2022-11-04_02_44_22.log ├── r101-2646 │ ├── config_semi.yaml │ └── seg_2022-11-04_23_43_51.log ├── r101-662 │ ├── config_semi.yaml │ └── seg_2022-11-04_23_45_28.log ├── r50-1323 │ ├── config_semi.yaml │ └── seg_2022-11-03_12_42_19.log ├── r50-2646 │ ├── config_semi.yaml │ └── seg_2022-11-03_13_58_37.log └── r50-662 │ ├── config_semi.yaml │ └── seg_2022-11-02_20_35_40.log ├── 5 Semi-Voc-blender-split-u2pl ├── r101-1323 │ ├── config_semi.yaml │ └── seg_2022-11-04_02_19_35.log ├── r101-2646 │ ├── config_semi.yaml │ └── seg_2022-11-04_15_38_27.log ├── r101-662 │ ├── config_semi.yaml │ └── seg_2022-11-04_02_23_01.log ├── r50-1323 │ ├── config_semi.yaml │ └── seg_2022-11-04_02_15_18.log ├── r50-2646 │ ├── config_semi.yaml │ └── seg_2022-11-04_02_15_17.log └── r50-662 │ ├── config_semi.yaml │ └── seg_2022-11-04_02_23_48.log └── 6 Semi-citys ├── R101-1488 ├── config_semi.yaml └── seg_2022-11-08_23_12_47.log ├── R101-186 ├── config_semi.yaml └── seg_2022-11-06_15_37_41.log ├── R101-372 ├── config_semi.yaml └── seg_2022-11-06_22_15_34.log ├── R101-744 ├── config_semi.yaml └── seg_2022-11-06_22:34:45.log ├── R50-1488 ├── config_semi.yaml └── seg_2022-11-03_14_17_21.log ├── R50-186 ├── config_semi.yaml └── seg_2022-11-03_14_38_38.log ├── R50-372 ├── config_semi.yaml └── seg_2022-11-06_13_28_34.log └── R50-744 ├── config_semi.yaml └── seg_2022-11-05_18_28_53.log /README.md: -------------------------------------------------------------------------------- 1 | # AugSeg 2 | 3 | > "Augmentation Matters: A Simple-yet-Effective Approach to Semi-supervised Semantic Segmentation". 4 | 5 | 6 | 7 | ### Introduction 8 | 9 | 10 | - Recent studies on semi-supervised semantic segmentation (SSS) have seen fast progress. Despite their promising performance, current state-of-the-art methods tend to increasingly complex designs at the cost of introducing more network components and additional training procedures. 11 | 12 | - Differently, in this work, we follow a standard teacher-student framework and propose AugSeg, a simple and clean approach that focuses mainly on data perturbations to boost the SSS performance. We argue that various data augmentations should be adjusted to better adapt to the semi-supervised scenarios instead of directly applying these techniques from supervised learning. Specifically, we adopt a simplified intensity-based augmentation that selects a random number of data transformations with uniformly sampling distortion strengths from a continuous space. Based on the estimated confidence of the model on different unlabeled samples, we also randomly inject labelled information to augment the unlabeled samples in an adaptive manner. 13 | 14 | - Without bells and whistles, our simple AugSeg can readily achieve new state-of-the-art performance on SSS benchmarks under different partition protocols. 15 | 16 | 17 | 18 | 19 | ### Diagram 20 | 21 | ![](./docs/Augseg-diagram.png) 22 | 23 | > Without any complicated designs, AugSeg readily obtains new SOTA performance on popular SSS benchmarks under different partition protocols. We hope our AugSeg can inspir future studies, and serve as a strong baseline for SSS. 24 | 25 | 26 | 27 | ### Performance 28 | 29 | Labeled images are sampled from the **original high-quality** training set. Results are obtained by DeepLabv3+ based on ResNet-101 with training size 512. 30 | 31 | | Method | 1/115 (92)| 1/57 (183)| 1/28 (366)| 1/14 (732)| 1/7 (1464) | 32 | | :-------------------------: | :-------: | :-------: | :-------: | :-------: | :---------: | 33 | | SupOnly | 43.92 | 59.10| 65.88 | 70.87 | 74.97 | 34 | | ST++ | 65.2 | 71.0 | 74.6 | 77.3 | 79.1 | 35 | | PS-MT | 65.80 | 69.58 | 76.57 | 78.42 | 80.01 | 36 | | U2PL | 67.98 | 69.15 | 73.66 | 76.16 | 79.49 | 37 | | **AugSeg** | **71.09** | **75.45** | **78.80** | **80.33** | **81.36** | 38 | 39 | 40 | Results are obtained by DeepLabv3+ based on ResNet-50/101. We reproduce U2PL results on ResNet-50. 41 | 42 | | R50 | 1/16 | 1/8 | 1/4 | 1/2 | R101 | 1/16 | 1/8 | 1/4 | 1/2 | 43 | | :-------------------------: | :-------: | :-------: | :-------: | :-------: | :---------: | :---------: | :---------: | --------------------------- | --------------------------- | 44 | | SupOnly | 63.34 | 68.73 | 74.14 | 76.62 | SupOnly | 64.77 | 71.64 | 75.24 | 78.03 | 45 | | U2PL | 69.03 | 73.02 | 76.31 | 78.64 | U2PL | 70.30 | 74.37 | 76.47 | 79.05 | 46 | | PS-MT | - | 75.76 | 76.92 | 77.64 | PS-MT | - | 76.89 | 77.60 | 79.09 | 47 | | **AugSeg** | **73.73** | **76.49** | **78.76** | **79.33** | **AugSeg** | **75.22** | **77.82** | **79.56** | **80.43** | 48 | 49 | > All the training logs of AugSeg and our reproduced SupOnly baselines are included under the directory of [training-logs](./training-logs) 50 | 51 | 52 | 53 | ## Running AugSeg 54 | 55 | ### Prepare datasets 56 | 57 | Please download the Pascal and Cityscapes, and set up the path to them properly in the configuration files. 58 | 59 | - Pascal: [JPEGImages](http://host.robots.ox.ac.uk/pascal/VOC/voc2012/VOCtrainval_11-May-2012.tar) | [SegmentationClass](https://drive.google.com/file/d/1ikrDlsai5QSf2GiSUR3f8PZUzyTubcuF/view?usp=sharing) 60 | - Cityscapes: [leftImg8bit](https://www.cityscapes-dataset.com/file-handling/?packageID=3) | [gtFine](https://drive.google.com/file/d/1E_27g9tuHm6baBqcA7jct_jqcGA89QPm/view?usp=sharing) 61 | 62 | - Splitall: included. 63 | 64 | Here is our adopted way， 65 | ``` 66 | ├── ./data 67 | ├── splitsall 68 | ├── cityscapes 69 | ├── pascal 70 | └── pascal_u2pl 71 | ├── VOC2012 72 | ├── JPEGImages 73 | ├── SegmentationClass 74 | └── SegmentationClassAug 75 | └── cityscapes 76 | ├── gtFine 77 | └── leftImg8bit 78 | ``` 79 | 80 | 81 | 82 | ### Prepare pre-trained encoder 83 | 84 | Please download the pretrained models, and set up the path to these models properly in the file of `config_xxx.yaml` . 85 | 86 | ~~[ResNet-50](https://drive.google.com/file/d/1AuyE_rCUSwDpjMJHMPklXeKdZpdH1-6F/view?usp=sharing) | [ResNet-101](https://drive.google.com/file/d/13jNMOEYkqBC3CimlSSw-sWRHVZEeROmK/view?usp=sharing)~~ 87 | 88 | [ResNet-50](https://drive.google.com/file/d/1mqUrqFvTQ0k5QEotk4oiOFyP6B9dVZXS/view?usp=sharing) | [ResNet-101](https://drive.google.com/file/d/1Rx0legsMolCWENpfvE2jUScT3ogalMO8/view?usp=sharing) 89 | 90 | Here is our adopted way， 91 | 92 | ``` 93 | ├── ./pretrained 94 | ├── resnet50.pth 95 | └── resnet101.pth 96 | ``` 97 | 98 | 99 | 100 | ### Prepare running Envs 101 | 102 | Nothing special 103 | - python: 3.7.13 104 | - pytorch: 1.7.1 105 | - cuda11.0.221_cudnn8.0.5_0 106 | - torchvision: 0.8.2 107 | 108 | 109 | 110 | ### Ready to Run 111 | 112 | Basically, you are recommanded to config the experimental runnings in a ".yaml" file firstly. 113 | We include various configuration files under the directory of "exps". 114 | 115 | 116 | ```bash 117 | # 1) configure your yaml file in a running script 118 | vim ./scripts/run_abls_citys.sh 119 | 120 | # 2) run directly 121 | sh ./scripts/run_abls_citys.sh 122 | 123 | ``` 124 | 125 | ## Citation 126 | 127 | If you find these projects useful, please consider citing: 128 | 129 | ```bibtex 130 | @inproceedings{zhao2023augmentation, 131 | title={Augmentation Matters: A Simple-yet-Effective Approach to Semi-supervised Semantic Segmentation}, 132 | author={Zhao, Zhen and Yang, Lihe and Long, Sifan and Pi, Jimin and Zhou, Luping and Wang, Jingdong}, 133 | booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition}, 134 | pages={11350--11359}, 135 | year={2023} 136 | } 137 | ``` 138 | 139 | We have other relevant semi-supervised semantic segmentation projects: 140 | - [ST++](https://github.com/LiheYoung/ST-PlusPlus) 141 | - [iMas](https://github.com/ZhenZHAO/iMAS) 142 | - [Unimatch](https://github.com/LiheYoung/UniMatch) 143 | 144 | 145 | ## Acknowledgement 146 | 147 | We thank [ST++](https://github.com/LiheYoung/ST-PlusPlus), [CPS](https://github.com/charlesCXK/TorchSemiSeg), and [U2PL](https://github.com/Haochen-Wang409/U2PL), for part of their codes, processed datasets, data partitions, and pretrained models. 148 | -------------------------------------------------------------------------------- /augseg/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ZhenZHAO/AugSeg/7ead8705cf6ce6c9234f52c414ec2237ed6743cd/augseg/__init__.py -------------------------------------------------------------------------------- /augseg/dataset/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ZhenZHAO/AugSeg/7ead8705cf6ce6c9234f52c414ec2237ed6743cd/augseg/dataset/__init__.py -------------------------------------------------------------------------------- /augseg/dataset/augs_ALIA.py: -------------------------------------------------------------------------------- 1 | import numpy as np 2 | import random 3 | import torch 4 | import scipy.stats as stats 5 | 6 | 7 | # # # # # # # # # # # # # # # # # # # # # 8 | # # 0 random box 9 | # # # # # # # # # # # # # # # # # # # # # 10 | def rand_bbox(size, lam=None): 11 | # past implementation 12 | if len(size) == 4: 13 | W = size[2] 14 | H = size[3] 15 | elif len(size) == 3: 16 | W = size[1] 17 | H = size[2] 18 | else: 19 | raise Exception 20 | B = size[0] 21 | 22 | cut_rat = np.sqrt(1. - lam) 23 | cut_w = int(W * cut_rat) 24 | cut_h = int(H * cut_rat) 25 | 26 | cx = np.random.randint(size=[B, ], low=int(W/8), high=W) 27 | cy = np.random.randint(size=[B, ], low=int(H/8), high=H) 28 | 29 | bbx1 = np.clip(cx - cut_w // 2, 0, W) 30 | bby1 = np.clip(cy - cut_h // 2, 0, H) 31 | 32 | bbx2 = np.clip(cx + cut_w // 2, 0, W) 33 | bby2 = np.clip(cy + cut_h // 2, 0, H) 34 | 35 | 36 | return bbx1, bby1, bbx2, bby2 37 | 38 | 39 | # # # # # # # # # # # # # # # # # # # # # 40 | # # 1 cutmix label-adaptive 41 | # # # # # # # # # # # # # # # # # # # # # 42 | def cut_mix_label_adaptive(unlabeled_image, unlabeled_mask, unlabeled_logits, 43 | labeled_image, labeled_mask, lst_confidences): 44 | assert len(lst_confidences) == len(unlabeled_image), "Ensure the confidence is properly obtained" 45 | assert labeled_image.shape == unlabeled_image.shape, "Ensure shape match between lb and unlb" 46 | mix_unlabeled_image = unlabeled_image.clone() 47 | mix_unlabeled_target = unlabeled_mask.clone() 48 | mix_unlabeled_logits = unlabeled_logits.clone() 49 | labeled_logits = torch.ones_like(labeled_mask) 50 | 51 | # 1) get the random mixing objects 52 | u_rand_index = torch.randperm(unlabeled_image.size()[0])[:unlabeled_image.size()[0]] 53 | 54 | # 2) get box 55 | l_bbx1, l_bby1, l_bbx2, l_bby2 = rand_bbox(unlabeled_image.size(), lam=np.random.beta(8, 2)) 56 | u_bbx1, u_bby1, u_bbx2, u_bby2 = rand_bbox(unlabeled_image.size(), lam=np.random.beta(4, 4)) 57 | 58 | # 3) labeled adaptive 59 | for i in range(0, mix_unlabeled_image.shape[0]): 60 | if np.random.random() > lst_confidences[i]: 61 | mix_unlabeled_image[i, :, l_bbx1[i]:l_bbx2[i], l_bby1[i]:l_bby2[i]] = \ 62 | labeled_image[u_rand_index[i], :, l_bbx1[i]:l_bbx2[i], l_bby1[i]:l_bby2[i]] 63 | 64 | mix_unlabeled_target[i, l_bbx1[i]:l_bbx2[i], l_bby1[i]:l_bby2[i]] = \ 65 | labeled_mask[u_rand_index[i], l_bbx1[i]:l_bbx2[i], l_bby1[i]:l_bby2[i]] 66 | 67 | mix_unlabeled_logits[i, l_bbx1[i]:l_bbx2[i], l_bby1[i]:l_bby2[i]] = \ 68 | labeled_logits[u_rand_index[i], l_bbx1[i]:l_bbx2[i], l_bby1[i]:l_bby2[i]] 69 | 70 | # 4) copy and paste 71 | for i in range(0, unlabeled_image.shape[0]): 72 | unlabeled_image[i, :, u_bbx1[i]:u_bbx2[i], u_bby1[i]:u_bby2[i]] = \ 73 | mix_unlabeled_image[u_rand_index[i], :, u_bbx1[i]:u_bbx2[i], u_bby1[i]:u_bby2[i]] 74 | 75 | unlabeled_mask[i, u_bbx1[i]:u_bbx2[i], u_bby1[i]:u_bby2[i]] = \ 76 | mix_unlabeled_target[u_rand_index[i], u_bbx1[i]:u_bbx2[i], u_bby1[i]:u_bby2[i]] 77 | 78 | unlabeled_logits[i, u_bbx1[i]:u_bbx2[i], u_bby1[i]:u_bby2[i]] = \ 79 | mix_unlabeled_logits[u_rand_index[i], u_bbx1[i]:u_bbx2[i], u_bby1[i]:u_bby2[i]] 80 | 81 | del mix_unlabeled_image, mix_unlabeled_target, mix_unlabeled_logits, labeled_logits 82 | 83 | return unlabeled_image, unlabeled_mask, unlabeled_logits 84 | -------------------------------------------------------------------------------- /augseg/dataset/augs_TIBA.py: -------------------------------------------------------------------------------- 1 | import numpy as np 2 | import scipy.stats as stats 3 | from PIL import Image, ImageOps, ImageFilter, ImageEnhance 4 | import random 5 | import collections 6 | import cv2 7 | import torch 8 | from torchvision import transforms 9 | 10 | 11 | # # # # # # # # # # # # # # # # # # # # # # # # 12 | # # # 1. Augmentation for image and labels 13 | # # # # # # # # # # # # # # # # # # # # # # # # 14 | class Compose(object): 15 | def __init__(self, segtransforms): 16 | self.segtransforms = segtransforms 17 | 18 | def __call__(self, image, label): 19 | for idx, t in enumerate(self.segtransforms): 20 | if isinstance(t, strong_img_aug): 21 | image = t(image) 22 | else: 23 | image, label = t(image, label) 24 | return image, label 25 | 26 | 27 | class ToTensorAndNormalize(object): 28 | def __init__(self, mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]): 29 | assert len(mean) == len(std) 30 | assert len(mean) == 3 31 | self.normalize = transforms.Normalize(mean, std) 32 | self.to_tensor = transforms.ToTensor() 33 | 34 | def __call__(self, in_image, in_label): 35 | in_image = Image.fromarray(np.uint8(in_image)) 36 | image = self.normalize(self.to_tensor(in_image)) 37 | label = torch.from_numpy(np.array(in_label, dtype=np.int32)).long() 38 | 39 | return image, label 40 | 41 | 42 | class Resize(object): 43 | def __init__(self, base_size, ratio_range, scale=True, bigger_side_to_base_size=True): 44 | # assert isinstance(ratio_range, collections.Iterable) and len(ratio_range) == 2 45 | assert isinstance(ratio_range, collections.abc.Iterable) and len(ratio_range) == 2 # for recent python version 46 | self.base_size = base_size 47 | self.ratio_range = ratio_range 48 | self.scale = scale 49 | self.bigger_side_to_base_size = bigger_side_to_base_size 50 | 51 | def __call__(self, in_image, in_label): 52 | w, h = in_image.size 53 | 54 | if isinstance(self.base_size, int): 55 | # obtain long_side 56 | if self.scale: 57 | long_side = random.randint(int(self.base_size * self.ratio_range[0]), 58 | int(self.base_size * self.ratio_range[1])) 59 | else: 60 | long_side = self.base_size 61 | 62 | # obtain new oh, ow 63 | if self.bigger_side_to_base_size: 64 | if h > w: 65 | oh = long_side 66 | ow = int(1.0 * long_side * w / h + 0.5) 67 | else: 68 | oh = int(1.0 * long_side * h / w + 0.5) 69 | ow = long_side 70 | else: 71 | oh, ow = (long_side, int(1.0 * long_side * w / h + 0.5)) if h < w else ( 72 | int(1.0 * long_side * h / w + 0.5), long_side) 73 | 74 | image = in_image.resize((ow, oh), Image.BILINEAR) 75 | label = in_label.resize((ow, oh), Image.NEAREST) 76 | return image, label 77 | elif (isinstance(self.base_size, list) or isinstance(self.base_size, tuple)) and len(self.base_size) == 2: 78 | if self.scale: 79 | # scale = random.random() * 1.5 + 0.5 # Scaling between [0.5, 2] 80 | scale = self.ratio_range[0] + random.random() * (self.ratio_range[1] - self.ratio_range[0]) 81 | # print("="*100, h, self.base_size[0]) 82 | # print("="*100, w, self.base_size[1]) 83 | oh, ow = int(self.base_size[0] * scale), int(self.base_size[1] * scale) 84 | else: 85 | oh, ow = self.base_size 86 | image = in_image.resize((ow, oh), Image.BILINEAR) 87 | label = in_label.resize((ow, oh), Image.NEAREST) 88 | # print("="*100, in_image.size, image.size) 89 | return image, label 90 | 91 | else: 92 | raise ValueError 93 | 94 | 95 | class Crop(object): 96 | def __init__(self, crop_size, crop_type="rand", mean=[0.485, 0.456, 0.406], ignore_value=255): 97 | if (isinstance(crop_size, list) or isinstance(crop_size, tuple)) and len(crop_size) == 2: 98 | self.crop_h, self.crop_w = crop_size 99 | elif isinstance(crop_size, int): 100 | self.crop_h, self.crop_w = crop_size, crop_size 101 | else: 102 | raise ValueError 103 | 104 | self.crop_type = crop_type 105 | self.image_padding = (np.array(mean) * 255.).tolist() 106 | self.ignore_value = ignore_value 107 | 108 | def __call__(self, in_image, in_label): 109 | # Padding to return the correct crop size 110 | w, h = in_image.size 111 | pad_h = max(self.crop_h - h, 0) 112 | pad_w = max(self.crop_w - w, 0) 113 | pad_kwargs = { 114 | "top": 0, 115 | "bottom": pad_h, 116 | "left": 0, 117 | "right": pad_w, 118 | "borderType": cv2.BORDER_CONSTANT, 119 | } 120 | if pad_h > 0 or pad_w > 0: 121 | image = cv2.copyMakeBorder(np.asarray(in_image, dtype=np.float32), 122 | value=self.image_padding, **pad_kwargs) 123 | label = cv2.copyMakeBorder(np.asarray(in_label, dtype=np.int32), 124 | value=self.ignore_value, **pad_kwargs) 125 | image = Image.fromarray(np.uint8(image)) 126 | label = Image.fromarray(np.uint8(label)) 127 | else: 128 | image = in_image 129 | label = in_label 130 | 131 | # cropping 132 | w, h = image.size 133 | if self.crop_type == "rand": 134 | x = random.randint(0, w - self.crop_w) 135 | y = random.randint(0, h - self.crop_h) 136 | else: 137 | x = (w - self.crop_w) // 2 138 | y = (h - self.crop_h) // 2 139 | image = image.crop((x, y, x + self.crop_w, y + self.crop_h)) 140 | label = label.crop((x, y, x + self.crop_w, y + self.crop_h)) 141 | return image, label 142 | 143 | 144 | class RandomFlip(object): 145 | def __init__(self, prob=0.5, flag_hflip=True,): 146 | self.prob = prob 147 | if flag_hflip: 148 | self.type_flip = Image.FLIP_LEFT_RIGHT 149 | else: 150 | self.type_flip = Image.FLIP_TOP_BOTTOM 151 | 152 | def __call__(self, in_image, in_label): 153 | if random.random() < self.prob: 154 | in_image = in_image.transpose(self.type_flip) 155 | in_label = in_label.transpose(self.type_flip) 156 | return in_image, in_label 157 | 158 | 159 | # # # # # # # # # # # # # # # # # # # # # # # # 160 | # # # 2. Strong Augmentation for image only 161 | # # # # # # # # # # # # # # # # # # # # # # # # 162 | 163 | def img_aug_identity(img, scale=None): 164 | return img 165 | 166 | 167 | def img_aug_autocontrast(img, scale=None): 168 | return ImageOps.autocontrast(img) 169 | 170 | 171 | def img_aug_equalize(img, scale=None): 172 | return ImageOps.equalize(img) 173 | 174 | 175 | def img_aug_invert(img, scale=None): 176 | return ImageOps.invert(img) 177 | 178 | 179 | def img_aug_blur(img, scale=[0.1, 2.0]): 180 | assert scale[0] < scale[1] 181 | sigma = np.random.uniform(scale[0], scale[1]) 182 | # print(f"sigma:{sigma}") 183 | return img.filter(ImageFilter.GaussianBlur(radius=sigma)) 184 | 185 | 186 | def img_aug_contrast(img, scale=[0.05, 0.95]): 187 | min_v, max_v = min(scale), max(scale) 188 | v = float(max_v - min_v)*random.random() 189 | v = max_v - v 190 | # # print(f"final:{v}") 191 | # v = np.random.uniform(scale[0], scale[1]) 192 | return ImageEnhance.Contrast(img).enhance(v) 193 | 194 | 195 | def img_aug_brightness(img, scale=[0.05, 0.95]): 196 | min_v, max_v = min(scale), max(scale) 197 | v = float(max_v - min_v)*random.random() 198 | v = max_v - v 199 | # print(f"final:{v}") 200 | return ImageEnhance.Brightness(img).enhance(v) 201 | 202 | 203 | def img_aug_color(img, scale=[0.05, 0.95]): 204 | min_v, max_v = min(scale), max(scale) 205 | v = float(max_v - min_v)*random.random() 206 | v = max_v - v 207 | # print(f"final:{v}") 208 | return ImageEnhance.Color(img).enhance(v) 209 | 210 | 211 | def img_aug_sharpness(img, scale=[0.05, 0.95]): 212 | min_v, max_v = min(scale), max(scale) 213 | v = float(max_v - min_v)*random.random() 214 | v = max_v - v 215 | # print(f"final:{v}") 216 | return ImageEnhance.Sharpness(img).enhance(v) 217 | 218 | 219 | def img_aug_hue(img, scale=[0, 0.5]): 220 | min_v, max_v = min(scale), max(scale) 221 | v = float(max_v - min_v)*random.random() 222 | v += min_v 223 | if np.random.random() < 0.5: 224 | hue_factor = -v 225 | else: 226 | hue_factor = v 227 | # print(f"Final-V:{hue_factor}") 228 | input_mode = img.mode 229 | if input_mode in {"L", "1", "I", "F"}: 230 | return img 231 | h, s, v = img.convert("HSV").split() 232 | np_h = np.array(h, dtype=np.uint8) 233 | # uint8 addition take cares of rotation across boundaries 234 | with np.errstate(over="ignore"): 235 | np_h += np.uint8(hue_factor * 255) 236 | h = Image.fromarray(np_h, "L") 237 | img = Image.merge("HSV", (h, s, v)).convert(input_mode) 238 | return img 239 | 240 | 241 | def img_aug_posterize(img, scale=[4, 8]): 242 | min_v, max_v = min(scale), max(scale) 243 | v = float(max_v - min_v)*random.random() 244 | # print(min_v, max_v, v) 245 | v = int(np.ceil(v)) 246 | v = max(1, v) 247 | v = max_v - v 248 | # print(f"final:{v}") 249 | return ImageOps.posterize(img, v) 250 | 251 | 252 | def img_aug_solarize(img, scale=[1, 256]): 253 | min_v, max_v = min(scale), max(scale) 254 | v = float(max_v - min_v)*random.random() 255 | # print(min_v, max_v, v) 256 | v = int(np.ceil(v)) 257 | v = max(1, v) 258 | v = max_v - v 259 | # print(f"final:{v}") 260 | return ImageOps.solarize(img, v) 261 | 262 | def get_augment_list(flag_using_wide=False): 263 | if flag_using_wide: 264 | l = [ 265 | (img_aug_identity, None), 266 | (img_aug_autocontrast, None), 267 | (img_aug_equalize, None), 268 | (img_aug_blur, [0.1, 2.0]), 269 | (img_aug_contrast, [0.1, 1.8]), 270 | (img_aug_brightness, [0.1, 1.8]), 271 | (img_aug_color, [0.1, 1.8]), 272 | (img_aug_sharpness, [0.1, 1.8]), 273 | (img_aug_posterize, [2, 8]), 274 | (img_aug_solarize, [1, 256]), 275 | (img_aug_hue, [0, 0.5]) 276 | ] 277 | else: 278 | l = [ 279 | (img_aug_identity, None), 280 | (img_aug_autocontrast, None), 281 | (img_aug_equalize, None), 282 | (img_aug_blur, [0.1, 2.0]), 283 | (img_aug_contrast, [0.05, 0.95]), 284 | (img_aug_brightness, [0.05, 0.95]), 285 | (img_aug_color, [0.05, 0.95]), 286 | (img_aug_sharpness, [0.05, 0.95]), 287 | (img_aug_posterize, [4, 8]), 288 | (img_aug_solarize, [1, 256]), 289 | (img_aug_hue, [0, 0.5]) 290 | ] 291 | return l 292 | 293 | 294 | class strong_img_aug: 295 | def __init__(self, num_augs, flag_using_random_num=False): 296 | assert 1<= num_augs <= 11 297 | self.n = num_augs 298 | self.augment_list = get_augment_list(flag_using_wide=False) 299 | self.flag_using_random_num = flag_using_random_num 300 | 301 | def __call__(self, img): 302 | if self.flag_using_random_num: 303 | max_num = np.random.randint(1, high=self.n + 1) 304 | else: 305 | max_num =self.n 306 | ops = random.choices(self.augment_list, k=max_num) 307 | for op, scales in ops: 308 | # print("="*20, str(op)) 309 | img = op(img, scales) 310 | return img 311 | -------------------------------------------------------------------------------- /augseg/dataset/base.py: -------------------------------------------------------------------------------- 1 | import logging 2 | 3 | from PIL import Image 4 | from torch.utils.data import Dataset 5 | 6 | 7 | class BaseDataset(Dataset): 8 | def __init__(self, d_list, **kwargs): 9 | # parse the input list 10 | self.parse_input_list(d_list, **kwargs) 11 | 12 | def parse_input_list(self, d_list, max_sample=-1, start_idx=-1, end_idx=-1): 13 | logger = logging.getLogger("global") 14 | assert isinstance(d_list, str) 15 | if "cityscapes" in d_list: 16 | self.list_sample = [ 17 | [ 18 | line.strip(), 19 | "gtFine/" + line.strip()[12:-15] + "gtFine_labelTrainIds.png", 20 | ] 21 | for line in open(d_list, "r") 22 | ] 23 | elif "pascal" in d_list or "VOC" in d_list: 24 | self.list_sample = [ 25 | [ 26 | "JPEGImages/{}.jpg".format(line.strip()), 27 | "SegmentationClassAug/{}.png".format(line.strip()), 28 | ] 29 | for line in open(d_list, "r") 30 | ] 31 | else: 32 | raise "unknown dataset!" 33 | 34 | if max_sample > 0: 35 | self.list_sample = self.list_sample[0:max_sample] 36 | if start_idx >= 0 and end_idx >= 0: 37 | self.list_sample = self.list_sample[start_idx:end_idx] 38 | 39 | self.num_sample = len(self.list_sample) 40 | assert self.num_sample > 0 41 | logger.info("# samples: {}".format(self.num_sample)) 42 | 43 | def img_loader(self, path, mode): 44 | with open(path, "rb") as f: 45 | img = Image.open(f) 46 | return img.convert(mode) 47 | 48 | def __len__(self): 49 | return self.num_sample 50 | -------------------------------------------------------------------------------- /augseg/dataset/builder.py: -------------------------------------------------------------------------------- 1 | import logging 2 | 3 | from .cityscapes import build_city_semi_loader, build_cityloader 4 | from .pascal_voc import build_voc_semi_loader, build_vocloader 5 | 6 | logger = logging.getLogger("global") 7 | 8 | 9 | def get_loader(cfg, seed=0): 10 | cfg_dataset = cfg["dataset"] 11 | 12 | if cfg_dataset["type"] == "cityscapes_semi": 13 | train_loader_sup, train_loader_unsup = build_city_semi_loader( 14 | "train", cfg, seed=seed 15 | ) 16 | val_loader = build_cityloader("val", cfg) 17 | logger.info("Get loader Done...") 18 | return train_loader_sup, train_loader_unsup, val_loader 19 | 20 | elif cfg_dataset["type"] == "cityscapes": 21 | train_loader_sup = build_cityloader("train", cfg, seed=seed) 22 | val_loader = build_cityloader("val", cfg) 23 | logger.info("Get loader Done...") 24 | return train_loader_sup, val_loader 25 | 26 | elif cfg_dataset["type"] == "pascal_semi": 27 | train_loader_sup, train_loader_unsup = build_voc_semi_loader( 28 | "train", cfg, seed=seed 29 | ) 30 | val_loader = build_vocloader("val", cfg) 31 | logger.info("Get loader Done...") 32 | return train_loader_sup, train_loader_unsup, val_loader 33 | 34 | elif cfg_dataset["type"] == "pascal": 35 | train_loader_sup = build_vocloader("train", cfg, seed=seed) 36 | val_loader = build_vocloader("val", cfg) 37 | logger.info("Get loader Done...") 38 | return train_loader_sup, val_loader 39 | 40 | else: 41 | raise NotImplementedError( 42 | "dataset type {} is not supported".format(cfg_dataset) 43 | ) 44 | -------------------------------------------------------------------------------- /augseg/dataset/cityscapes.py: -------------------------------------------------------------------------------- 1 | import copy 2 | import math 3 | import os 4 | import os.path 5 | import random 6 | 7 | import numpy as np 8 | import torch 9 | import torch.distributed as dist 10 | from torch.utils.data import DataLoader 11 | from torch.utils.data.distributed import DistributedSampler 12 | from torchvision import transforms 13 | 14 | # from . import augmentations as img_trsform 15 | from . import augs_TIBA as img_trsform 16 | from .base import BaseDataset 17 | 18 | # https://pytorch.org/docs/stable/notes/randomness.html 19 | def seed_worker(worker_id): 20 | cur_seed = np.random.get_state()[1][0] 21 | cur_seed += worker_id 22 | np.random.seed(cur_seed) 23 | random.seed(cur_seed) 24 | 25 | 26 | class city_dset(BaseDataset): 27 | def __init__(self, data_root, data_list, trs_form, trs_form_strong=None, 28 | seed=0, n_sup=2975, split="val", flag_semi=False, 29 | mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225] 30 | ): 31 | super(city_dset, self).__init__(data_list) 32 | self.data_root = data_root 33 | self.transform_weak = trs_form 34 | self.transform_strong = trs_form_strong 35 | self.flag_semi = flag_semi 36 | self.split = split 37 | # random.seed(seed) 38 | 39 | self.trf_normalize = self._get_to_tensor_and_normalize(mean, std) 40 | 41 | # oversamplying labeled data for semi-supervised training 42 | if len(self.list_sample) >= n_sup and split == "train": 43 | self.list_sample_new = random.sample(self.list_sample, n_sup) 44 | elif len(self.list_sample) < n_sup and split == "train": 45 | num_repeat = math.ceil(n_sup / len(self.list_sample)) 46 | self.list_sample = self.list_sample * num_repeat 47 | 48 | self.list_sample_new = random.sample(self.list_sample, n_sup) 49 | else: 50 | self.list_sample_new = self.list_sample 51 | 52 | @staticmethod 53 | def _get_to_tensor_and_normalize(mean, std): 54 | return img_trsform.ToTensorAndNormalize(mean, std) 55 | 56 | def __getitem__(self, index): 57 | # load image and its label 58 | image_path = os.path.join(self.data_root, self.list_sample_new[index][0]) 59 | label_path = os.path.join(self.data_root, self.list_sample_new[index][1]) 60 | image = self.img_loader(image_path, "RGB") 61 | label = self.img_loader(label_path, "L") 62 | 63 | if self.transform_strong is None: 64 | image, label = self.transform_weak(image, label) 65 | # print(image.shape, label.shape) 66 | image, label = self.trf_normalize(image, label) 67 | if not self.flag_semi: 68 | return index, image, label 69 | else: 70 | return index, image, image.clone(), label 71 | else: 72 | # apply augmentation 73 | image_weak, label = self.transform_weak(image, label) 74 | image_strong = self.transform_strong(image_weak) 75 | # print("="*100) 76 | # print(index, image_weak.size, image_strong.size, label.size) 77 | # print("="*100) 78 | 79 | image_weak, label = self.trf_normalize(image_weak, label) 80 | image_strong, _ = self.trf_normalize(image_strong, label) 81 | # print(index, image_weak.shape, image_strong.shape,label.shape) 82 | 83 | return index, image_weak, image_strong, label 84 | 85 | # image, label = self.transform(image, label) 86 | # return image[0], label[0, 0].long() 87 | 88 | def __len__(self): 89 | return len(self.list_sample_new) 90 | 91 | 92 | def build_additional_strong_transform(cfg): 93 | assert cfg.get("strong_aug", False) != False 94 | strong_aug_nums = cfg["strong_aug"].get("num_augs", 2) 95 | flag_use_rand_num = cfg["strong_aug"].get("flag_use_random_num_sampling", True) 96 | strong_img_aug = img_trsform.strong_img_aug(strong_aug_nums, 97 | flag_using_random_num=flag_use_rand_num) 98 | return strong_img_aug 99 | 100 | 101 | def build_basic_transfrom(cfg, split="val", mean=[0.485, 0.456, 0.406]): 102 | ignore_label = cfg["ignore_label"] 103 | trs_form = [] 104 | if split != "val": 105 | if cfg.get("rand_resize", False): 106 | trs_form.append(img_trsform.Resize(cfg.get("resize_base_size", [1024, 2048]), cfg["rand_resize"])) 107 | 108 | if cfg.get("flip", False): 109 | trs_form.append(img_trsform.RandomFlip(prob=0.5, flag_hflip=True)) 110 | 111 | # crop also sometime for validating 112 | if cfg.get("crop", False): 113 | crop_size, crop_type = cfg["crop"]["size"], cfg["crop"]["type"] 114 | trs_form.append(img_trsform.Crop(crop_size, crop_type=crop_type, mean=mean, ignore_value=ignore_label)) 115 | 116 | return img_trsform.Compose(trs_form) 117 | 118 | 119 | def build_cityloader(split, all_cfg, seed=0): 120 | # extract augs config from "train"/"val" into the higher level. 121 | cfg_dset = all_cfg["dataset"] 122 | cfg = copy.deepcopy(cfg_dset) 123 | cfg.update(cfg.get(split, {})) 124 | 125 | # set up workers and batchsize 126 | workers = cfg.get("workers", 2) 127 | batch_size = cfg.get("batch_size", 1) 128 | n_sup = cfg.get("n_sup", 2975) 129 | 130 | # build transform 131 | mean, std = cfg["mean"], cfg["std"] 132 | trs_form = build_basic_transfrom(cfg, split=split, mean=mean) 133 | 134 | # create dataset 135 | dset = city_dset(cfg["data_root"], cfg["data_list"], trs_form, None, 136 | seed, n_sup, mean=mean, std=std) 137 | 138 | # build sampler 139 | sample = DistributedSampler(dset) 140 | loader = DataLoader( 141 | dset, 142 | batch_size=batch_size, 143 | num_workers=workers, 144 | sampler=sample, 145 | shuffle=False, 146 | pin_memory=False, 147 | worker_init_fn=seed_worker, 148 | ) 149 | return loader 150 | 151 | 152 | def build_city_semi_loader(split, all_cfg, seed=0): 153 | split = "train" 154 | # extract augs config from "train" into the higher level. 155 | cfg_dset = all_cfg["dataset"] 156 | cfg = copy.deepcopy(cfg_dset) 157 | cfg.update(cfg.get(split, {})) 158 | 159 | # set up workers and batchsize 160 | workers = cfg.get("workers", 2) 161 | batch_size = cfg.get("batch_size", 2) 162 | n_sup = 2975 - cfg.get("n_sup", 2975) 163 | 164 | # build transform 165 | mean, std = cfg["mean"], cfg["std"] 166 | trs_form_weak = build_basic_transfrom(cfg, split=split, mean=mean) 167 | if cfg.get("strong_aug", False): 168 | trs_form_strong = build_additional_strong_transform(cfg) 169 | else: 170 | trs_form_strong = None 171 | 172 | dset = city_dset(cfg["data_root"], cfg["data_list"], trs_form_weak, None, 173 | seed, n_sup, split=split, mean=mean, std=std) 174 | sample_sup = DistributedSampler(dset) 175 | 176 | data_list_unsup = cfg["data_list"].replace("labeled.txt", "unlabeled.txt") 177 | dset_unsup = city_dset(cfg["data_root"], data_list_unsup, trs_form_weak, trs_form_strong, 178 | seed, n_sup, split, 179 | flag_semi=True, 180 | mean=mean, std=std) 181 | sample_unsup = DistributedSampler(dset_unsup) 182 | 183 | # create dataloader 184 | loader_sup = DataLoader( 185 | dset, 186 | batch_size=batch_size, 187 | num_workers=workers, 188 | sampler=sample_sup, 189 | shuffle=False, 190 | pin_memory=True, 191 | drop_last=True, 192 | worker_init_fn=seed_worker, 193 | ) 194 | loader_unsup = DataLoader( 195 | dset_unsup, 196 | batch_size=batch_size, 197 | num_workers=workers, 198 | sampler=sample_unsup, 199 | shuffle=False, 200 | pin_memory=True, 201 | drop_last=True, 202 | worker_init_fn=seed_worker, 203 | ) 204 | return loader_sup, loader_unsup 205 | -------------------------------------------------------------------------------- /augseg/dataset/pascal_voc.py: -------------------------------------------------------------------------------- 1 | import copy 2 | import math 3 | import os 4 | import os.path 5 | import random 6 | 7 | import numpy as np 8 | import torch 9 | from torch.utils.data import DataLoader, Dataset 10 | from torch.utils.data.distributed import DistributedSampler 11 | from torchvision import transforms 12 | 13 | from . import augs_TIBA as img_trsform 14 | from .base import BaseDataset 15 | 16 | # https://pytorch.org/docs/stable/notes/randomness.html 17 | def seed_worker(worker_id): 18 | cur_seed = np.random.get_state()[1][0] 19 | cur_seed += worker_id 20 | np.random.seed(cur_seed) 21 | random.seed(cur_seed) 22 | 23 | 24 | class voc_dset(BaseDataset): 25 | def __init__( 26 | self, data_root, data_list, trs_form, trs_form_strong=None, 27 | seed=0, n_sup=10582, split="val", flag_semi=False, 28 | mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225] 29 | ): 30 | super(voc_dset, self).__init__(data_list) 31 | self.data_root = data_root 32 | self.transform_weak = trs_form 33 | self.transform_strong = trs_form_strong 34 | self.flag_semi = flag_semi 35 | self.split = split 36 | # random.seed(seed) # set on the top level 37 | 38 | self.trf_normalize = self._get_to_tensor_and_normalize(mean, std) 39 | 40 | # oversamplying labeled data for semi-supervised training 41 | if len(self.list_sample) >= n_sup and split == "train": 42 | self.list_sample_new = random.sample(self.list_sample, n_sup) 43 | elif len(self.list_sample) < n_sup and split == "train": 44 | num_repeat = math.ceil(n_sup / len(self.list_sample)) 45 | self.list_sample = self.list_sample * num_repeat 46 | self.list_sample_new = random.sample(self.list_sample, n_sup) 47 | else: 48 | self.list_sample_new = self.list_sample 49 | 50 | # # ADD: shuffle the image list ===> not neccesary, random.sample also shuffle it. 51 | # if split == "train": 52 | # np.random.shuffle(self.list_sample) 53 | 54 | @staticmethod 55 | def _get_to_tensor_and_normalize(mean, std): 56 | return img_trsform.ToTensorAndNormalize(mean, std) 57 | 58 | def __getitem__(self, index): 59 | # load image and its label 60 | image_path = os.path.join(self.data_root, self.list_sample_new[index][0]) 61 | label_path = os.path.join(self.data_root, self.list_sample_new[index][1]) 62 | image = self.img_loader(image_path, "RGB") 63 | label = self.img_loader(label_path, "L") 64 | 65 | if self.transform_strong is None: 66 | image, label = self.transform_weak(image, label) 67 | # print(image.shape, label.shape) 68 | image, label = self.trf_normalize(image, label) 69 | if not self.flag_semi: 70 | return index, image, label 71 | else: 72 | return index, image, image.clone(), label 73 | else: 74 | # apply augmentation 75 | image_weak, label = self.transform_weak(image, label) 76 | image_strong = self.transform_strong(image_weak) 77 | # print("="*100) 78 | # print(index, image_weak.size, image_strong.size, label.size) 79 | # print("="*100) 80 | 81 | image_weak, label = self.trf_normalize(image_weak, label) 82 | image_strong, _ = self.trf_normalize(image_strong, label) 83 | # print(index, image_weak.shape, image_strong.shape,label.shape) 84 | 85 | return index, image_weak, image_strong, label 86 | 87 | def __len__(self): 88 | return len(self.list_sample_new) 89 | 90 | 91 | def build_additional_strong_transform(cfg): 92 | assert cfg.get("strong_aug", False) != False 93 | strong_aug_nums = cfg["strong_aug"].get("num_augs", 2) 94 | flag_use_rand_num = cfg["strong_aug"].get("flag_use_random_num_sampling", True) 95 | strong_img_aug = img_trsform.strong_img_aug(strong_aug_nums, 96 | flag_using_random_num=flag_use_rand_num) 97 | return strong_img_aug 98 | 99 | 100 | def build_basic_transfrom(cfg, split="val", mean=[0.485, 0.456, 0.406]): 101 | ignore_label = cfg["ignore_label"] 102 | trs_form = [] 103 | if split != "val": 104 | if cfg.get("rand_resize", False): 105 | trs_form.append(img_trsform.Resize(cfg.get("resize_base_size", 600), cfg["rand_resize"])) 106 | 107 | if cfg.get("flip", False): 108 | trs_form.append(img_trsform.RandomFlip(prob=0.5, flag_hflip=True)) 109 | 110 | # crop also sometime for validating 111 | if cfg.get("crop", False): 112 | crop_size, crop_type = cfg["crop"]["size"], cfg["crop"]["type"] 113 | trs_form.append(img_trsform.Crop(crop_size, crop_type=crop_type, mean=mean, ignore_value=ignore_label)) 114 | 115 | return img_trsform.Compose(trs_form) 116 | 117 | 118 | def build_vocloader(split, all_cfg, seed=0): 119 | # extract augs config from "train"/"val" into the higher level. 120 | cfg_dset = all_cfg["dataset"] 121 | cfg = copy.deepcopy(cfg_dset) 122 | cfg.update(cfg.get(split, {})) 123 | 124 | # set up workers and batchsize 125 | workers = cfg.get("workers", 2) 126 | batch_size = cfg.get("batch_size", 1) 127 | n_sup = cfg.get("n_sup", 10582) 128 | 129 | # build transform 130 | mean, std = cfg["mean"], cfg["std"] 131 | trs_form = build_basic_transfrom(cfg, split=split, mean=mean) 132 | 133 | # create dataset 134 | dset = voc_dset(cfg["data_root"], cfg["data_list"], trs_form, None, 135 | seed, n_sup, mean=mean, std=std) 136 | 137 | # build sampler 138 | sample = DistributedSampler(dset) 139 | loader = DataLoader( 140 | dset, 141 | batch_size=batch_size, 142 | num_workers=workers, 143 | sampler=sample, 144 | shuffle=False, 145 | pin_memory=False, 146 | worker_init_fn=seed_worker, 147 | ) 148 | return loader 149 | 150 | 151 | def build_voc_semi_loader(split, all_cfg, seed=0): 152 | split = "train" 153 | # extract augs config from "train" into the higher level. 154 | cfg_dset = all_cfg["dataset"] 155 | cfg = copy.deepcopy(cfg_dset) 156 | cfg.update(cfg.get(split, {})) 157 | 158 | # set up workers and batchsize 159 | workers = cfg.get("workers", 2) 160 | batch_size = cfg.get("batch_size", 2) 161 | n_sup = 10582 - cfg.get("n_sup", 10582) # oversample labeled data to the amount of unlabeled data 162 | 163 | # build transform 164 | mean, std = cfg["mean"], cfg["std"] 165 | trs_form_weak = build_basic_transfrom(cfg, split=split, mean=mean) 166 | if cfg.get("strong_aug", False): 167 | trs_form_strong = build_additional_strong_transform(cfg) 168 | else: 169 | trs_form_strong = None 170 | 171 | dset = voc_dset(cfg["data_root"], cfg["data_list"], trs_form_weak, None, 172 | seed, n_sup, split=split, mean=mean, std=std) 173 | sample_sup = DistributedSampler(dset) 174 | 175 | data_list_unsup = cfg["data_list"].replace("labeled.txt", "unlabeled.txt") 176 | dset_unsup = voc_dset(cfg["data_root"], data_list_unsup, trs_form_weak, trs_form_strong, 177 | seed, n_sup, split, 178 | flag_semi=True, 179 | mean=mean, std=std) 180 | sample_unsup = DistributedSampler(dset_unsup) 181 | 182 | # create dataloader 183 | loader_sup = DataLoader( 184 | dset, 185 | batch_size=batch_size, 186 | num_workers=workers, 187 | sampler=sample_sup, 188 | shuffle=False, 189 | pin_memory=True, 190 | drop_last=True, 191 | worker_init_fn=seed_worker, 192 | ) 193 | loader_unsup = DataLoader( 194 | dset_unsup, 195 | batch_size=batch_size, 196 | num_workers=workers, 197 | sampler=sample_unsup, 198 | shuffle=False, 199 | pin_memory=True, 200 | drop_last=True, 201 | worker_init_fn=seed_worker, 202 | ) 203 | return loader_sup, loader_unsup 204 | -------------------------------------------------------------------------------- /augseg/models/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ZhenZHAO/AugSeg/7ead8705cf6ce6c9234f52c414ec2237ed6743cd/augseg/models/__init__.py -------------------------------------------------------------------------------- /augseg/models/decoder.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import torch.nn as nn 3 | from torch.nn import functional as F 4 | 5 | def get_syncbn(): 6 | # return nn.BatchNorm2d 7 | return nn.SyncBatchNorm 8 | 9 | 10 | class dec_deeplabv3_plus(nn.Module): 11 | def __init__( 12 | self, 13 | in_planes, 14 | num_classes=19, 15 | inner_planes=256, 16 | sync_bn=False, 17 | dilations=(12, 24, 36), 18 | low_conv_planes=48, 19 | ): 20 | super(dec_deeplabv3_plus, self).__init__() 21 | 22 | norm_layer = get_syncbn() if sync_bn else nn.BatchNorm2d 23 | 24 | self.low_conv = nn.Sequential( 25 | nn.Conv2d(256, low_conv_planes, kernel_size=1), 26 | norm_layer(low_conv_planes), 27 | nn.ReLU(inplace=True) 28 | ) 29 | 30 | self.aspp = ASPP( 31 | in_planes, inner_planes=inner_planes, sync_bn=sync_bn, dilations=dilations 32 | ) 33 | 34 | self.head = nn.Sequential( 35 | nn.Conv2d(self.aspp.get_outplanes(), 256, 1, bias=False), 36 | norm_layer(256), 37 | nn.ReLU(inplace=True), 38 | ) 39 | 40 | self.classifier = nn.Sequential( 41 | nn.Conv2d(256+int(low_conv_planes), 256, kernel_size=3, stride=1, padding=1, bias=False), 42 | norm_layer(256), 43 | nn.ReLU(inplace=True), 44 | nn.Conv2d(256, 256, kernel_size=3, stride=1, padding=1, bias=False), 45 | norm_layer(256), 46 | nn.ReLU(inplace=True), 47 | nn.Conv2d(256, num_classes, kernel_size=1, stride=1, padding=0, bias=True), 48 | ) 49 | 50 | 51 | def forward(self, x): 52 | x1, x2, x3, x4 = x 53 | low_feat = self.low_conv(x1) 54 | h, w = low_feat.size()[-2:] 55 | 56 | aspp_out = self.aspp(x4) 57 | aspp_out = self.head(aspp_out) 58 | aspp_out = F.interpolate( 59 | aspp_out, size=(h, w), mode="bilinear", align_corners=True 60 | ) 61 | 62 | aspp_out = torch.cat((low_feat, aspp_out), dim=1) 63 | 64 | return self.classifier(aspp_out) 65 | 66 | 67 | class Aux_Module(nn.Module): 68 | def __init__(self, in_planes, num_classes=19, sync_bn=False): 69 | super(Aux_Module, self).__init__() 70 | 71 | norm_layer = get_syncbn() if sync_bn else nn.BatchNorm2d 72 | self.aux = nn.Sequential( 73 | nn.Conv2d(in_planes, 256, kernel_size=3, stride=1, padding=1, bias=False), 74 | norm_layer(256), 75 | nn.ReLU(inplace=True), 76 | nn.Conv2d(256, num_classes, kernel_size=1, stride=1, padding=0, bias=True), 77 | ) 78 | 79 | def forward(self, x): 80 | res = self.aux(x) 81 | return res 82 | 83 | 84 | class ASPP(nn.Module): 85 | """ 86 | Reference: 87 | Chen, Liang-Chieh, et al. *"Rethinking Atrous Convolution for Semantic Image Segmentation."* 88 | """ 89 | 90 | def __init__( 91 | self, in_planes, inner_planes=256, sync_bn=False, dilations=(12, 24, 36) 92 | ): 93 | super(ASPP, self).__init__() 94 | 95 | norm_layer = get_syncbn() if sync_bn else nn.BatchNorm2d 96 | self.conv1 = nn.Sequential( 97 | nn.AdaptiveAvgPool2d((1, 1)), 98 | nn.Conv2d( 99 | in_planes, 100 | inner_planes, 101 | kernel_size=1, 102 | padding=0, 103 | dilation=1, 104 | bias=False, 105 | ), 106 | norm_layer(inner_planes), 107 | nn.ReLU(inplace=True), 108 | ) 109 | self.conv2 = nn.Sequential( 110 | nn.Conv2d( 111 | in_planes, 112 | inner_planes, 113 | kernel_size=1, 114 | padding=0, 115 | dilation=1, 116 | bias=False, 117 | ), 118 | norm_layer(inner_planes), 119 | nn.ReLU(inplace=True), 120 | ) 121 | self.conv3 = nn.Sequential( 122 | nn.Conv2d( 123 | in_planes, 124 | inner_planes, 125 | kernel_size=3, 126 | padding=dilations[0], 127 | dilation=dilations[0], 128 | bias=False, 129 | ), 130 | norm_layer(inner_planes), 131 | nn.ReLU(inplace=True), 132 | ) 133 | self.conv4 = nn.Sequential( 134 | nn.Conv2d( 135 | in_planes, 136 | inner_planes, 137 | kernel_size=3, 138 | padding=dilations[1], 139 | dilation=dilations[1], 140 | bias=False, 141 | ), 142 | norm_layer(inner_planes), 143 | nn.ReLU(inplace=True), 144 | ) 145 | self.conv5 = nn.Sequential( 146 | nn.Conv2d( 147 | in_planes, 148 | inner_planes, 149 | kernel_size=3, 150 | padding=dilations[2], 151 | dilation=dilations[2], 152 | bias=False, 153 | ), 154 | norm_layer(inner_planes), 155 | nn.ReLU(inplace=True), 156 | ) 157 | 158 | self.out_planes = (len(dilations) + 2) * inner_planes 159 | 160 | def get_outplanes(self): 161 | return self.out_planes 162 | 163 | def forward(self, x): 164 | _, _, h, w = x.size() 165 | feat1 = F.interpolate( 166 | self.conv1(x), size=(h, w), mode="bilinear", align_corners=True 167 | ) 168 | feat2 = self.conv2(x) 169 | feat3 = self.conv3(x) 170 | feat4 = self.conv4(x) 171 | feat5 = self.conv5(x) 172 | aspp_out = torch.cat((feat1, feat2, feat3, feat4, feat5), 1) 173 | return aspp_out 174 | -------------------------------------------------------------------------------- /augseg/models/model_helper.py: -------------------------------------------------------------------------------- 1 | import importlib 2 | import torch.nn as nn 3 | from torch.nn import functional as F 4 | from .decoder import Aux_Module 5 | 6 | 7 | class ModelBuilder(nn.Module): 8 | def __init__(self, net_cfg): 9 | super(ModelBuilder, self).__init__() 10 | self._sync_bn = net_cfg["sync_bn"] 11 | self._num_classes = net_cfg["num_classes"] 12 | 13 | self.encoder = self._build_encoder(net_cfg["encoder"]) 14 | self.decoder = self._build_decoder(net_cfg["decoder"]) 15 | 16 | self._use_auxloss = True if net_cfg.get("aux_loss", False) else False 17 | if self._use_auxloss: 18 | cfg_aux = net_cfg["aux_loss"] 19 | self.loss_weight = cfg_aux["loss_weight"] 20 | self.auxor = Aux_Module( 21 | cfg_aux["aux_plane"], self._num_classes, self._sync_bn 22 | ) 23 | 24 | def _build_encoder(self, enc_cfg): 25 | enc_cfg["kwargs"].update({"sync_bn": self._sync_bn}) 26 | pretrained_model_url = enc_cfg["pretrain"] 27 | encoder = self._build_module(enc_cfg["type"], enc_cfg["kwargs"], pretrain_model_url=pretrained_model_url) 28 | return encoder 29 | 30 | def _build_decoder(self, dec_cfg): 31 | dec_cfg["kwargs"].update( 32 | { 33 | "in_planes": self.encoder.get_outplanes(), 34 | "sync_bn": self._sync_bn, 35 | "num_classes": self._num_classes, 36 | } 37 | ) 38 | decoder = self._build_module(dec_cfg["type"], dec_cfg["kwargs"]) 39 | return decoder 40 | 41 | def _build_module(self, mtype, kwargs, pretrain_model_url=None): 42 | module_name, class_name = mtype.rsplit(".", 1) 43 | module = importlib.import_module(module_name) 44 | cls = getattr(module, class_name) 45 | if pretrain_model_url is None: 46 | return cls(**kwargs) 47 | else: 48 | return cls(pretrain_model_url=pretrain_model_url, **kwargs) 49 | 50 | def forward(self, x, flag_use_fdrop=False): 51 | h, w = x.shape[-2:] 52 | if self._use_auxloss: 53 | f1, f2, feat1, feat2 = self.encoder(x) 54 | outs = self.decoder([f1, f2, feat1, feat2]) 55 | pred_aux = self.auxor(feat1) 56 | 57 | # upsampling 58 | outs = F.interpolate(outs, (h, w), mode="bilinear", align_corners=True) 59 | pred_aux = F.interpolate(pred_aux, (h, w), mode="bilinear", align_corners=True) 60 | 61 | return outs, pred_aux 62 | else: 63 | if flag_use_fdrop: 64 | f1, f2, feat1, feat2 = self.encoder(x) 65 | f1 = nn.Dropout2d(0.5)(f1) 66 | feat2 = nn.Dropout2d(0.5)(feat2) 67 | outs = self.decoder([f1, f2, feat1, feat2]) 68 | else: 69 | feat = self.encoder(x) 70 | outs = self.decoder(feat) 71 | 72 | outs = F.interpolate(outs, (h, w), mode="bilinear", align_corners=True) 73 | 74 | return outs, None 75 | -------------------------------------------------------------------------------- /augseg/models/resnet.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import torch.nn as nn 3 | 4 | import os, sys 5 | 6 | __all__ = [ 7 | "ResNet", 8 | "resnet18", 9 | "resnet34", 10 | "resnet50", 11 | "resnet101", 12 | "resnet152", 13 | ] 14 | 15 | # HOME_NAME = "SSS" 16 | # cur_path_file = os.getcwd() 17 | # lst_dir = cur_path_file.split(os.path.sep) 18 | # CODE_DIR = os.path.sep.join(lst_dir[:lst_dir.index(HOME_NAME)+1]) 19 | 20 | model_urls = { 21 | "resnet18": "/path/to/resnet18.pth", 22 | "resnet34": "/path/to/resnet34.pth", 23 | "resnet50": "/path/to/resnet50.pth", 24 | "resnet101": "/path/to/resnet101.pth", 25 | # "resnet50": os.path.join(CODE_DIR, "pretrained", "resnet50.pth"), 26 | # "resnet101": os.path.join(CODE_DIR, "pretrained", "resnet101.pth"), 27 | "resnet152": "/path/to/resnet152.pth", 28 | } 29 | 30 | 31 | def get_syncbn(): 32 | # return nn.BatchNorm2d 33 | return nn.SyncBatchNorm 34 | 35 | def conv3x3(in_planes, out_planes, stride=1, groups=1, dilation=1): 36 | """3x3 convolution with padding""" 37 | return nn.Conv2d( 38 | in_planes, 39 | out_planes, 40 | kernel_size=3, 41 | stride=stride, 42 | padding=dilation, 43 | groups=groups, 44 | bias=False, 45 | dilation=dilation, 46 | ) 47 | 48 | 49 | def conv1x1(in_planes, out_planes, stride=1): 50 | """1x1 convolution""" 51 | return nn.Conv2d(in_planes, out_planes, kernel_size=1, stride=stride, bias=False) 52 | 53 | 54 | class BasicBlock(nn.Module): 55 | expansion = 1 56 | 57 | def __init__( 58 | self, 59 | inplanes, 60 | planes, 61 | stride=1, 62 | downsample=None, 63 | groups=1, 64 | base_width=64, 65 | dilation=1, 66 | norm_layer=None, 67 | ): 68 | super(BasicBlock, self).__init__() 69 | if norm_layer is None: 70 | norm_layer = nn.BatchNorm2d 71 | if groups != 1 or base_width != 64: 72 | raise ValueError("BasicBlock only supports groups=1 and base_width=64") 73 | if dilation > 1: 74 | raise NotImplementedError("Dilation > 1 not supported in BasicBlock") 75 | # Both self.conv1 and self.downsample layers downsample the input when stride != 1 76 | self.conv1 = conv3x3(inplanes, planes, stride) 77 | self.bn1 = norm_layer(planes) 78 | self.relu = nn.ReLU(inplace=True) 79 | self.conv2 = conv3x3(planes, planes) 80 | self.bn2 = norm_layer(planes) 81 | self.downsample = downsample 82 | self.stride = stride 83 | 84 | def forward(self, x): 85 | identity = x 86 | 87 | out = self.conv1(x) 88 | out = self.bn1(out) 89 | out = self.relu(out) 90 | 91 | out = self.conv2(out) 92 | out = self.bn2(out) 93 | 94 | if self.downsample is not None: 95 | identity = self.downsample(x) 96 | 97 | out += identity 98 | out = self.relu(out) 99 | 100 | return out 101 | 102 | 103 | class Bottleneck(nn.Module): 104 | expansion = 4 105 | 106 | def __init__( 107 | self, 108 | inplanes, 109 | planes, 110 | stride=1, 111 | downsample=None, 112 | groups=1, 113 | base_width=64, 114 | dilation=1, 115 | norm_layer=nn.BatchNorm2d, 116 | ): 117 | super(Bottleneck, self).__init__() 118 | width = int(planes * (base_width / 64.0)) * groups 119 | # Both self.conv2 and self.downsample layers downsample the input when stride != 1 120 | self.conv1 = conv1x1(inplanes, width) 121 | self.bn1 = norm_layer(width) 122 | self.conv2 = conv3x3(width, width, stride, groups, dilation) 123 | self.bn2 = norm_layer(width) 124 | self.conv3 = conv1x1(width, planes * self.expansion) 125 | self.bn3 = norm_layer(planes * self.expansion) 126 | self.relu = nn.ReLU(inplace=True) 127 | self.downsample = downsample 128 | self.stride = stride 129 | 130 | def forward(self, x): 131 | identity = x 132 | 133 | out = self.conv1(x) 134 | out = self.bn1(out) 135 | out = self.relu(out) 136 | 137 | out = self.conv2(out) 138 | out = self.bn2(out) 139 | out = self.relu(out) 140 | 141 | out = self.conv3(out) 142 | out = self.bn3(out) 143 | 144 | if self.downsample is not None: 145 | identity = self.downsample(x) 146 | 147 | out += identity 148 | out = self.relu(out) 149 | 150 | return out 151 | 152 | 153 | class ResNet(nn.Module): 154 | def __init__( 155 | self, 156 | block, 157 | layers, 158 | zero_init_residual=False, 159 | groups=1, 160 | width_per_group=64, 161 | replace_stride_with_dilation=[False, False, False], 162 | sync_bn=False, 163 | multi_grid=False, 164 | ): 165 | super(ResNet, self).__init__() 166 | 167 | norm_layer = get_syncbn() if sync_bn else nn.BatchNorm2d 168 | self._norm_layer = norm_layer 169 | 170 | self.inplanes = 128 171 | self.dilation = 1 172 | 173 | if replace_stride_with_dilation is None: 174 | # each element in the tuple indicates if we should replace 175 | # the 2x2 stride with a dilated convolution instead 176 | replace_stride_with_dilation = [False, False, False] 177 | 178 | if len(replace_stride_with_dilation) != 3: 179 | raise ValueError( 180 | "replace_stride_with_dilation should be None " 181 | "or a 3-element tuple, got {}".format(replace_stride_with_dilation) 182 | ) 183 | 184 | self.groups = groups 185 | self.base_width = width_per_group 186 | self.conv1 = nn.Sequential( 187 | conv3x3(3, 64, stride=2), 188 | norm_layer(64), 189 | nn.ReLU(inplace=True), 190 | conv3x3(64, 64), 191 | norm_layer(64), 192 | nn.ReLU(inplace=True), 193 | conv3x3(64, self.inplanes), 194 | ) 195 | self.bn1 = norm_layer(self.inplanes) 196 | self.relu = nn.ReLU(inplace=True) 197 | self.maxpool = nn.MaxPool2d( 198 | kernel_size=3, stride=2, padding=1, ceil_mode=True 199 | ) # change 200 | 201 | self.layer1 = self._make_layer(block, 64, layers[0]) 202 | self.layer2 = self._make_layer( 203 | block, 128, layers[1], stride=2, dilate=replace_stride_with_dilation[0] 204 | ) 205 | self.layer3 = self._make_layer( 206 | block, 256, layers[2], stride=2, dilate=replace_stride_with_dilation[1] 207 | ) 208 | self.layer4 = self._make_layer( 209 | block, 210 | 512, 211 | layers[3], 212 | stride=2, 213 | dilate=replace_stride_with_dilation[2], 214 | multi_grid=multi_grid, 215 | ) 216 | 217 | for m in self.modules(): 218 | if isinstance(m, nn.Conv2d): 219 | nn.init.kaiming_normal_(m.weight, mode="fan_out", nonlinearity="relu") 220 | elif isinstance(m, (nn.BatchNorm2d, nn.GroupNorm, nn.SyncBatchNorm)): 221 | nn.init.constant_(m.weight, 1) 222 | nn.init.constant_(m.bias, 0) 223 | 224 | # Zero-initialize the last BN in each residual branch, 225 | # so that the residual branch starts with zeros, and each residual block behaves like an identity. 226 | # This improves the model by 0.2~0.3% according to https://arxiv.org/abs/1706.02677 227 | if zero_init_residual: 228 | for m in self.modules(): 229 | if isinstance(m, Bottleneck): 230 | nn.init.constant_(m.bn3.weight, 0) 231 | elif isinstance(m, BasicBlock): 232 | nn.init.constant_(m.bn2.weight, 0) 233 | 234 | def get_outplanes(self): 235 | return self.inplanes 236 | 237 | def get_auxplanes(self): 238 | return self.inplanes // 2 239 | 240 | def _make_layer( 241 | self, block, planes, blocks, stride=1, dilate=False, multi_grid=False 242 | ): 243 | norm_layer = self._norm_layer 244 | downsample = None 245 | previous_dilation = self.dilation 246 | if dilate: 247 | self.dilation *= stride 248 | stride = 1 249 | if stride != 1 or self.inplanes != planes * block.expansion: 250 | downsample = nn.Sequential( 251 | conv1x1(self.inplanes, planes * block.expansion, stride), 252 | norm_layer(planes * block.expansion), 253 | ) 254 | 255 | grids = [1] * blocks 256 | if multi_grid: 257 | grids = [2, 2, 4] 258 | 259 | layers = [] 260 | layers.append( 261 | block( 262 | self.inplanes, 263 | planes, 264 | stride, 265 | downsample, 266 | self.groups, 267 | self.base_width, 268 | previous_dilation * grids[0], 269 | norm_layer, 270 | ) 271 | ) 272 | self.inplanes = planes * block.expansion 273 | for i in range(1, blocks): 274 | layers.append( 275 | block( 276 | self.inplanes, 277 | planes, 278 | groups=self.groups, 279 | base_width=self.base_width, 280 | dilation=self.dilation * grids[i], 281 | norm_layer=norm_layer, 282 | ) 283 | ) 284 | 285 | return nn.Sequential(*layers) 286 | 287 | def forward(self, x): 288 | x = self.relu(self.bn1(self.conv1(x))) 289 | x = self.maxpool(x) 290 | 291 | x1 = self.layer1(x) 292 | x2 = self.layer2(x1) 293 | x3 = self.layer3(x2) 294 | x4 = self.layer4(x3) 295 | return [x1, x2, x3, x4] 296 | 297 | 298 | def resnet18(pretrained=True, pretrain_model_url=None, **kwargs): 299 | """Constructs a ResNet-18 model. 300 | 301 | Args: 302 | pretrained (bool): If True, returns a model pre-trained on ImageNet 303 | """ 304 | model = ResNet(BasicBlock, [2, 2, 2, 2], **kwargs) 305 | if pretrained: 306 | if pretrain_model_url is None: 307 | model_url = model_urls["resnet18"] 308 | else: 309 | model_url = pretrain_model_url 310 | state_dict = torch.load(model_url) 311 | 312 | missing_keys, unexpected_keys = model.load_state_dict(state_dict, strict=False) 313 | print( 314 | f"[Info] Load ImageNet pretrain from '{model_url}'", 315 | "\nmissing_keys: ", 316 | missing_keys, 317 | "\nunexpected_keys: ", 318 | unexpected_keys, 319 | ) 320 | return model 321 | 322 | 323 | def resnet34(pretrained=True, pretrain_model_url=None, **kwargs): 324 | """Constructs a ResNet-34 model. 325 | 326 | Args: 327 | pretrained (bool): If True, returns a model pre-trained on ImageNet 328 | """ 329 | model = ResNet(BasicBlock, [3, 4, 6, 3], **kwargs) 330 | if pretrained: 331 | if pretrain_model_url is None: 332 | model_url = model_urls["resnet34"] 333 | else: 334 | model_url = pretrain_model_url 335 | 336 | state_dict = torch.load(model_url) 337 | 338 | missing_keys, unexpected_keys = model.load_state_dict(state_dict, strict=False) 339 | print( 340 | f"[Info] Load ImageNet pretrain from '{model_url}'", 341 | "\nmissing_keys: ", 342 | missing_keys, 343 | "\nunexpected_keys: ", 344 | unexpected_keys, 345 | ) 346 | return model 347 | 348 | 349 | def resnet50(pretrained=True, pretrain_model_url=None, **kwargs): 350 | """Constructs a ResNet-50 model. 351 | 352 | Args: 353 | pretrained (bool): If True, returns a model pre-trained on ImageNet 354 | """ 355 | model = ResNet(Bottleneck, [3, 4, 6, 3], **kwargs) 356 | if pretrained: 357 | if pretrain_model_url is None: 358 | model_url = model_urls["resnet50"] 359 | else: 360 | model_url = pretrain_model_url 361 | 362 | state_dict = torch.load(model_url) 363 | 364 | missing_keys, unexpected_keys = model.load_state_dict(state_dict, strict=False) 365 | print( 366 | f"[Info] Load ImageNet pretrain from '{model_url}'", 367 | "\nmissing_keys: ", 368 | missing_keys, 369 | "\nunexpected_keys: ", 370 | unexpected_keys, 371 | ) 372 | return model 373 | 374 | 375 | def resnet101(pretrained=True, pretrain_model_url=None, **kwargs): 376 | """Constructs a ResNet-101 model. 377 | 378 | Args: 379 | pretrained (bool): If True, returns a model pre-trained on ImageNet 380 | """ 381 | model = ResNet(Bottleneck, [3, 4, 23, 3], **kwargs) 382 | if pretrained: 383 | if pretrain_model_url is None: 384 | model_url = model_urls["resnet101"] 385 | else: 386 | model_url = pretrain_model_url 387 | state_dict = torch.load(model_url) 388 | 389 | missing_keys, unexpected_keys = model.load_state_dict(state_dict, strict=False) 390 | print( 391 | f"[Info] Load ImageNet pretrain from '{model_url}'", 392 | "\nmissing_keys: ", 393 | missing_keys, 394 | "\nunexpected_keys: ", 395 | unexpected_keys, 396 | ) 397 | return model 398 | 399 | 400 | def resnet152(pretrained=True, pretrain_model_url=None, **kwargs): 401 | """Constructs a ResNet-152 model. 402 | 403 | Args: 404 | pretrained (bool): If True, returns a model pre-trained on ImageNet 405 | """ 406 | model = ResNet(Bottleneck, [3, 8, 36, 3], **kwargs) 407 | if pretrained: 408 | if pretrain_model_url is None: 409 | model_url = model_urls["resnet152"] 410 | else: 411 | model_url = pretrain_model_url 412 | state_dict = torch.load(model_url) 413 | 414 | missing_keys, unexpected_keys = model.load_state_dict(state_dict, strict=False) 415 | print( 416 | f"[Info] Load ImageNet pretrain from '{model_url}'", 417 | "\nmissing_keys: ", 418 | missing_keys, 419 | "\nunexpected_keys: ", 420 | unexpected_keys, 421 | ) 422 | return model 423 | -------------------------------------------------------------------------------- /augseg/utils/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ZhenZHAO/AugSeg/7ead8705cf6ce6c9234f52c414ec2237ed6743cd/augseg/utils/__init__.py -------------------------------------------------------------------------------- /augseg/utils/dist_helper.py: -------------------------------------------------------------------------------- 1 | import copy 2 | import math 3 | import os 4 | import random 5 | import subprocess 6 | 7 | import numpy as np 8 | import torch 9 | import torch.distributed as dist 10 | from torch.utils.data.sampler import Sampler 11 | 12 | 13 | def setup_distributed(backend="nccl", port=None): 14 | """AdaHessian Optimizer 15 | Lifted from https://github.com/BIGBALLON/distribuuuu/blob/master/distribuuuu/utils.py 16 | Originally licensed MIT, Copyright (c) 2020 Wei Li 17 | """ 18 | num_gpus = torch.cuda.device_count() 19 | 20 | if "SLURM_JOB_ID" in os.environ: 21 | rank = int(os.environ["SLURM_PROCID"]) 22 | world_size = int(os.environ["SLURM_NTASKS"]) 23 | node_list = os.environ["SLURM_NODELIST"] 24 | addr = subprocess.getoutput(f"scontrol show hostname {node_list} | head -n1") 25 | # specify master port 26 | if port is not None: 27 | os.environ["MASTER_PORT"] = str(port) 28 | elif "MASTER_PORT" not in os.environ: 29 | os.environ["MASTER_PORT"] = "10685" 30 | if "MASTER_ADDR" not in os.environ: 31 | os.environ["MASTER_ADDR"] = addr 32 | os.environ["WORLD_SIZE"] = str(world_size) 33 | os.environ["LOCAL_RANK"] = str(rank % num_gpus) 34 | os.environ["RANK"] = str(rank) 35 | else: 36 | rank = int(os.environ["RANK"]) 37 | world_size = int(os.environ["WORLD_SIZE"]) 38 | 39 | torch.cuda.set_device(rank % num_gpus) 40 | 41 | dist.init_process_group( 42 | backend=backend, 43 | world_size=world_size, 44 | rank=rank, 45 | ) 46 | return rank, world_size 47 | 48 | 49 | def gather_together(data): 50 | world_size = dist.get_world_size() 51 | gather_data = [torch.zeros_like(data).cuda() for _ in range(world_size)] 52 | dist.all_gather(gather_data, data) 53 | return gather_data 54 | 55 | 56 | class DistributedGivenIterationSampler(Sampler): 57 | def __init__( 58 | self, dataset, total_iter, batch_size, world_size=None, rank=None, last_iter=-1 59 | ): 60 | if world_size is None: 61 | world_size = dist.get_world_size() 62 | if rank is None: 63 | rank = dist.get_rank() 64 | assert rank < world_size 65 | self.dataset = dataset 66 | self.total_iter = total_iter 67 | self.batch_size = batch_size 68 | self.world_size = world_size 69 | self.rank = rank 70 | self.last_iter = last_iter 71 | 72 | self.total_size = self.total_iter * self.batch_size 73 | 74 | self.indices = self.gen_new_list() 75 | self.call = 0 76 | 77 | def __iter__(self): 78 | if self.call == 0: 79 | self.call = 1 80 | return iter(self.indices[(self.last_iter + 1) * self.batch_size :]) 81 | else: 82 | raise RuntimeError( 83 | "this sampler is not designed to be called more than once!!" 84 | ) 85 | 86 | def gen_new_list(self): 87 | # each process shuffle all list with same seed, and pick one piece according to rank 88 | np.random.seed(0) 89 | 90 | all_size = self.total_size * self.world_size 91 | indices = np.arange(len(self.dataset)) 92 | indices = indices[:all_size] 93 | num_repeat = (all_size - 1) // indices.shape[0] + 1 94 | indices = np.tile(indices, num_repeat) 95 | indices = indices[:all_size] 96 | 97 | np.random.shuffle(indices) 98 | beg = self.total_size * self.rank 99 | indices = indices[beg : beg + self.total_size] 100 | 101 | assert len(indices) == self.total_size 102 | 103 | return indices 104 | 105 | def __len__(self): 106 | # note here we do not take last iter into consideration, since __len__ 107 | # should only be used for displaying, the correct remaining size is 108 | # handled by dataloader 109 | # return self.total_size - (self.last_iter+1)*self.batch_size 110 | return self.total_size 111 | -------------------------------------------------------------------------------- /augseg/utils/loss_helper.py: -------------------------------------------------------------------------------- 1 | import numpy as np 2 | import scipy.ndimage as nd 3 | import torch 4 | import torch.nn as nn 5 | from torch.nn import functional as F 6 | 7 | 8 | # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # 9 | # # # # # 1. get training criterion 10 | # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # 11 | def get_criterion(cfg): 12 | cfg_criterion = cfg["criterion"] 13 | aux_weight = ( 14 | cfg["net"]["aux_loss"]["loss_weight"] 15 | if cfg["net"].get("aux_loss", False) 16 | else 0 17 | ) 18 | ignore_index = cfg["dataset"]["ignore_label"] 19 | if cfg_criterion["type"] == "ohem": 20 | criterion = CriterionOhem( 21 | aux_weight, ignore_index=ignore_index, **cfg_criterion["kwargs"] 22 | ) 23 | else: 24 | criterion = Criterion( 25 | aux_weight, ignore_index=ignore_index, **cfg_criterion["kwargs"] 26 | ) 27 | 28 | return criterion 29 | 30 | 31 | class Criterion(nn.Module): 32 | def __init__(self, aux_weight, ignore_index=255, use_weight=False): 33 | super(Criterion, self).__init__() 34 | self._aux_weight = aux_weight 35 | self._ignore_index = ignore_index 36 | self.use_weight = use_weight 37 | if not use_weight: 38 | self._criterion = nn.CrossEntropyLoss(ignore_index=ignore_index) 39 | else: 40 | weights = torch.FloatTensor( 41 | [ 42 | 0.0, 43 | 0.0, 44 | 0.0, 45 | 1.0, 46 | 1.0, 47 | 1.0, 48 | 1.0, 49 | 0.0, 50 | 0.0, 51 | 1.0, 52 | 0.0, 53 | 0.0, 54 | 1.0, 55 | 0.0, 56 | 1.0, 57 | 0.0, 58 | 1.0, 59 | 1.0, 60 | 1.0, 61 | ] 62 | ).cuda() 63 | self._criterion = nn.CrossEntropyLoss(ignore_index=ignore_index) 64 | self._criterion1 = nn.CrossEntropyLoss( 65 | ignore_index=ignore_index, weight=weights 66 | ) 67 | 68 | def forward(self, preds, target): 69 | h, w = target.size(1), target.size(2) 70 | if self._aux_weight > 0: # require aux loss 71 | main_pred, aux_pred = preds 72 | main_h, main_w = main_pred.size(2), main_pred.size(3) 73 | aux_h, aux_w = aux_pred.size(2), aux_pred.size(3) 74 | assert ( 75 | len(preds) == 2 76 | and main_h == aux_h 77 | and main_w == aux_w 78 | and main_h == h 79 | and main_w == w 80 | ) 81 | if self.use_weight: 82 | loss1 = self._criterion(main_pred, target) + self._criterion1( 83 | main_pred, target 84 | ) 85 | else: 86 | loss1 = self._criterion(main_pred, target) 87 | loss2 = self._criterion(aux_pred, target) 88 | loss = loss1 + self._aux_weight * loss2 89 | else: 90 | pred_h, pred_w = preds.size(2), preds.size(3) 91 | assert pred_h == h and pred_w == w 92 | loss = self._criterion(preds, target) 93 | return loss 94 | 95 | 96 | class CriterionOhem(nn.Module): 97 | def __init__( 98 | self, 99 | aux_weight, 100 | thresh=0.7, 101 | min_kept=100000, 102 | ignore_index=255, 103 | use_weight=False, 104 | ): 105 | super(CriterionOhem, self).__init__() 106 | self._aux_weight = aux_weight 107 | self._criterion1 = OhemCrossEntropy2dTensor( 108 | ignore_index, thresh, min_kept, use_weight 109 | ) 110 | self._criterion2 = OhemCrossEntropy2dTensor(ignore_index, thresh, min_kept) 111 | 112 | def forward(self, preds, target): 113 | h, w = target.size(1), target.size(2) 114 | if self._aux_weight > 0: # require aux loss 115 | main_pred, aux_pred = preds 116 | main_h, main_w = main_pred.size(2), main_pred.size(3) 117 | aux_h, aux_w = aux_pred.size(2), aux_pred.size(3) 118 | assert ( 119 | len(preds) == 2 120 | and main_h == aux_h 121 | and main_w == aux_w 122 | and main_h == h 123 | and main_w == w 124 | ) 125 | 126 | loss1 = self._criterion1(main_pred, target) 127 | loss2 = self._criterion2(aux_pred, target) 128 | loss = loss1 + self._aux_weight * loss2 129 | else: 130 | pred_h, pred_w = preds.size(2), preds.size(3) 131 | assert pred_h == h and pred_w == w 132 | loss = self._criterion1(preds, target) 133 | return loss 134 | 135 | 136 | class OhemCrossEntropy2d(nn.Module): 137 | def __init__(self, ignore_label=255, thresh=0.7, min_kept=100000, factor=8): 138 | super(OhemCrossEntropy2d, self).__init__() 139 | self.ignore_label = ignore_label 140 | self.thresh = float(thresh) 141 | self.min_kept = int(min_kept) 142 | self.factor = factor 143 | self.criterion = torch.nn.CrossEntropyLoss(ignore_index=ignore_label) 144 | 145 | def find_threshold(self, np_predict, np_target): 146 | # downsample 1/8 147 | factor = self.factor 148 | predict = nd.zoom(np_predict, (1.0, 1.0, 1.0 / factor, 1.0 / factor), order=1) 149 | target = nd.zoom(np_target, (1.0, 1.0 / factor, 1.0 / factor), order=0) 150 | 151 | n, c, h, w = predict.shape 152 | min_kept = self.min_kept // ( 153 | factor * factor 154 | ) # int(self.min_kept_ratio * n * h * w) 155 | 156 | input_label = target.ravel().astype(np.int32) 157 | input_prob = np.rollaxis(predict, 1).reshape((c, -1)) 158 | 159 | valid_flag = input_label != self.ignore_label 160 | valid_inds = np.where(valid_flag)[0] 161 | label = input_label[valid_flag] 162 | num_valid = valid_flag.sum() 163 | if min_kept >= num_valid: 164 | threshold = 1.0 165 | elif num_valid > 0: 166 | prob = input_prob[:, valid_flag] 167 | pred = prob[label, np.arange(len(label), dtype=np.int32)] 168 | threshold = self.thresh 169 | if min_kept > 0: 170 | k_th = min(len(pred), min_kept) - 1 171 | new_array = np.partition(pred, k_th) 172 | new_threshold = new_array[k_th] 173 | if new_threshold > self.thresh: 174 | threshold = new_threshold 175 | return threshold 176 | 177 | def generate_new_target(self, predict, target): 178 | np_predict = predict.data.cpu().numpy() 179 | np_target = target.data.cpu().numpy() 180 | n, c, h, w = np_predict.shape 181 | 182 | threshold = self.find_threshold(np_predict, np_target) 183 | 184 | input_label = np_target.ravel().astype(np.int32) 185 | input_prob = np.rollaxis(np_predict, 1).reshape((c, -1)) 186 | 187 | valid_flag = input_label != self.ignore_label 188 | valid_inds = np.where(valid_flag)[0] 189 | label = input_label[valid_flag] 190 | num_valid = valid_flag.sum() 191 | 192 | if num_valid > 0: 193 | prob = input_prob[:, valid_flag] 194 | pred = prob[label, np.arange(len(label), dtype=np.int32)] 195 | kept_flag = pred <= threshold 196 | valid_inds = valid_inds[kept_flag] 197 | 198 | label = input_label[valid_inds].copy() 199 | input_label.fill(self.ignore_label) 200 | input_label[valid_inds] = label 201 | new_target = ( 202 | torch.from_numpy(input_label.reshape(target.size())) 203 | .long() 204 | .cuda(target.get_device()) 205 | ) 206 | 207 | return new_target 208 | 209 | def forward(self, predict, target, weight=None): 210 | """ 211 | Args: 212 | predict:(n, c, h, w) 213 | target:(n, h, w) 214 | weight (Tensor, optional): a manual rescaling weight given to each class. 215 | If given, has to be a Tensor of size "nclasses" 216 | """ 217 | assert not target.requires_grad 218 | 219 | input_prob = F.softmax(predict, 1) 220 | target = self.generate_new_target(input_prob, target) 221 | return self.criterion(predict, target) 222 | 223 | 224 | class OhemCrossEntropy2dTensor(nn.Module): 225 | """ 226 | Ohem Cross Entropy Tensor Version 227 | """ 228 | 229 | def __init__( 230 | self, ignore_index=255, thresh=0.7, min_kept=256, use_weight=False, reduce=False 231 | ): 232 | super(OhemCrossEntropy2dTensor, self).__init__() 233 | self.ignore_index = ignore_index 234 | self.thresh = float(thresh) 235 | self.min_kept = int(min_kept) 236 | if use_weight: 237 | weight = torch.FloatTensor( 238 | [ 239 | 0.8373, 240 | 0.918, 241 | 0.866, 242 | 1.0345, 243 | 1.0166, 244 | 0.9969, 245 | 0.9754, 246 | 1.0489, 247 | 0.8786, 248 | 1.0023, 249 | 0.9539, 250 | 0.9843, 251 | 1.1116, 252 | 0.9037, 253 | 1.0865, 254 | 1.0955, 255 | 1.0865, 256 | 1.1529, 257 | 1.0507, 258 | ] 259 | ).cuda() 260 | # weight = torch.FloatTensor( 261 | # [0.4762, 0.5, 0.4762, 1.4286, 1.1111, 0.4762, 0.8333, 0.5, 0.5, 0.8333, 0.5263, 0.5882, 262 | # 1.4286, 0.5, 3.3333,5.0, 10.0, 2.5, 0.8333]).cuda() 263 | self.criterion = torch.nn.CrossEntropyLoss( 264 | reduction="mean", weight=weight, ignore_index=ignore_index 265 | ) 266 | elif reduce: 267 | self.criterion = torch.nn.CrossEntropyLoss( 268 | reduction="none", ignore_index=ignore_index 269 | ) 270 | else: 271 | self.criterion = torch.nn.CrossEntropyLoss( 272 | reduction="mean", ignore_index=ignore_index 273 | ) 274 | 275 | def forward(self, pred, target): 276 | b, c, h, w = pred.size() 277 | target = target.view(-1) 278 | valid_mask = target.ne(self.ignore_index) 279 | target = target * valid_mask.long() 280 | num_valid = valid_mask.sum() 281 | 282 | prob = F.softmax(pred, dim=1) 283 | prob = (prob.transpose(0, 1)).reshape(c, -1) 284 | 285 | if self.min_kept > num_valid: 286 | pass 287 | # print('Labels: {}'.format(num_valid)) 288 | elif num_valid > 0: 289 | prob = prob.masked_fill_(~valid_mask, 1) 290 | mask_prob = prob[target, torch.arange(len(target), dtype=torch.long)] 291 | threshold = self.thresh 292 | if self.min_kept > 0: 293 | _, index = mask_prob.sort() 294 | threshold_index = index[min(len(index), self.min_kept) - 1] 295 | if mask_prob[threshold_index] > self.thresh: 296 | threshold = mask_prob[threshold_index] 297 | kept_mask = mask_prob.le(threshold) 298 | target = target * kept_mask.long() 299 | valid_mask = valid_mask * kept_mask 300 | 301 | target = target.masked_fill_(~valid_mask, self.ignore_index) 302 | target = target.view(b, h, w) 303 | 304 | return self.criterion(pred, target) 305 | 306 | 307 | # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # 308 | # # # # # 2. calculate unsupervised loss 309 | # # # # # # # # # # # # # # # # # # # # # # # # # # # # # # 310 | def compute_unsupervised_loss_by_threshold(predict, target, logits, thresh=0.95): 311 | batch_size, num_class, h, w = predict.shape 312 | thresh_mask = logits.ge(thresh).bool() * (target != 255).bool() 313 | target[~thresh_mask] = 255 314 | loss = F.cross_entropy(predict, target, ignore_index=255, reduction="none") 315 | return loss.mean(), thresh_mask.float().mean() 316 | 317 | 318 | def compute_unsupervised_loss_by_threshold_hardness(predict, target, logits, thresh=0.95, hardness_tensor=None): 319 | batch_size, num_class, h, w = predict.shape 320 | thresh_mask = logits.ge(thresh).bool() * (target != 255).bool() 321 | target[~thresh_mask] = 255 322 | loss = F.cross_entropy(predict, target, ignore_index=255, reduction="none") 323 | if hardness_tensor is None: 324 | return loss.mean() 325 | loss = loss.mean(dim=[1,2]) 326 | assert loss.shape == hardness_tensor.shape, "wrong hardness calculation!" 327 | loss *= hardness_tensor 328 | return loss.mean() 329 | -------------------------------------------------------------------------------- /augseg/utils/lr_helper.py: -------------------------------------------------------------------------------- 1 | """Learning Rate Schedulers""" 2 | from __future__ import division 3 | 4 | import copy 5 | import logging 6 | import warnings 7 | from math import cos, pi 8 | 9 | import torch.optim as optim 10 | 11 | 12 | def get_optimizer(parms, cfg_optim): 13 | """ 14 | Get the optimizer 15 | """ 16 | optim_type = cfg_optim["type"] 17 | optim_kwargs = cfg_optim["kwargs"] 18 | if optim_type == "SGD": 19 | optimizer = optim.SGD(parms, **optim_kwargs) 20 | elif optim_type == "adam": 21 | optimizer = optim.Adam(parms, **optim_kwargs) 22 | else: 23 | optimizer = None 24 | 25 | assert optimizer is not None, "optimizer type is not supported by LightSeg" 26 | 27 | return optimizer 28 | 29 | 30 | def get_scheduler(cfg_trainer, len_data, optimizer, start_epoch=0, use_iteration=False): 31 | epochs = ( 32 | cfg_trainer["epochs"] if not use_iteration else 1 33 | ) # if use_iteration = True, only one epoch be use 34 | lr_mode = cfg_trainer["lr_scheduler"]["mode"] 35 | lr_args = cfg_trainer["lr_scheduler"]["kwargs"] 36 | lr_scheduler = LRScheduler( 37 | lr_mode, lr_args, len_data, optimizer, epochs, start_epoch 38 | ) 39 | return lr_scheduler 40 | 41 | 42 | class LRScheduler(object): 43 | def __init__(self, mode, lr_args, data_size, optimizer, num_epochs, start_epochs): 44 | super(LRScheduler, self).__init__() 45 | logger = logging.getLogger("global") 46 | 47 | assert mode in ["multistep", "poly", "cosine"] 48 | self.mode = mode 49 | self.optimizer = optimizer 50 | self.data_size = data_size 51 | 52 | self.cur_iter = start_epochs * data_size 53 | self.max_iter = num_epochs * data_size 54 | 55 | # set learning rate 56 | self.base_lr = [ 57 | param_group["lr"] for param_group in self.optimizer.param_groups 58 | ] 59 | self.cur_lr = [lr for lr in self.base_lr] 60 | 61 | # poly kwargs 62 | # TODO 63 | if mode == "poly": 64 | self.power = lr_args["power"] if lr_args.get("power", False) else 0.9 65 | # logger.info("The kwargs for lr scheduler: {}".format(self.power)) 66 | if mode == "milestones": 67 | default_mist = list(range(0, num_epochs, num_epochs // 3))[1:] 68 | self.milestones = ( 69 | lr_args["milestones"] 70 | if lr_args.get("milestones", False) 71 | else default_mist 72 | ) 73 | # logger.info("The kwargs for lr scheduler: {}".format(self.milestones)) 74 | if mode == "cosine": 75 | self.targetlr = lr_args["targetlr"] 76 | # logger.info("The kwargs for lr scheduler: {}".format(self.targetlr)) 77 | 78 | def step(self): 79 | self._step() 80 | self.update_lr() 81 | self.cur_iter += 1 82 | 83 | def _step(self): 84 | if self.mode == "step": 85 | epoch = self.cur_iter // self.data_size 86 | power = sum([1 for s in self.milestones if s <= epoch]) 87 | for i, lr in enumerate(self.base_lr): 88 | adj_lr = lr * pow(0.1, power) 89 | self.cur_lr[i] = adj_lr 90 | elif self.mode == "poly": 91 | for i, lr in enumerate(self.base_lr): 92 | adj_lr = lr * ( 93 | (1 - float(self.cur_iter) / self.max_iter) ** (self.power) 94 | ) 95 | self.cur_lr[i] = adj_lr 96 | elif self.mode == "cosine": 97 | for i, lr in enumerate(self.base_lr): 98 | adj_lr = ( 99 | self.targetlr 100 | + (lr - self.targetlr) 101 | * (1 + cos(pi * self.cur_iter / self.max_iter)) 102 | / 2 103 | ) 104 | self.cur_lr[i] = adj_lr 105 | else: 106 | raise NotImplementedError 107 | 108 | def get_lr(self): 109 | return self.cur_lr 110 | 111 | def update_lr(self): 112 | for param_group, lr in zip(self.optimizer.param_groups, self.cur_lr): 113 | param_group["lr"] = lr 114 | -------------------------------------------------------------------------------- /data/splitsall.tar.gz: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ZhenZHAO/AugSeg/7ead8705cf6ce6c9234f52c414ec2237ed6743cd/data/splitsall.tar.gz -------------------------------------------------------------------------------- /docs/Augseg-diagram.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ZhenZHAO/AugSeg/7ead8705cf6ce6c9234f52c414ec2237ed6743cd/docs/Augseg-diagram.png -------------------------------------------------------------------------------- /docs/augs-cutmix.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ZhenZHAO/AugSeg/7ead8705cf6ce6c9234f52c414ec2237ed6743cd/docs/augs-cutmix.png -------------------------------------------------------------------------------- /docs/augs-intensity.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ZhenZHAO/AugSeg/7ead8705cf6ce6c9234f52c414ec2237ed6743cd/docs/augs-intensity.png -------------------------------------------------------------------------------- /exps/zrun_citys/citys_semi744/config_semi.yaml: -------------------------------------------------------------------------------- 1 | # # # # # # # # # # # # # # 2 | # 1. datasets 3 | # # # # # # # # # # # # # # 4 | dataset: # Required. 5 | type: cityscapes_semi 6 | train: 7 | data_root: ./data/cityscapes 8 | data_list: ./data/splitsall/cityscapes/744/labeled.txt 9 | batch_size: 4 10 | flip: True 11 | rand_resize: [0.5, 2.0] 12 | resize_base_size: [1024, 2048] 13 | crop: 14 | type: rand 15 | size: [800, 800] 16 | #---- 1) strong data augs ----# 17 | strong_aug: 18 | num_augs: 3 19 | flag_use_random_num_sampling: True 20 | val: 21 | data_root: ./data/cityscapes 22 | data_list: ./data/splitsall/cityscapes/val.txt 23 | batch_size: 4 24 | crop: 25 | size: [800, 800] 26 | n_sup: 744 27 | workers: 4 28 | mean: [0.485, 0.456, 0.406] 29 | std: [0.229, 0.224, 0.225] 30 | ignore_label: 255 31 | 32 | # # # # # # # # # # # # # # 33 | # 2. training params 34 | # # # # # # # # # # # # # # 35 | trainer: # Required. 36 | epochs: 240 37 | sup_only_epoch: 0 # 0, 1 38 | evaluate_student: False 39 | optimizer: 40 | type: SGD 41 | kwargs: 42 | lr: 0.01 # 4GPUs 43 | momentum: 0.9 44 | weight_decay: 0.0005 45 | lr_scheduler: 46 | mode: poly 47 | kwargs: 48 | power: 0.9 49 | unsupervised: 50 | flag_extra_weak: False 51 | threshold: -0.7 # ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ params1: 0.7, 0.75, 0.8, 0.85, 0.9, 0.95 52 | loss_weight: 1.0 # ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ params2: 0.7, 1.0, 1.5, 2.0, 3.0, 4.0 53 | #------ 2)cutmix augs ------# 54 | use_cutmix: True 55 | use_cutmix_adaptive: True 56 | use_cutmix_trigger_prob: 1.0 # wide range, but trigger by 100% 57 | 58 | # # # # # # # # # # # # # # 59 | # 3. output files, and loss 60 | # # # # # # # # # # # # # # 61 | saver: 62 | snapshot_dir: checkpoints 63 | pretrain: '' 64 | use_tb: False 65 | 66 | criterion: 67 | type: ohem 68 | kwargs: 69 | thresh: 0.7 70 | min_kept: 100000 71 | 72 | # # # # # # # # # # # # # # 73 | # 4. models 74 | # # # # # # # # # # # # # # 75 | net: # Required. 76 | num_classes: 19 77 | sync_bn: True 78 | ema_decay: 0.996 79 | # aux_loss: 80 | # aux_plane: 1024 81 | # loss_weight: 0.4 82 | encoder: 83 | type: augseg.models.resnet.resnet101 84 | pretrain: ./pretrained/resnet101.pth 85 | # type: augseg.models.resnet.resnet50 86 | # pretrain: ./pretrained/resnet50.pth 87 | kwargs: 88 | zero_init_residual: True 89 | multi_grid: True 90 | replace_stride_with_dilation: [False, False, True] # [output_stride = 16] 91 | # replace_stride_with_dilation: [False, True, True] # [output_stride = 8] 92 | decoder: 93 | type: augseg.models.decoder.dec_deeplabv3_plus 94 | kwargs: 95 | inner_planes: 256 96 | low_conv_planes: 48 # 256 97 | dilations: [6, 12, 18] 98 | # dilations: [12, 24, 36] # [output_stride = 8] 99 | -------------------------------------------------------------------------------- /exps/zrun_citys/r50_citys_semi744/config_semi.yaml: -------------------------------------------------------------------------------- 1 | # # # # # # # # # # # # # # 2 | # 1. datasets 3 | # # # # # # # # # # # # # # 4 | dataset: # Required. 5 | type: cityscapes_semi 6 | train: 7 | data_root: ./data/cityscapes 8 | data_list: ./data/splitsall/cityscapes/744/labeled.txt 9 | batch_size: 4 10 | flip: True 11 | rand_resize: [0.5, 2.0] 12 | resize_base_size: [1024, 2048] 13 | crop: 14 | type: rand 15 | size: [800, 800] 16 | #---- 1) strong data augs ----# 17 | strong_aug: 18 | num_augs: 3 19 | flag_use_random_num_sampling: True 20 | val: 21 | data_root: ./data/cityscapes 22 | data_list: ./data/splitsall/cityscapes/val.txt 23 | batch_size: 4 24 | crop: 25 | size: [800, 800] 26 | n_sup: 744 27 | workers: 4 28 | mean: [0.485, 0.456, 0.406] 29 | std: [0.229, 0.224, 0.225] 30 | ignore_label: 255 31 | 32 | # # # # # # # # # # # # # # 33 | # 2. training params 34 | # # # # # # # # # # # # # # 35 | trainer: # Required. 36 | epochs: 240 37 | sup_only_epoch: 0 # 0, 1 38 | evaluate_student: False 39 | optimizer: 40 | type: SGD 41 | kwargs: 42 | lr: 0.01 # 4GPUs 43 | momentum: 0.9 44 | weight_decay: 0.0005 45 | lr_scheduler: 46 | mode: poly 47 | kwargs: 48 | power: 0.9 49 | unsupervised: 50 | flag_extra_weak: False 51 | threshold: -0.7 # ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ params1: 0.7, 0.75, 0.8, 0.85, 0.9, 0.95 52 | loss_weight: 2.0 # ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ params2: 0.7, 1.0, 1.5, 2.0, 3.0, 4.0 53 | #------ 2)cutmix augs ------# 54 | use_cutmix: True 55 | use_cutmix_adaptive: True 56 | use_cutmix_trigger_prob: 1.0 # wide range, but trigger by 100% 57 | 58 | # # # # # # # # # # # # # # 59 | # 3. output files, and loss 60 | # # # # # # # # # # # # # # 61 | saver: 62 | snapshot_dir: checkpoints 63 | pretrain: '' 64 | use_tb: False 65 | 66 | criterion: 67 | type: ohem 68 | kwargs: 69 | thresh: 0.7 70 | min_kept: 100000 71 | 72 | # # # # # # # # # # # # # # 73 | # 4. models 74 | # # # # # # # # # # # # # # 75 | net: # Required. 76 | num_classes: 19 77 | sync_bn: True 78 | ema_decay: 0.996 79 | # aux_loss: 80 | # aux_plane: 1024 81 | # loss_weight: 0.4 82 | encoder: 83 | # type: augseg.models.resnet.resnet101 84 | # pretrain: ./pretrained/resnet101.pth 85 | type: augseg.models.resnet.resnet50 86 | pretrain: ./pretrained/resnet50.pth 87 | kwargs: 88 | zero_init_residual: True 89 | multi_grid: True 90 | replace_stride_with_dilation: [False, False, True] # [output_stride = 16] 91 | # replace_stride_with_dilation: [False, True, True] # [output_stride = 8] 92 | decoder: 93 | type: augseg.models.decoder.dec_deeplabv3_plus 94 | kwargs: 95 | inner_planes: 256 96 | low_conv_planes: 48 # 256 97 | dilations: [6, 12, 18] 98 | # dilations: [12, 24, 36] # [output_stride = 8] 99 | -------------------------------------------------------------------------------- /exps/zrun_vocs/r50_voc_semi662/config_semi.yaml: -------------------------------------------------------------------------------- 1 | # # # # # # # # # # # # # # 2 | # 1. datasets 3 | # # # # # # # # # # # # # # 4 | dataset: # Required. 5 | type: pascal_semi 6 | train: 7 | data_root: ./data/VOC2012 8 | data_list: ./data/splitsall/pascal/662/labeled.txt 9 | batch_size: 8 10 | flip: True 11 | rand_resize: [0.5, 2.0] 12 | resize_base_size: 500 13 | crop: 14 | type: rand 15 | size: [513, 513] 16 | #---- 1) strong data augs ----# 17 | strong_aug: 18 | num_augs: 3 # !!!!!!!!!!!!!!!!!!!! key param1 (4): 1, 2, 3, 4 !!!!!!!!!!!!!!!!!!!! 19 | flag_use_random_num_sampling: True # !!!!!!!!!!!!!!!!!!!! key param1 (2): True, False !!!!!!!!!!!!!!!!!!!! 20 | val: 21 | data_root: ./data/VOC2012 22 | data_list: ./data/splitsall/pascal/val.txt 23 | batch_size: 1 24 | n_sup: 662 25 | workers: 4 26 | mean: [0.485, 0.456, 0.406] 27 | std: [0.229, 0.224, 0.225] 28 | ignore_label: 255 29 | 30 | # # # # # # # # # # # # # # 31 | # 2. training params 32 | # # # # # # # # # # # # # # 33 | trainer: # Required. 34 | epochs: 80 35 | sup_only_epoch: 0 # 0, 1 36 | evaluate_student: True 37 | optimizer: 38 | type: SGD 39 | kwargs: 40 | lr: 0.001 # 0.001:8*2gpus, 4*4gpus; 0.002:8*4gpus. 41 | momentum: 0.9 42 | weight_decay: 0.0001 43 | lr_scheduler: 44 | mode: poly 45 | kwargs: 46 | power: 0.9 47 | # # # # # # # # # # # # # # 48 | # unsupervised loss 49 | # # # # # # # # # # # # # # 50 | unsupervised: 51 | flag_extra_weak: False 52 | threshold: 0.95 # ============================================================================= <>: [0.7, 0.75, 0.8, 0.85, 0.9, 0.95] 53 | loss_weight: 1.0 # ============================================================================= << abl2-weight >> : [1.0, 1.5, 2.0, 2.5, 3.0, 4.0] 54 | #------ 2)cutmix augs ------# 55 | use_cutmix: True 56 | use_cutmix_adaptive: True 57 | use_cutmix_trigger_prob: 1.0 58 | 59 | # # # # # # # # # # # # # # 60 | # 3. output files, and loss 61 | # # # # # # # # # # # # # # 62 | saver: 63 | snapshot_dir: checkpoints 64 | pretrain: '' 65 | use_tb: False 66 | 67 | criterion: 68 | type: CELoss 69 | kwargs: 70 | use_weight: False 71 | 72 | # # # # # # # # # # # # # # 73 | # 4. models 74 | # # # # # # # # # # # # # # 75 | net: # Required. 76 | num_classes: 21 77 | sync_bn: True 78 | ema_decay: 0.999 79 | encoder: 80 | type: augseg.models.resnet.resnet50 81 | pretrain: ./pretrained/resnet50.pth 82 | # type: augseg.models.resnet.resnet101 83 | # pretrain: ./pretrained/resnet101.pth 84 | kwargs: 85 | zero_init_residual: True 86 | multi_grid: True 87 | replace_stride_with_dilation: [False, False, True] # [output_stride = 16] 88 | # replace_stride_with_dilation: [False, True, True] # [output_stride = 8] 89 | decoder: 90 | type: augseg.models.decoder.dec_deeplabv3_plus 91 | kwargs: 92 | inner_planes: 256 93 | low_conv_planes: 48 94 | dilations: [6, 12, 18] # [output_stride = 16] 95 | # dilations: [12, 24, 36] # [output_stride = 8] 96 | -------------------------------------------------------------------------------- /exps/zrun_vocs/voc_semi_fine92_r50/config_semi.yaml: -------------------------------------------------------------------------------- 1 | # # # # # # # # # # # # # # 2 | # 1. datasets 3 | # # # # # # # # # # # # # # 4 | dataset: # Required. 5 | type: pascal_semi 6 | train: 7 | data_root: ./data/VOC2012 8 | data_list: ./data/splitsall/pascal/92/labeled.txt 9 | batch_size: 8 10 | flip: True 11 | rand_resize: [0.5, 2.0] 12 | resize_base_size: 500 13 | crop: 14 | type: rand 15 | size: [513, 513] 16 | #---- 1) strong data augs ----# 17 | strong_aug: 18 | num_augs: 3 # !!!!!!!!!!!!!!!!!!!! key param1 (4): 1, 2, 3, 4 !!!!!!!!!!!!!!!!!!!! 19 | flag_use_random_num_sampling: True # !!!!!!!!!!!!!!!!!!!! key param1 (2): True, False !!!!!!!!!!!!!!!!!!!! 20 | val: 21 | data_root: ./data/VOC2012 22 | data_list: ./data/splitsall/pascal/val.txt 23 | batch_size: 1 24 | n_sup: 92 25 | workers: 4 26 | mean: [0.485, 0.456, 0.406] 27 | std: [0.229, 0.224, 0.225] 28 | ignore_label: 255 29 | 30 | # # # # # # # # # # # # # # 31 | # 2. training params 32 | # # # # # # # # # # # # # # 33 | trainer: # Required. 34 | epochs: 80 35 | sup_only_epoch: 0 # 0, 1 36 | evaluate_student: True 37 | optimizer: 38 | type: SGD 39 | kwargs: 40 | lr: 0.001 # 0.001:8*2gpus, 4*4gpus; 0.002:8*4gpus. 41 | momentum: 0.9 42 | weight_decay: 0.0001 43 | lr_scheduler: 44 | mode: poly 45 | kwargs: 46 | power: 0.9 47 | # # # # # # # # # # # # # # 48 | # unsupervised loss 49 | # # # # # # # # # # # # # # 50 | unsupervised: 51 | flag_extra_weak: True 52 | threshold: 0.95 # ============================================================================= <>: [0.7, 0.75, 0.8, 0.85, 0.9, 0.95] 53 | loss_weight: 1.0 # ============================================================================= << abl2-weight >> : [1.0, 1.5, 2.0, 2.5, 3.0, 4.0] 54 | #------ 2)cutmix augs ------# 55 | use_cutmix: True 56 | use_cutmix_adaptive: True 57 | use_cutmix_trigger_prob: 1.0 58 | 59 | # # # # # # # # # # # # # # 60 | # 3. output files, and loss 61 | # # # # # # # # # # # # # # 62 | saver: 63 | snapshot_dir: checkpoints 64 | pretrain: '' 65 | use_tb: False 66 | 67 | criterion: 68 | type: CELoss 69 | kwargs: 70 | use_weight: False 71 | 72 | # # # # # # # # # # # # # # 73 | # 4. models 74 | # # # # # # # # # # # # # # 75 | net: # Required. 76 | num_classes: 21 77 | sync_bn: True 78 | ema_decay: 0.999 79 | encoder: 80 | type: augseg.models.resnet.resnet50 81 | pretrain: ./pretrained/resnet50.pth 82 | # type: augseg.models.resnet.resnet101 83 | # pretrain: ./pretrained/resnet101.pth 84 | kwargs: 85 | zero_init_residual: True 86 | multi_grid: True 87 | replace_stride_with_dilation: [False, False, True] # [output_stride = 16] 88 | # replace_stride_with_dilation: [False, True, True] # [output_stride = 8] 89 | decoder: 90 | type: augseg.models.decoder.dec_deeplabv3_plus 91 | kwargs: 92 | inner_planes: 256 93 | low_conv_planes: 48 94 | dilations: [6, 12, 18] # [output_stride = 16] 95 | # dilations: [12, 24, 36] # [output_stride = 8] 96 | -------------------------------------------------------------------------------- /exps/zrun_vocs_u2pl/r50_voc_semi662/config_semi.yaml: -------------------------------------------------------------------------------- 1 | # # # # # # # # # # # # # # 2 | # 1. datasets 3 | # # # # # # # # # # # # # # 4 | dataset: # Required. 5 | type: pascal_semi 6 | train: 7 | data_root: ./data/VOC2012 8 | data_list: ./data/splitsall/pascal_u2pl/662/labeled.txt 9 | batch_size: 8 10 | flip: True 11 | rand_resize: [0.5, 2.0] 12 | resize_base_size: 500 13 | crop: 14 | type: rand 15 | size: [513, 513] 16 | #---- 1) strong data augs ----# 17 | strong_aug: 18 | num_augs: 3 # !!!!!!!!!!!!!!!!!!!! key param1 (4): 1, 2, 3, 4 !!!!!!!!!!!!!!!!!!!! 19 | flag_use_random_num_sampling: True # !!!!!!!!!!!!!!!!!!!! key param1 (2): True, False !!!!!!!!!!!!!!!!!!!! 20 | val: 21 | data_root: ./data/VOC2012 22 | data_list: ./data/splitsall/pascal_u2pl/val.txt 23 | batch_size: 1 24 | n_sup: 662 25 | workers: 4 26 | mean: [0.485, 0.456, 0.406] 27 | std: [0.229, 0.224, 0.225] 28 | ignore_label: 255 29 | 30 | # # # # # # # # # # # # # # 31 | # 2. training params 32 | # # # # # # # # # # # # # # 33 | trainer: # Required. 34 | epochs: 80 35 | sup_only_epoch: 0 # 0, 1 36 | evaluate_student: True 37 | optimizer: 38 | type: SGD 39 | kwargs: 40 | lr: 0.001 # 0.001:8*2gpus, 4*4gpus; 0.002:8*4gpus. 41 | momentum: 0.9 42 | weight_decay: 0.0001 43 | lr_scheduler: 44 | mode: poly 45 | kwargs: 46 | power: 0.9 47 | # # # # # # # # # # # # # # 48 | # unsupervised loss 49 | # # # # # # # # # # # # # # 50 | unsupervised: 51 | flag_extra_weak: False 52 | threshold: 0.95 # ============================================================================= <>: [0.7, 0.75, 0.8, 0.85, 0.9, 0.95] 53 | loss_weight: 1.0 # ============================================================================= << abl2-weight >> : [1.0, 1.5, 2.0, 2.5, 3.0, 4.0] 54 | #------ 2)cutmix augs ------# 55 | use_cutmix: True 56 | use_cutmix_adaptive: True 57 | use_cutmix_trigger_prob: 1.0 58 | 59 | # # # # # # # # # # # # # # 60 | # 3. output files, and loss 61 | # # # # # # # # # # # # # # 62 | saver: 63 | snapshot_dir: checkpoints 64 | pretrain: '' 65 | use_tb: False 66 | 67 | criterion: 68 | type: CELoss 69 | kwargs: 70 | use_weight: False 71 | 72 | # # # # # # # # # # # # # # 73 | # 4. models 74 | # # # # # # # # # # # # # # 75 | net: # Required. 76 | num_classes: 21 77 | sync_bn: True 78 | ema_decay: 0.999 79 | encoder: 80 | type: augseg.models.resnet.resnet50 81 | pretrain: ./pretrained/resnet50.pth 82 | # type: augseg.models.resnet.resnet101 83 | # pretrain: ./pretrained/resnet101.pth 84 | kwargs: 85 | zero_init_residual: True 86 | multi_grid: True 87 | replace_stride_with_dilation: [False, False, True] # [output_stride = 16] 88 | # replace_stride_with_dilation: [False, True, True] # [output_stride = 8] 89 | decoder: 90 | type: augseg.models.decoder.dec_deeplabv3_plus 91 | kwargs: 92 | inner_planes: 256 93 | low_conv_planes: 48 94 | dilations: [6, 12, 18] # [output_stride = 16] 95 | # dilations: [12, 24, 36] # [output_stride = 8] 96 | -------------------------------------------------------------------------------- /exps/zrun_vocs_u2pl/voc_semi662/config_semi.yaml: -------------------------------------------------------------------------------- 1 | # # # # # # # # # # # # # # 2 | # 1. datasets 3 | # # # # # # # # # # # # # # 4 | dataset: # Required. 5 | type: pascal_semi 6 | train: 7 | data_root: ./data/VOC2012 8 | data_list: ./data/splitsall/pascal_u2pl/662/labeled.txt 9 | batch_size: 8 10 | flip: True 11 | rand_resize: [0.5, 2.0] 12 | resize_base_size: 500 13 | crop: 14 | type: rand 15 | size: [513, 513] 16 | #---- 1) strong data augs ----# 17 | strong_aug: 18 | num_augs: 3 # !!!!!!!!!!!!!!!!!!!! key param1 (4): 1, 2, 3, 4 !!!!!!!!!!!!!!!!!!!! 19 | flag_use_random_num_sampling: True # !!!!!!!!!!!!!!!!!!!! key param1 (2): True, False !!!!!!!!!!!!!!!!!!!! 20 | val: 21 | data_root: ./data/VOC2012 22 | data_list: ./data/splitsall/pascal_u2pl/val.txt 23 | batch_size: 1 24 | n_sup: 662 25 | workers: 4 26 | mean: [0.485, 0.456, 0.406] 27 | std: [0.229, 0.224, 0.225] 28 | ignore_label: 255 29 | 30 | # # # # # # # # # # # # # # 31 | # 2. training params 32 | # # # # # # # # # # # # # # 33 | trainer: # Required. 34 | epochs: 80 35 | sup_only_epoch: 0 # 0, 1 36 | evaluate_student: True 37 | optimizer: 38 | type: SGD 39 | kwargs: 40 | lr: 0.001 # 0.001:8*2gpus, 4*4gpus; 0.002:8*4gpus. 41 | momentum: 0.9 42 | weight_decay: 0.0001 43 | lr_scheduler: 44 | mode: poly 45 | kwargs: 46 | power: 0.9 47 | # # # # # # # # # # # # # # 48 | # unsupervised loss 49 | # # # # # # # # # # # # # # 50 | unsupervised: 51 | flag_extra_weak: False 52 | threshold: 0.95 # ============================================================================= <>: [0.7, 0.75, 0.8, 0.85, 0.9, 0.95] 53 | loss_weight: 1.0 # ============================================================================= << abl2-weight >> : [1.0, 1.5, 2.0, 2.5, 3.0, 4.0] 54 | #------ 2)cutmix augs ------# 55 | use_cutmix: True 56 | use_cutmix_adaptive: True 57 | use_cutmix_trigger_prob: 1.0 58 | 59 | # # # # # # # # # # # # # # 60 | # 3. output files, and loss 61 | # # # # # # # # # # # # # # 62 | saver: 63 | snapshot_dir: checkpoints 64 | pretrain: '' 65 | use_tb: False 66 | 67 | criterion: 68 | type: CELoss 69 | kwargs: 70 | use_weight: False 71 | 72 | # # # # # # # # # # # # # # 73 | # 4. models 74 | # # # # # # # # # # # # # # 75 | net: # Required. 76 | num_classes: 21 77 | sync_bn: True 78 | ema_decay: 0.999 79 | encoder: 80 | # type: augseg.models.resnet.resnet50 81 | # pretrain: ./pretrained/resnet50.pth 82 | type: augseg.models.resnet.resnet101 83 | pretrain: ./pretrained/resnet101.pth 84 | kwargs: 85 | zero_init_residual: True 86 | multi_grid: True 87 | replace_stride_with_dilation: [False, False, True] # [output_stride = 16] 88 | # replace_stride_with_dilation: [False, True, True] # [output_stride = 8] 89 | decoder: 90 | type: augseg.models.decoder.dec_deeplabv3_plus 91 | kwargs: 92 | inner_planes: 256 93 | low_conv_planes: 48 94 | dilations: [6, 12, 18] # [output_stride = 16] 95 | # dilations: [12, 24, 36] # [output_stride = 8] 96 | -------------------------------------------------------------------------------- /requirements.txt: -------------------------------------------------------------------------------- 1 | pyyaml>=5.4.0 2 | easydict 3 | tensorboard 4 | opencv-python 5 | scipy 6 | scikit-learn 7 | scikit-image 8 | einops 9 | h5py 10 | pandas 11 | tqdm 12 | simpleitk 13 | medpy 14 | nibabel 15 | -------------------------------------------------------------------------------- /scripts/bashtorch: -------------------------------------------------------------------------------- 1 | # bashtorch file 2 | # export TORCH_HOME=./envs/torch 3 | export TORCH_HOME=./envs/a100torch 4 | export PATH=${TORCH_HOME}/bin:$PATH 5 | export LD_LIBRARY_PATH=${TORCH_HOME}/lib/:${LD_LIBRARY_PATH} 6 | -------------------------------------------------------------------------------- /scripts/zsing_run_citys.sh: -------------------------------------------------------------------------------- 1 | tport=52009 2 | ngpu=4 3 | ROOT=. 4 | 5 | # CUDA_VISIBLE_DEVICES=4,5,6,7 \ 6 | python -m torch.distributed.launch \ 7 | --nproc_per_node=${ngpu} \ 8 | --node_rank=0 \ 9 | --master_port=${tport} \ 10 | $ROOT/train_semi.py \ 11 | --config=$ROOT/exps/zrun_citys/citys_semi744/config_semi.yaml --seed 2 --port ${tport} 12 | 13 | # --- --- --- 14 | # --config=$ROOT/exps/zrun_citys/citys_semi186/config_semi.yaml --seed 2 --port ${tport} 15 | # --config=$ROOT/exps/zrun_citys/citys_semi372/config_semi.yaml --seed 2 --port ${tport} 16 | # --config=$ROOT/exps/zrun_citys/citys_semi744/config_semi.yaml --seed 2 --port ${tport} 17 | # --config=$ROOT/exps/zrun_citys/citys_semi1488/config_semi.yaml --seed 2 --port ${tport} 18 | 19 | # --config=$ROOT/exps/zrun_citys/r50_citys_semi186/config_semi.yaml --seed 2 --port ${tport} 20 | # --config=$ROOT/exps/zrun_citys/r50_citys_semi372/config_semi.yaml --seed 2 --port ${tport} 21 | # --config=$ROOT/exps/zrun_citys/r50_citys_semi744/config_semi.yaml --seed 2 --port ${tport} 22 | # --config=$ROOT/exps/zrun_citys/r50_citys_semi1488/config_semi.yaml --seed 2 --port ${tport} 23 | -------------------------------------------------------------------------------- /scripts/zsing_run_voc.sh: -------------------------------------------------------------------------------- 1 | tport=53907 2 | ngpu=2 3 | ROOT=. 4 | 5 | # CUDA_VISIBLE_DEVICES=4,5, 6 | python -m torch.distributed.launch \ 7 | --nproc_per_node=${ngpu} \ 8 | --node_rank=0 \ 9 | --master_port=${tport} \ 10 | $ROOT/train_semi.py \ 11 | --config=$ROOT/exps/zrun_vocs_u2pl/voc_semi662/config_semi.yaml --seed 2 --port ${tport} 12 | 13 | # ---- ----- 14 | # --config=$ROOT/exps/zrun_vocs/voc_semi_fine92/config_semi.yaml --seed 2 --port ${tport} 15 | # --config=$ROOT/exps/zrun_vocs/voc_semi_fine183/config_semi.yaml --seed 2 --port ${tport} 16 | # --config=$ROOT/exps/zrun_vocs/voc_semi_fine366/config_semi.yaml --seed 2 --port ${tport} 17 | # --config=$ROOT/exps/zrun_vocs/voc_semi_fine732/config_semi.yaml --seed 2 --port ${tport} 18 | # --config=$ROOT/exps/zrun_vocs/voc_semi_fine1464/config_semi.yaml --seed 2 --port ${tport} 19 | 20 | # --config=$ROOT/exps/zrun_vocs/voc_semi662/config_semi.yaml --seed 2 --port ${tport} 21 | # --config=$ROOT/exps/zrun_vocs/voc_semi1323/config_semi.yaml --seed 2 --port ${tport} 22 | # --config=$ROOT/exps/zrun_vocs/voc_semi2646/config_semi.yaml --seed 2 --port ${tport} 23 | # --config=$ROOT/exps/zrun_vocs/voc_semi5291/config_semi.yaml --seed 2 --port ${tport} 24 | 25 | # --config=$ROOT/exps/zrun_vocs/r50_voc_semi662/config_semi.yaml --seed 2 --port ${tport} 26 | # --config=$ROOT/exps/zrun_vocs/r50_voc_semi1323/config_semi.yaml --seed 2 --port ${tport} 27 | # --config=$ROOT/exps/zrun_vocs/r50_voc_semi2646/config_semi.yaml --seed 2 --port ${tport} 28 | # --config=$ROOT/exps/zrun_vocs/r50_voc_semi5291/config_semi.yaml --seed 2 --port ${tport} 29 | 30 | # ---- ---- u2pl 31 | # --config=$ROOT/exps/zrun_vocs_u2pl/voc_semi662/config_semi.yaml --seed 2 --port ${tport} 32 | # --config=$ROOT/exps/zrun_vocs_u2pl/voc_semi1323/config_semi.yaml --seed 2 --port ${tport} 33 | # --config=$ROOT/exps/zrun_vocs_u2pl/voc_semi2646/config_semi.yaml --seed 2 --port ${tport} 34 | # --config=$ROOT/exps/zrun_vocs_u2pl/voc_semi5291/config_semi.yaml --seed 2 --port ${tport} 35 | 36 | # --config=$ROOT/exps/zrun_vocs_u2pl/r50_voc_semi662/config_semi.yaml --seed 2 --port ${tport} 37 | # --config=$ROOT/exps/zrun_vocs_u2pl/r50_voc_semi1323/config_semi.yaml --seed 2 --port ${tport} 38 | # --config=$ROOT/exps/zrun_vocs_u2pl/r50_voc_semi2646/config_semi.yaml --seed 2 --port ${tport} 39 | # --config=$ROOT/exps/zrun_vocs_u2pl/r50_voc_semi5291/config_semi.yaml --seed 2 --port ${tport} -------------------------------------------------------------------------------- /single_run.sh: -------------------------------------------------------------------------------- 1 | tport=53907 2 | ngpu=2 3 | ROOT=. 4 | 5 | CUDA_VISIBLE_DEVICES=0,1 \ 6 | python -m torch.distributed.launch \ 7 | --nproc_per_node=${ngpu} \ 8 | --node_rank=0 \ 9 | --master_port=${tport} \ 10 | $ROOT/train_semi.py \ 11 | --config=$ROOT/exps/zrun_vocs_u2pl/voc_semi662/config_semi.yaml --seed 2 --port ${tport} 12 | -------------------------------------------------------------------------------- /train_semi.py: -------------------------------------------------------------------------------- 1 | import argparse 2 | import yaml 3 | import os, sys 4 | import os.path as osp 5 | import pprint 6 | import time 7 | import pickle 8 | 9 | import random 10 | import numpy as np 11 | import pandas as pd 12 | import torch 13 | import torch.backends.cudnn as cudnn 14 | import torch.distributed as dist 15 | import torch.nn.functional as F 16 | from torch.utils.tensorboard import SummaryWriter 17 | 18 | from augseg.dataset.augs_ALIA import cut_mix_label_adaptive 19 | from augseg.dataset.builder import get_loader 20 | from augseg.models.model_helper import ModelBuilder 21 | from augseg.utils.dist_helper import setup_distributed 22 | from augseg.utils.loss_helper import get_criterion, compute_unsupervised_loss_by_threshold 23 | from augseg.utils.lr_helper import get_optimizer, get_scheduler 24 | from augseg.utils.utils import AverageMeter, intersectionAndUnion, load_state 25 | from augseg.utils.utils import init_log, get_rank, get_world_size, set_random_seed, setup_default_logging 26 | 27 | import warnings 28 | warnings.filterwarnings('ignore') 29 | 30 | 31 | def main(in_args): 32 | args = in_args 33 | if args.seed is not None: 34 | # print("set random seed to", args.seed) 35 | set_random_seed(args.seed, deterministic=True) 36 | # set_random_seed(args.seed) 37 | cfg = yaml.load(open(args.config, "r"), Loader=yaml.Loader) 38 | rank, word_size = setup_distributed(port=args.port) 39 | 40 | ########################### 41 | # 1. output settings 42 | ########################### 43 | cfg["exp_path"] = osp.dirname(args.config) 44 | cfg["save_path"] = osp.join(cfg["exp_path"], cfg["saver"]["snapshot_dir"]) 45 | cfg["log_path"] = osp.join(cfg["exp_path"], "log") 46 | flag_use_tb = cfg["saver"]["use_tb"] 47 | 48 | if not os.path.exists(cfg["log_path"]) and rank == 0: 49 | os.makedirs(cfg["log_path"]) 50 | if not osp.exists(cfg["save_path"]) and rank == 0: 51 | os.makedirs(cfg["save_path"]) 52 | # my favorate: logs 53 | if rank == 0: 54 | logger, curr_timestr = setup_default_logging("global", cfg["log_path"]) 55 | csv_path = os.path.join(cfg["log_path"], "seg_{}_stat.csv".format(curr_timestr)) 56 | else: 57 | logger, curr_timestr = None, "" 58 | csv_path = None 59 | # tensorboard 60 | if rank == 0: 61 | logger.info("{}".format(pprint.pformat(cfg))) 62 | if flag_use_tb: 63 | tb_logger = SummaryWriter( 64 | osp.join(cfg["log_path"], "events_seg",curr_timestr) 65 | ) 66 | else: 67 | tb_logger = None 68 | else: 69 | tb_logger = None 70 | # make sure all folders and csv handler are correctly created on rank ==0. 71 | dist.barrier() 72 | 73 | ########################### 74 | # 2. prepare model 1 75 | ########################### 76 | model = ModelBuilder(cfg["net"]) 77 | modules_back = [model.encoder] 78 | modules_head = [model.decoder] 79 | if cfg["net"].get("aux_loss", False): 80 | modules_head.append(model.auxor) 81 | if cfg["net"].get("sync_bn", True): 82 | model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model) 83 | model.cuda() 84 | 85 | ########################### 86 | # 3. data 87 | ########################### 88 | sup_loss_fn = get_criterion(cfg) 89 | train_loader_sup, train_loader_unsup, val_loader = get_loader(cfg, seed=args.seed) 90 | 91 | ############################## 92 | # 4. optimizer & scheduler 93 | ############################## 94 | cfg_trainer = cfg["trainer"] 95 | cfg_optim = cfg_trainer["optimizer"] 96 | times = 10 if "pascal" in cfg["dataset"]["type"] else 1 97 | 98 | params_list = [] 99 | for module in modules_back: 100 | params_list.append( 101 | dict(params=module.parameters(), lr=cfg_optim["kwargs"]["lr"]) 102 | ) 103 | for module in modules_head: 104 | params_list.append( 105 | dict(params=module.parameters(), lr=cfg_optim["kwargs"]["lr"] * times) 106 | ) 107 | optimizer = get_optimizer(params_list, cfg_optim) 108 | 109 | ########################### 110 | # 5. prepare model more 111 | ########################### 112 | local_rank = int(os.environ["LOCAL_RANK"]) 113 | model = torch.nn.parallel.DistributedDataParallel( 114 | model, 115 | device_ids=[local_rank], 116 | output_device=local_rank, 117 | find_unused_parameters=False, 118 | ) 119 | 120 | # Teacher model -- freeze training 121 | model_teacher = ModelBuilder(cfg["net"]) 122 | model_teacher.cuda() 123 | model_teacher = torch.nn.parallel.DistributedDataParallel( 124 | model_teacher, 125 | device_ids=[local_rank], 126 | output_device=local_rank, 127 | find_unused_parameters=False, 128 | ) 129 | for p in model_teacher.parameters(): 130 | p.requires_grad = False 131 | 132 | # initialize teacher model -- not neccesary if using warmup 133 | with torch.no_grad(): 134 | for t_params, s_params in zip(model_teacher.parameters(), model.parameters()): 135 | t_params.data = s_params.data 136 | 137 | ###################################### 138 | # 6. resume 139 | ###################################### 140 | last_epoch = 0 141 | best_prec = 0 142 | best_epoch = -1 143 | best_prec_stu = 0 144 | best_epoch_stu = -1 145 | # auto_resume > pretrain 146 | if cfg["saver"].get("auto_resume", False): 147 | lastest_model = os.path.join(cfg["save_path"], "ckpt.pth") 148 | if not os.path.exists(lastest_model): 149 | "No checkpoint found in '{}'".format(lastest_model) 150 | else: 151 | print(f"Resume model from: '{lastest_model}'") 152 | best_prec, last_epoch = load_state( 153 | lastest_model, model, optimizer=optimizer, key="model_state" 154 | ) 155 | _, _ = load_state( 156 | lastest_model, model_teacher, optimizer=optimizer, key="teacher_state" 157 | ) 158 | 159 | optimizer_start = get_optimizer(params_list, cfg_optim) 160 | lr_scheduler = get_scheduler( 161 | cfg_trainer, len(train_loader_sup), optimizer_start, start_epoch=last_epoch 162 | ) 163 | 164 | ###################################### 165 | # 7. training loop 166 | ###################################### 167 | if rank == 0: 168 | logger.info('-------------------------- start training --------------------------') 169 | # Start to train model 170 | for epoch in range(last_epoch, cfg_trainer["epochs"]): 171 | # Training 172 | res_loss_sup, res_loss_unsup = train( 173 | model, 174 | model_teacher, 175 | optimizer, 176 | lr_scheduler, 177 | sup_loss_fn, 178 | train_loader_sup, 179 | train_loader_unsup, 180 | epoch, 181 | tb_logger, 182 | logger, 183 | cfg 184 | ) 185 | 186 | # Validation and store checkpoint 187 | if "cityscapes" in cfg["dataset"].get("type", "pascal"): 188 | if epoch % 10 == 0 or epoch > (cfg_trainer["epochs"]-50): 189 | if cfg_trainer.get("evaluate_student", True): 190 | prec_stu = validate_citys(model, val_loader, epoch, logger, cfg) 191 | else: 192 | prec_stu =-1000.0 193 | prec_tea = validate_citys(model_teacher, val_loader, epoch, logger, cfg) 194 | prec = prec_tea 195 | else: 196 | prec_stu = -1000.0 197 | prec_tea = -1000.0 198 | prec = prec_tea 199 | else: 200 | if cfg_trainer.get("evaluate_student", True): 201 | prec_stu = validate(model, val_loader, epoch, logger, cfg) 202 | else: 203 | prec_stu = -1000.0 204 | prec_tea = validate(model_teacher, val_loader, epoch, logger, cfg) 205 | prec = prec_tea 206 | 207 | if rank == 0: 208 | state = { 209 | "epoch": epoch + 1, 210 | "model_state": model.state_dict(), 211 | "optimizer_state": optimizer.state_dict(), 212 | "teacher_state": model_teacher.state_dict(), 213 | "best_miou": best_prec, 214 | } 215 | if prec_stu > best_prec_stu: 216 | best_prec_stu = prec_stu 217 | best_epoch_stu = epoch 218 | 219 | if prec > best_prec: 220 | best_prec = prec 221 | best_epoch = epoch 222 | state["best_miou"] = prec 223 | torch.save(state, osp.join(cfg["save_path"], "ckpt_best.pth")) 224 | 225 | torch.save(state, osp.join(cfg["save_path"], "ckpt.pth")) 226 | # save statistics 227 | tmp_results = { 228 | 'loss_lb': res_loss_sup, 229 | 'loss_ub': res_loss_unsup, 230 | 'miou_stu': prec_stu, 231 | 'miou_tea': prec_tea, 232 | "best": best_prec, 233 | "best-stu":best_prec_stu} 234 | data_frame = pd.DataFrame(data=tmp_results, index=range(epoch, epoch+1)) 235 | if epoch > 0 and osp.exists(csv_path): 236 | data_frame.to_csv(csv_path, mode='a', header=None, index_label='epoch') 237 | else: 238 | data_frame.to_csv(csv_path, index_label='epoch') 239 | 240 | logger.info(" <> - Epoch: {}. MIoU: {:.2f}/{:.2f}. \033[34mBest-STU:{:.2f}/{} \033[31mBest-EMA: {:.2f}/{}\033[0m".format(epoch, 241 | prec_stu * 100, prec_tea * 100, best_prec_stu * 100, best_epoch_stu, best_prec * 100, best_epoch)) 242 | if tb_logger is not None: 243 | tb_logger.add_scalar("mIoU val", prec, epoch) 244 | 245 | 246 | 247 | 248 | 249 | def train( 250 | model, 251 | model_teacher, 252 | optimizer, 253 | lr_scheduler, 254 | sup_loss_fn, 255 | loader_l, 256 | loader_u, 257 | epoch, 258 | tb_logger, 259 | logger, 260 | cfg, 261 | ): 262 | 263 | ema_decay_origin = cfg["net"]["ema_decay"] 264 | rank, world_size = dist.get_rank(), dist.get_world_size() 265 | flag_extra_weak = cfg["trainer"]["unsupervised"].get("flag_extra_weak", False) 266 | model.train() 267 | 268 | # data loader 269 | loader_l.sampler.set_epoch(epoch) 270 | loader_u.sampler.set_epoch(epoch) 271 | loader_l_iter = iter(loader_l) 272 | loader_u_iter = iter(loader_u) 273 | assert len(loader_l) == len(loader_u), f"labeled data {len(loader_l)} unlabeled data {len(loader_u)}, mixmatch!" 274 | 275 | # metric indicators 276 | sup_losses = AverageMeter(20) 277 | uns_losses = AverageMeter(20) 278 | batch_times = AverageMeter(20) 279 | learning_rates = AverageMeter(20) 280 | meter_high_pseudo_ratio = AverageMeter(20) 281 | 282 | # print freq 8 times for a epoch 283 | print_freq = len(loader_u) // 8 # 8 for semi 4 for sup 284 | print_freq_lst = [i * print_freq for i in range(1,8)] 285 | print_freq_lst.append(len(loader_u) -1) 286 | 287 | # start iterations 288 | model.train() 289 | model_teacher.eval() 290 | for step in range(len(loader_l)): 291 | batch_start = time.time() 292 | 293 | i_iter = epoch * len(loader_l) + step # total iters till now 294 | lr = lr_scheduler.get_lr() 295 | learning_rates.update(lr[0]) 296 | lr_scheduler.step() # lr is updated at the iteration level 297 | 298 | # obtain labeled and unlabeled data 299 | _, image_l, label_l = loader_l_iter.next() 300 | image_l, label_l = image_l.cuda(), label_l.cuda() 301 | _, image_u_weak, image_u_aug, _ = loader_u_iter.next() 302 | image_u_weak, image_u_aug = image_u_weak.cuda(), image_u_aug.cuda() 303 | 304 | # start the training 305 | if epoch < cfg["trainer"].get("sup_only_epoch", 0): 306 | # forward 307 | pred, aux = model(image_l) 308 | # supervised loss 309 | if "aux_loss" in cfg["net"].keys(): 310 | sup_loss = sup_loss_fn([pred, aux], label_l) 311 | del aux 312 | else: 313 | sup_loss = sup_loss_fn(pred, label_l) 314 | del pred 315 | 316 | # no unlabeled data during the warmup period 317 | unsup_loss = torch.tensor(0.0).cuda() 318 | pseduo_high_ratio = torch.tensor(0.0).cuda() 319 | 320 | else: 321 | # 1. generate pseudo labels 322 | p_threshold = cfg["trainer"]["unsupervised"].get("threshold", 0.95) 323 | with torch.no_grad(): 324 | model_teacher.eval() 325 | pred_u, _ = model_teacher(image_u_weak.detach()) 326 | pred_u = F.softmax(pred_u, dim=1) 327 | # obtain pseudos 328 | logits_u_aug, label_u_aug = torch.max(pred_u, dim=1) 329 | 330 | # obtain confidence 331 | entropy = -torch.sum(pred_u * torch.log(pred_u + 1e-10), dim=1) 332 | entropy /= np.log(cfg["net"]["num_classes"]) 333 | confidence = 1.0 - entropy 334 | confidence = confidence * logits_u_aug 335 | confidence = confidence.mean(dim=[1,2]) # 1*C 336 | confidence = confidence.cpu().numpy().tolist() 337 | # confidence = logits_u_aug.ge(p_threshold).float().mean(dim=[1,2]).cpu().numpy().tolist() 338 | del pred_u 339 | model.train() 340 | 341 | # 2. apply cutmix 342 | trigger_prob = cfg["trainer"]["unsupervised"].get("use_cutmix_trigger_prob", 1.0) 343 | if np.random.uniform(0, 1) < trigger_prob and cfg["trainer"]["unsupervised"].get("use_cutmix", False): 344 | if cfg["trainer"]["unsupervised"].get("use_cutmix_adaptive", False): 345 | image_u_aug, label_u_aug, logits_u_aug = cut_mix_label_adaptive( 346 | image_u_aug, 347 | label_u_aug, 348 | logits_u_aug, 349 | image_l, 350 | label_l, 351 | confidence 352 | ) 353 | 354 | # 3. forward concate labeled + unlabeld into student networks 355 | num_labeled = len(image_l) 356 | if flag_extra_weak: 357 | pred_all, aux_all = model(torch.cat((image_l, image_u_weak, image_u_aug), dim=0)) 358 | del image_l, image_u_weak, image_u_aug 359 | pred_l= pred_all[:num_labeled] 360 | _, pred_u_strong = pred_all[num_labeled:].chunk(2) 361 | del pred_all 362 | else: 363 | pred_all, aux_all = model(torch.cat((image_l, image_u_aug), dim=0)) 364 | del image_l, image_u_weak, image_u_aug 365 | pred_l= pred_all[:num_labeled] 366 | pred_u_strong = pred_all[num_labeled:] 367 | del pred_all 368 | 369 | # 4. supervised loss 370 | if "aux_loss" in cfg["net"].keys(): 371 | aux = aux_all[:num_labeled] 372 | sup_loss = sup_loss_fn([pred_l, aux], label_l) 373 | del aux_all, aux 374 | else: 375 | sup_loss = sup_loss_fn(pred_l, label_l) 376 | 377 | # 5. unsupervised loss 378 | unsup_loss, pseduo_high_ratio = compute_unsupervised_loss_by_threshold( 379 | pred_u_strong, label_u_aug.detach(), 380 | logits_u_aug.detach(), thresh=p_threshold) 381 | unsup_loss *= cfg["trainer"]["unsupervised"].get("loss_weight", 1.0) 382 | del pred_l, pred_u_strong, label_u_aug, logits_u_aug 383 | 384 | loss = sup_loss + unsup_loss 385 | 386 | # update student model 387 | optimizer.zero_grad() 388 | loss.backward() 389 | optimizer.step() 390 | 391 | # update teacher model with EMA 392 | with torch.no_grad(): 393 | if epoch > cfg["trainer"].get("sup_only_epoch", 0): 394 | ema_decay = min( 395 | 1 396 | - 1 397 | / ( 398 | i_iter 399 | - len(loader_l) * cfg["trainer"].get("sup_only_epoch", 0) 400 | + 1 401 | ), 402 | ema_decay_origin, 403 | ) 404 | else: 405 | ema_decay = 0.0 406 | # update weight 407 | for param_train, param_eval in zip(model.parameters(), model_teacher.parameters()): 408 | param_eval.data = param_eval.data * ema_decay + param_train.data * (1 - ema_decay) 409 | # update bn 410 | for buffer_train, buffer_eval in zip(model.buffers(), model_teacher.buffers()): 411 | buffer_eval.data = buffer_eval.data * ema_decay + buffer_train.data * (1 - ema_decay) 412 | # buffer_eval.data = buffer_train.data 413 | 414 | # gather all loss from different gpus 415 | reduced_sup_loss = sup_loss.clone().detach() 416 | dist.all_reduce(reduced_sup_loss) 417 | sup_losses.update(reduced_sup_loss.item() / world_size) 418 | 419 | reduced_uns_loss = unsup_loss.clone().detach() 420 | dist.all_reduce(reduced_uns_loss) 421 | uns_losses.update(reduced_uns_loss.item() / world_size) 422 | 423 | reduced_pseudo_high_ratio = pseduo_high_ratio.clone().detach() 424 | dist.all_reduce(reduced_pseudo_high_ratio) 425 | meter_high_pseudo_ratio.update(reduced_pseudo_high_ratio.item() / world_size) 426 | 427 | # 12. print log information 428 | batch_end = time.time() 429 | batch_times.update(batch_end - batch_start) 430 | # if i_iter % 10 == 0 and rank == 0: 431 | if step in print_freq_lst and rank == 0: 432 | logger.info( 433 | "Epoch/Iter [{}:{:3}/{:3}]. " 434 | "Sup:{sup_loss.val:.3f}({sup_loss.avg:.3f}) " 435 | "Uns:{uns_loss.val:.3f}({uns_loss.avg:.3f}) " 436 | "Pseudo:{high_ratio.val:.3f}({high_ratio.avg:.3f}) " 437 | "Time:{batch_time.avg:.2f} " 438 | "LR:{lr.val:.5f}".format( 439 | cfg["trainer"]["epochs"], epoch, step, 440 | sup_loss=sup_losses, 441 | uns_loss=uns_losses, 442 | high_ratio=meter_high_pseudo_ratio, 443 | batch_time=batch_times, 444 | lr=learning_rates, 445 | ) 446 | ) 447 | if tb_logger is not None: 448 | tb_logger.add_scalar("lr", learning_rates.avg, i_iter) 449 | tb_logger.add_scalar("Sup Loss", sup_losses.avg, i_iter) 450 | tb_logger.add_scalar("Uns Loss", uns_losses.avg, i_iter) 451 | tb_logger.add_scalar("High ratio", meter_high_pseudo_ratio.avg, i_iter) 452 | 453 | return sup_losses.avg, uns_losses.avg 454 | 455 | 456 | def validate( 457 | model, 458 | data_loader, 459 | epoch, 460 | logger, 461 | cfg 462 | ): 463 | model.eval() 464 | data_loader.sampler.set_epoch(epoch) 465 | 466 | num_classes, ignore_label = ( 467 | cfg["net"]["num_classes"], 468 | cfg["dataset"]["ignore_label"], 469 | ) 470 | rank, world_size = dist.get_rank(), dist.get_world_size() 471 | 472 | intersection_meter = AverageMeter() 473 | union_meter = AverageMeter() 474 | 475 | for step, batch in enumerate(data_loader): 476 | _, images, labels = batch 477 | images = images.cuda() 478 | labels = labels.long().cuda() 479 | 480 | with torch.no_grad(): 481 | output, _ = model(images) 482 | 483 | # get the output produced by model_teacher 484 | output = output.data.max(1)[1].cpu().numpy() 485 | target_origin = labels.cpu().numpy() 486 | 487 | # start to calculate miou 488 | intersection, union, target = intersectionAndUnion( 489 | output, target_origin, num_classes, ignore_label 490 | ) 491 | 492 | # gather all validation information 493 | reduced_intersection = torch.from_numpy(intersection).cuda() 494 | reduced_union = torch.from_numpy(union).cuda() 495 | reduced_target = torch.from_numpy(target).cuda() 496 | 497 | dist.all_reduce(reduced_intersection) 498 | dist.all_reduce(reduced_union) 499 | dist.all_reduce(reduced_target) 500 | 501 | intersection_meter.update(reduced_intersection.cpu().numpy()) 502 | union_meter.update(reduced_union.cpu().numpy()) 503 | 504 | iou_class = intersection_meter.sum / (union_meter.sum + 1e-10) 505 | mIoU = np.mean(iou_class) 506 | 507 | if rank == 0: 508 | for i, iou in enumerate(iou_class): 509 | logger.info(" [Test] - class [{}] IoU {:.2f}".format(i, iou * 100)) 510 | 511 | return mIoU 512 | 513 | 514 | def validate_citys( 515 | model, 516 | data_loader, 517 | epoch, 518 | logger, 519 | cfg 520 | ): 521 | model.eval() 522 | data_loader.sampler.set_epoch(epoch) 523 | rank, world_size = dist.get_rank(), dist.get_world_size() 524 | 525 | num_classes = cfg["net"]["num_classes"] 526 | ignore_label = cfg["dataset"]["ignore_label"] 527 | if cfg["dataset"]["val"].get("crop", False): 528 | crop_size, _ = cfg["dataset"]["val"]["crop"].get("size", [800, 800]) 529 | else: 530 | crop_size = 800 531 | 532 | intersection_meter = AverageMeter() 533 | union_meter = AverageMeter() 534 | 535 | for step, batch in enumerate(data_loader): 536 | _, images, labels = batch 537 | images = images.cuda() 538 | labels = labels.long() 539 | batch_size, h, w = labels.shape 540 | 541 | with torch.no_grad(): 542 | final = torch.zeros(batch_size, num_classes, h, w).cuda() 543 | row = 0 544 | while row < h: 545 | col = 0 546 | while col < w: 547 | pred, _ = model(images[:, :, row: min(h, row + crop_size), col: min(w, col + crop_size)]) 548 | final[:, :, row: min(h, row + crop_size), col: min(w, col + crop_size)] += pred.softmax(dim=1) 549 | col += int(crop_size * 2 / 3) 550 | row += int(crop_size * 2 / 3) 551 | # get the output 552 | output = final.argmax(dim=1).cpu().numpy() 553 | target_origin = labels.numpy() 554 | # print("="*50, output.shape, output.dtype, target_origin.shape, target_origin.dtype) 555 | 556 | # start to calculate miou 557 | intersection, union, target = intersectionAndUnion( 558 | output, target_origin, num_classes, ignore_label 559 | ) 560 | # # return ndarray, b*clas 561 | # print("="*20, type(intersection), type(union), type(target), intersection, union, target) 562 | 563 | # gather all validation information 564 | reduced_intersection = torch.from_numpy(intersection).cuda() 565 | reduced_union = torch.from_numpy(union).cuda() 566 | reduced_target = torch.from_numpy(target).cuda() 567 | 568 | dist.all_reduce(reduced_intersection) 569 | dist.all_reduce(reduced_union) 570 | dist.all_reduce(reduced_target) 571 | 572 | intersection_meter.update(reduced_intersection.cpu().numpy()) 573 | union_meter.update(reduced_union.cpu().numpy()) 574 | 575 | iou_class = intersection_meter.sum / (union_meter.sum + 1e-10) 576 | mIoU = np.mean(iou_class) 577 | 578 | if rank == 0: 579 | for i, iou in enumerate(iou_class): 580 | logger.info(" [Test] - class [{}] IoU {:.2f}".format(i, iou * 100)) 581 | return mIoU 582 | 583 | 584 | if __name__ == "__main__": 585 | parser = argparse.ArgumentParser(description="Semi-Supervised Semantic Segmentation") 586 | parser.add_argument("--config", type=str, default="config.yaml") 587 | parser.add_argument("--local_rank", type=int, default=0) 588 | parser.add_argument("--seed", type=int, default=0) 589 | parser.add_argument("--port", default=None, type=int) 590 | args = parser.parse_args() 591 | main(args) 592 | -------------------------------------------------------------------------------- /training-logs/3 Semi-Voc-fine/1464/config_semi.yaml: -------------------------------------------------------------------------------- 1 | # # # # # # # # # # # # # # 2 | # 1. datasets 3 | # # # # # # # # # # # # # # 4 | dataset: # Required. 5 | type: pascal_semi 6 | train: 7 | data_root: ./data/VOC2012 8 | data_list: ./data/splitsall/pascal/1464/labeled.txt 9 | batch_size: 8 10 | flip: True 11 | rand_resize: [0.5, 2.0] 12 | resize_base_size: 500 13 | crop: 14 | type: rand 15 | size: [513, 513] 16 | #---- 1) strong data augs ----# 17 | strong_aug: 18 | num_augs: 3 # !!!!!!!!!!!!!!!!!!!! key param1 (4): 1, 2, 3, 4 !!!!!!!!!!!!!!!!!!!! 19 | flag_use_random_num_sampling: True # !!!!!!!!!!!!!!!!!!!! key param1 (2): True, False !!!!!!!!!!!!!!!!!!!! 20 | val: 21 | data_root: ./data/VOC2012 22 | data_list: ./data/splitsall/pascal/val.txt 23 | batch_size: 1 24 | n_sup: 1464 25 | workers: 4 26 | mean: [0.485, 0.456, 0.406] 27 | std: [0.229, 0.224, 0.225] 28 | ignore_label: 255 29 | 30 | # # # # # # # # # # # # # # 31 | # 2. training params 32 | # # # # # # # # # # # # # # 33 | trainer: # Required. 34 | epochs: 80 35 | sup_only_epoch: 0 # 0, 1 36 | evaluate_student: True 37 | optimizer: 38 | type: SGD 39 | kwargs: 40 | lr: 0.001 # 0.001:8*2gpus, 4*4gpus; 0.002:8*4gpus. 41 | momentum: 0.9 42 | weight_decay: 0.0001 43 | lr_scheduler: 44 | mode: poly 45 | kwargs: 46 | power: 0.9 47 | # # # # # # # # # # # # # # 48 | # unsupervised loss 49 | # # # # # # # # # # # # # # 50 | unsupervised: 51 | flag_extra_weak: False 52 | threshold: 0.95 # ============================================================================= <>: [0.7, 0.75, 0.8, 0.85, 0.9, 0.95] 53 | loss_weight: 1.0 # ============================================================================= << abl2-weight >> : [1.0, 1.5, 2.0, 2.5, 3.0, 4.0] 54 | #------ 2)cutmix augs ------# 55 | use_cutmix: True 56 | use_cutmix_adaptive: True 57 | use_cutmix_trigger_prob: 1.0 58 | 59 | # # # # # # # # # # # # # # 60 | # 3. output files, and loss 61 | # # # # # # # # # # # # # # 62 | saver: 63 | snapshot_dir: checkpoints 64 | pretrain: '' 65 | use_tb: False 66 | 67 | criterion: 68 | type: CELoss 69 | kwargs: 70 | use_weight: False 71 | 72 | # # # # # # # # # # # # # # 73 | # 4. models 74 | # # # # # # # # # # # # # # 75 | net: # Required. 76 | num_classes: 21 77 | sync_bn: True 78 | ema_decay: 0.999 79 | encoder: 80 | # type: augseg.models.resnet.resnet50 81 | # pretrain: ./pretrained/resnet50.pth 82 | type: augseg.models.resnet.resnet101 83 | pretrain: ./pretrained/resnet101.pth 84 | kwargs: 85 | zero_init_residual: True 86 | multi_grid: True 87 | replace_stride_with_dilation: [False, False, True] # [output_stride = 16] 88 | # replace_stride_with_dilation: [False, True, True] # [output_stride = 8] 89 | decoder: 90 | type: augseg.models.decoder.dec_deeplabv3_plus 91 | kwargs: 92 | inner_planes: 256 93 | low_conv_planes: 48 94 | dilations: [6, 12, 18] # [output_stride = 16] 95 | # dilations: [12, 24, 36] # [output_stride = 8] 96 | -------------------------------------------------------------------------------- /training-logs/3 Semi-Voc-fine/183/config_semi.yaml: -------------------------------------------------------------------------------- 1 | # # # # # # # # # # # # # # 2 | # 1. datasets 3 | # # # # # # # # # # # # # # 4 | dataset: # Required. 5 | type: pascal_semi 6 | train: 7 | data_root: ./data/VOC2012 8 | data_list: ./data/splitsall/pascal/183/labeled.txt 9 | batch_size: 8 10 | flip: True 11 | rand_resize: [0.5, 2.0] 12 | resize_base_size: 500 13 | crop: 14 | type: rand 15 | size: [513, 513] 16 | #---- 1) strong data augs ----# 17 | strong_aug: 18 | num_augs: 3 # !!!!!!!!!!!!!!!!!!!! key param1 (4): 1, 2, 3, 4 !!!!!!!!!!!!!!!!!!!! 19 | flag_use_random_num_sampling: True # !!!!!!!!!!!!!!!!!!!! key param1 (2): True, False !!!!!!!!!!!!!!!!!!!! 20 | val: 21 | data_root: ./data/VOC2012 22 | data_list: ./data/splitsall/pascal/val.txt 23 | batch_size: 1 24 | n_sup: 183 25 | workers: 4 26 | mean: [0.485, 0.456, 0.406] 27 | std: [0.229, 0.224, 0.225] 28 | ignore_label: 255 29 | 30 | # # # # # # # # # # # # # # 31 | # 2. training params 32 | # # # # # # # # # # # # # # 33 | trainer: # Required. 34 | epochs: 80 35 | sup_only_epoch: 0 # 0, 1 36 | evaluate_student: True 37 | optimizer: 38 | type: SGD 39 | kwargs: 40 | lr: 0.001 # 0.001:8*2gpus, 4*4gpus; 0.002:8*4gpus. 41 | momentum: 0.9 42 | weight_decay: 0.0001 43 | lr_scheduler: 44 | mode: poly 45 | kwargs: 46 | power: 0.9 47 | # # # # # # # # # # # # # # 48 | # unsupervised loss 49 | # # # # # # # # # # # # # # 50 | unsupervised: 51 | flag_extra_weak: False 52 | threshold: 0.95 # ============================================================================= <>: [0.7, 0.75, 0.8, 0.85, 0.9, 0.95] 53 | loss_weight: 1.0 # ============================================================================= << abl2-weight >> : [1.0, 1.5, 2.0, 2.5, 3.0, 4.0] 54 | #------ 2)cutmix augs ------# 55 | use_cutmix: True 56 | use_cutmix_adaptive: True 57 | use_cutmix_trigger_prob: 1.0 58 | 59 | # # # # # # # # # # # # # # 60 | # 3. output files, and loss 61 | # # # # # # # # # # # # # # 62 | saver: 63 | snapshot_dir: checkpoints 64 | pretrain: '' 65 | use_tb: False 66 | 67 | criterion: 68 | type: CELoss 69 | kwargs: 70 | use_weight: False 71 | 72 | # # # # # # # # # # # # # # 73 | # 4. models 74 | # # # # # # # # # # # # # # 75 | net: # Required. 76 | num_classes: 21 77 | sync_bn: True 78 | ema_decay: 0.999 79 | encoder: 80 | # type: augseg.models.resnet.resnet50 81 | # pretrain: ./pretrained/resnet50.pth 82 | type: augseg.models.resnet.resnet101 83 | pretrain: ./pretrained/resnet101.pth 84 | kwargs: 85 | zero_init_residual: True 86 | multi_grid: True 87 | replace_stride_with_dilation: [False, False, True] # [output_stride = 16] 88 | # replace_stride_with_dilation: [False, True, True] # [output_stride = 8] 89 | decoder: 90 | type: augseg.models.decoder.dec_deeplabv3_plus 91 | kwargs: 92 | inner_planes: 256 93 | low_conv_planes: 48 94 | dilations: [6, 12, 18] # [output_stride = 16] 95 | # dilations: [12, 24, 36] # [output_stride = 8] 96 | -------------------------------------------------------------------------------- /training-logs/3 Semi-Voc-fine/366/config_semi.yaml: -------------------------------------------------------------------------------- 1 | # # # # # # # # # # # # # # 2 | # 1. datasets 3 | # # # # # # # # # # # # # # 4 | dataset: # Required. 5 | type: pascal_semi 6 | train: 7 | data_root: ./data/VOC2012 8 | data_list: ./data/splitsall/pascal/366/labeled.txt 9 | batch_size: 8 10 | flip: True 11 | rand_resize: [0.5, 2.0] 12 | resize_base_size: 500 13 | crop: 14 | type: rand 15 | size: [513, 513] 16 | #---- 1) strong data augs ----# 17 | strong_aug: 18 | num_augs: 3 # !!!!!!!!!!!!!!!!!!!! key param1 (4): 1, 2, 3, 4 !!!!!!!!!!!!!!!!!!!! 19 | flag_use_random_num_sampling: True # !!!!!!!!!!!!!!!!!!!! key param1 (2): True, False !!!!!!!!!!!!!!!!!!!! 20 | val: 21 | data_root: ./data/VOC2012 22 | data_list: ./data/splitsall/pascal/val.txt 23 | batch_size: 1 24 | n_sup: 366 25 | workers: 4 26 | mean: [0.485, 0.456, 0.406] 27 | std: [0.229, 0.224, 0.225] 28 | ignore_label: 255 29 | 30 | # # # # # # # # # # # # # # 31 | # 2. training params 32 | # # # # # # # # # # # # # # 33 | trainer: # Required. 34 | epochs: 80 35 | sup_only_epoch: 0 # 0, 1 36 | evaluate_student: True 37 | optimizer: 38 | type: SGD 39 | kwargs: 40 | lr: 0.001 # 0.001:8*2gpus, 4*4gpus; 0.002:8*4gpus. 41 | momentum: 0.9 42 | weight_decay: 0.0001 43 | lr_scheduler: 44 | mode: poly 45 | kwargs: 46 | power: 0.9 47 | # # # # # # # # # # # # # # 48 | # unsupervised loss 49 | # # # # # # # # # # # # # # 50 | unsupervised: 51 | flag_extra_weak: False 52 | threshold: 0.95 # ============================================================================= <>: [0.7, 0.75, 0.8, 0.85, 0.9, 0.95] 53 | loss_weight: 1.0 # ============================================================================= << abl2-weight >> : [1.0, 1.5, 2.0, 2.5, 3.0, 4.0] 54 | #------ 2)cutmix augs ------# 55 | use_cutmix: True 56 | use_cutmix_adaptive: True 57 | use_cutmix_trigger_prob: 1.0 58 | 59 | # # # # # # # # # # # # # # 60 | # 3. output files, and loss 61 | # # # # # # # # # # # # # # 62 | saver: 63 | snapshot_dir: checkpoints 64 | pretrain: '' 65 | use_tb: False 66 | 67 | criterion: 68 | type: CELoss 69 | kwargs: 70 | use_weight: False 71 | 72 | # # # # # # # # # # # # # # 73 | # 4. models 74 | # # # # # # # # # # # # # # 75 | net: # Required. 76 | num_classes: 21 77 | sync_bn: True 78 | ema_decay: 0.999 79 | encoder: 80 | # type: augseg.models.resnet.resnet50 81 | # pretrain: ./pretrained/resnet50.pth 82 | type: augseg.models.resnet.resnet101 83 | pretrain: ./pretrained/resnet101.pth 84 | kwargs: 85 | zero_init_residual: True 86 | multi_grid: True 87 | replace_stride_with_dilation: [False, False, True] # [output_stride = 16] 88 | # replace_stride_with_dilation: [False, True, True] # [output_stride = 8] 89 | decoder: 90 | type: augseg.models.decoder.dec_deeplabv3_plus 91 | kwargs: 92 | inner_planes: 256 93 | low_conv_planes: 48 94 | dilations: [6, 12, 18] # [output_stride = 16] 95 | # dilations: [12, 24, 36] # [output_stride = 8] 96 | -------------------------------------------------------------------------------- /training-logs/3 Semi-Voc-fine/732/config_semi.yaml: -------------------------------------------------------------------------------- 1 | # # # # # # # # # # # # # # 2 | # 1. datasets 3 | # # # # # # # # # # # # # # 4 | dataset: # Required. 5 | type: pascal_semi 6 | train: 7 | data_root: ./data/VOC2012 8 | data_list: ./data/splitsall/pascal/732/labeled.txt 9 | batch_size: 8 10 | flip: True 11 | rand_resize: [0.5, 2.0] 12 | resize_base_size: 500 13 | crop: 14 | type: rand 15 | size: [513, 513] 16 | #---- 1) strong data augs ----# 17 | strong_aug: 18 | num_augs: 3 # !!!!!!!!!!!!!!!!!!!! key param1 (4): 1, 2, 3, 4 !!!!!!!!!!!!!!!!!!!! 19 | flag_use_random_num_sampling: True # !!!!!!!!!!!!!!!!!!!! key param1 (2): True, False !!!!!!!!!!!!!!!!!!!! 20 | val: 21 | data_root: ./data/VOC2012 22 | data_list: ./data/splitsall/pascal/val.txt 23 | batch_size: 1 24 | n_sup: 732 25 | workers: 4 26 | mean: [0.485, 0.456, 0.406] 27 | std: [0.229, 0.224, 0.225] 28 | ignore_label: 255 29 | 30 | # # # # # # # # # # # # # # 31 | # 2. training params 32 | # # # # # # # # # # # # # # 33 | trainer: # Required. 34 | epochs: 80 35 | sup_only_epoch: 0 # 0, 1 36 | evaluate_student: True 37 | optimizer: 38 | type: SGD 39 | kwargs: 40 | lr: 0.001 # 0.001:8*2gpus, 4*4gpus; 0.002:8*4gpus. 41 | momentum: 0.9 42 | weight_decay: 0.0001 43 | lr_scheduler: 44 | mode: poly 45 | kwargs: 46 | power: 0.9 47 | # # # # # # # # # # # # # # 48 | # unsupervised loss 49 | # # # # # # # # # # # # # # 50 | unsupervised: 51 | flag_extra_weak: False 52 | threshold: 0.95 # ============================================================================= <>: [0.7, 0.75, 0.8, 0.85, 0.9, 0.95] 53 | loss_weight: 1.0 # ============================================================================= << abl2-weight >> : [1.0, 1.5, 2.0, 2.5, 3.0, 4.0] 54 | #------ 2)cutmix augs ------# 55 | use_cutmix: True 56 | use_cutmix_adaptive: True 57 | use_cutmix_trigger_prob: 1.0 58 | 59 | # # # # # # # # # # # # # # 60 | # 3. output files, and loss 61 | # # # # # # # # # # # # # # 62 | saver: 63 | snapshot_dir: checkpoints 64 | pretrain: '' 65 | use_tb: False 66 | 67 | criterion: 68 | type: CELoss 69 | kwargs: 70 | use_weight: False 71 | 72 | # # # # # # # # # # # # # # 73 | # 4. models 74 | # # # # # # # # # # # # # # 75 | net: # Required. 76 | num_classes: 21 77 | sync_bn: True 78 | ema_decay: 0.999 79 | encoder: 80 | # type: augseg.models.resnet.resnet50 81 | # pretrain: ./pretrained/resnet50.pth 82 | type: augseg.models.resnet.resnet101 83 | pretrain: ./pretrained/resnet101.pth 84 | kwargs: 85 | zero_init_residual: True 86 | multi_grid: True 87 | replace_stride_with_dilation: [False, False, True] # [output_stride = 16] 88 | # replace_stride_with_dilation: [False, True, True] # [output_stride = 8] 89 | decoder: 90 | type: augseg.models.decoder.dec_deeplabv3_plus 91 | kwargs: 92 | inner_planes: 256 93 | low_conv_planes: 48 94 | dilations: [6, 12, 18] # [output_stride = 16] 95 | # dilations: [12, 24, 36] # [output_stride = 8] 96 | -------------------------------------------------------------------------------- /training-logs/3 Semi-Voc-fine/92/config_semi.yaml: -------------------------------------------------------------------------------- 1 | # # # # # # # # # # # # # # 2 | # 1. datasets 3 | # # # # # # # # # # # # # # 4 | dataset: # Required. 5 | type: pascal_semi 6 | train: 7 | data_root: ./data/VOC2012 8 | data_list: ./data/splitsall/pascal/92/labeled.txt 9 | batch_size: 8 10 | flip: True 11 | rand_resize: [0.5, 2.0] 12 | resize_base_size: 500 13 | crop: 14 | type: rand 15 | size: [513, 513] 16 | #---- 1) strong data augs ----# 17 | strong_aug: 18 | num_augs: 3 # !!!!!!!!!!!!!!!!!!!! key param1 (4): 1, 2, 3, 4 !!!!!!!!!!!!!!!!!!!! 19 | flag_use_random_num_sampling: True # !!!!!!!!!!!!!!!!!!!! key param1 (2): True, False !!!!!!!!!!!!!!!!!!!! 20 | val: 21 | data_root: ./data/VOC2012 22 | data_list: ./data/splitsall/pascal/val.txt 23 | batch_size: 1 24 | n_sup: 92 25 | workers: 4 26 | mean: [0.485, 0.456, 0.406] 27 | std: [0.229, 0.224, 0.225] 28 | ignore_label: 255 29 | 30 | # # # # # # # # # # # # # # 31 | # 2. training params 32 | # # # # # # # # # # # # # # 33 | trainer: # Required. 34 | epochs: 80 35 | sup_only_epoch: 0 # 0, 1 36 | evaluate_student: True 37 | optimizer: 38 | type: SGD 39 | kwargs: 40 | lr: 0.001 # 0.001:8*2gpus, 4*4gpus; 0.002:8*4gpus. 41 | momentum: 0.9 42 | weight_decay: 0.0001 43 | lr_scheduler: 44 | mode: poly 45 | kwargs: 46 | power: 0.9 47 | # # # # # # # # # # # # # # 48 | # unsupervised loss 49 | # # # # # # # # # # # # # # 50 | unsupervised: 51 | flag_extra_weak: False 52 | threshold: 0.95 # ============================================================================= <>: [0.7, 0.75, 0.8, 0.85, 0.9, 0.95] 53 | loss_weight: 1.0 # ============================================================================= << abl2-weight >> : [1.0, 1.5, 2.0, 2.5, 3.0, 4.0] 54 | #------ 2)cutmix augs ------# 55 | use_cutmix: True 56 | use_cutmix_adaptive: True 57 | use_cutmix_trigger_prob: 1.0 58 | 59 | # # # # # # # # # # # # # # 60 | # 3. output files, and loss 61 | # # # # # # # # # # # # # # 62 | saver: 63 | snapshot_dir: checkpoints 64 | pretrain: '' 65 | use_tb: False 66 | 67 | criterion: 68 | type: CELoss 69 | kwargs: 70 | use_weight: False 71 | 72 | # # # # # # # # # # # # # # 73 | # 4. models 74 | # # # # # # # # # # # # # # 75 | net: # Required. 76 | num_classes: 21 77 | sync_bn: True 78 | ema_decay: 0.999 79 | encoder: 80 | # type: augseg.models.resnet.resnet50 81 | # pretrain: ./pretrained/resnet50.pth 82 | type: augseg.models.resnet.resnet101 83 | pretrain: ./pretrained/resnet101.pth 84 | kwargs: 85 | zero_init_residual: True 86 | multi_grid: True 87 | replace_stride_with_dilation: [False, False, True] # [output_stride = 16] 88 | # replace_stride_with_dilation: [False, True, True] # [output_stride = 8] 89 | decoder: 90 | type: augseg.models.decoder.dec_deeplabv3_plus 91 | kwargs: 92 | inner_planes: 256 93 | low_conv_planes: 48 94 | dilations: [6, 12, 18] # [output_stride = 16] 95 | # dilations: [12, 24, 36] # [output_stride = 8] 96 | -------------------------------------------------------------------------------- /training-logs/4 Semi-Voc-blender/r101-1323/config_semi.yaml: -------------------------------------------------------------------------------- 1 | # # # # # # # # # # # # # # 2 | # 1. datasets 3 | # # # # # # # # # # # # # # 4 | dataset: # Required. 5 | type: pascal_semi 6 | train: 7 | data_root: ./data/VOC2012 8 | data_list: ./data/splitsall/pascal/1323/labeled.txt 9 | batch_size: 8 10 | flip: True 11 | rand_resize: [0.5, 2.0] 12 | resize_base_size: 500 13 | crop: 14 | type: rand 15 | size: [513, 513] 16 | #---- 1) strong data augs ----# 17 | strong_aug: 18 | num_augs: 3 # !!!!!!!!!!!!!!!!!!!! key param1 (4): 1, 2, 3, 4 !!!!!!!!!!!!!!!!!!!! 19 | flag_use_random_num_sampling: True # !!!!!!!!!!!!!!!!!!!! key param1 (2): True, False !!!!!!!!!!!!!!!!!!!! 20 | val: 21 | data_root: ./data/VOC2012 22 | data_list: ./data/splitsall/pascal/val.txt 23 | batch_size: 1 24 | n_sup: 1323 25 | workers: 4 26 | mean: [0.485, 0.456, 0.406] 27 | std: [0.229, 0.224, 0.225] 28 | ignore_label: 255 29 | 30 | # # # # # # # # # # # # # # 31 | # 2. training params 32 | # # # # # # # # # # # # # # 33 | trainer: # Required. 34 | epochs: 80 35 | sup_only_epoch: 0 # 0, 1 36 | evaluate_student: True 37 | optimizer: 38 | type: SGD 39 | kwargs: 40 | lr: 0.001 # 0.001:8*2gpus, 4*4gpus; 0.002:8*4gpus. 41 | momentum: 0.9 42 | weight_decay: 0.0001 43 | lr_scheduler: 44 | mode: poly 45 | kwargs: 46 | power: 0.9 47 | # # # # # # # # # # # # # # 48 | # unsupervised loss 49 | # # # # # # # # # # # # # # 50 | unsupervised: 51 | flag_extra_weak: False 52 | threshold: 0.95 # ============================================================================= <>: [0.7, 0.75, 0.8, 0.85, 0.9, 0.95] 53 | loss_weight: 1.0 # ============================================================================= << abl2-weight >> : [1.0, 1.5, 2.0, 2.5, 3.0, 4.0] 54 | #------ 2)cutmix augs ------# 55 | use_cutmix: True 56 | use_cutmix_adaptive: True 57 | use_cutmix_trigger_prob: 1.0 58 | 59 | # # # # # # # # # # # # # # 60 | # 3. output files, and loss 61 | # # # # # # # # # # # # # # 62 | saver: 63 | snapshot_dir: checkpoints 64 | pretrain: '' 65 | use_tb: False 66 | 67 | criterion: 68 | type: CELoss 69 | kwargs: 70 | use_weight: False 71 | 72 | # # # # # # # # # # # # # # 73 | # 4. models 74 | # # # # # # # # # # # # # # 75 | net: # Required. 76 | num_classes: 21 77 | sync_bn: True 78 | ema_decay: 0.999 79 | encoder: 80 | # type: augseg.models.resnet.resnet50 81 | # pretrain: ./pretrained/resnet50.pth 82 | type: augseg.models.resnet.resnet101 83 | pretrain: ./pretrained/resnet101.pth 84 | kwargs: 85 | zero_init_residual: True 86 | multi_grid: True 87 | replace_stride_with_dilation: [False, False, True] # [output_stride = 16] 88 | # replace_stride_with_dilation: [False, True, True] # [output_stride = 8] 89 | decoder: 90 | type: augseg.models.decoder.dec_deeplabv3_plus 91 | kwargs: 92 | inner_planes: 256 93 | low_conv_planes: 48 94 | dilations: [6, 12, 18] # [output_stride = 16] 95 | # dilations: [12, 24, 36] # [output_stride = 8] 96 | -------------------------------------------------------------------------------- /training-logs/4 Semi-Voc-blender/r101-2646/config_semi.yaml: -------------------------------------------------------------------------------- 1 | # # # # # # # # # # # # # # 2 | # 1. datasets 3 | # # # # # # # # # # # # # # 4 | dataset: # Required. 5 | type: pascal_semi 6 | train: 7 | data_root: ./data/VOC2012 8 | data_list: ./data/splitsall/pascal/2646/labeled.txt 9 | batch_size: 8 10 | flip: True 11 | rand_resize: [0.5, 2.0] 12 | resize_base_size: 500 13 | crop: 14 | type: rand 15 | size: [513, 513] 16 | #---- 1) strong data augs ----# 17 | strong_aug: 18 | num_augs: 3 # !!!!!!!!!!!!!!!!!!!! key param1 (4): 1, 2, 3, 4 !!!!!!!!!!!!!!!!!!!! 19 | flag_use_random_num_sampling: True # !!!!!!!!!!!!!!!!!!!! key param1 (2): True, False !!!!!!!!!!!!!!!!!!!! 20 | val: 21 | data_root: ./data/VOC2012 22 | data_list: ./data/splitsall/pascal/val.txt 23 | batch_size: 1 24 | n_sup: 2646 25 | workers: 4 26 | mean: [0.485, 0.456, 0.406] 27 | std: [0.229, 0.224, 0.225] 28 | ignore_label: 255 29 | 30 | # # # # # # # # # # # # # # 31 | # 2. training params 32 | # # # # # # # # # # # # # # 33 | trainer: # Required. 34 | epochs: 80 35 | sup_only_epoch: 0 # 0, 1 36 | evaluate_student: True 37 | optimizer: 38 | type: SGD 39 | kwargs: 40 | lr: 0.001 # 0.001:8*2gpus, 4*4gpus; 0.002:8*4gpus. 41 | momentum: 0.9 42 | weight_decay: 0.0001 43 | lr_scheduler: 44 | mode: poly 45 | kwargs: 46 | power: 0.9 47 | # # # # # # # # # # # # # # 48 | # unsupervised loss 49 | # # # # # # # # # # # # # # 50 | unsupervised: 51 | flag_extra_weak: False 52 | threshold: 0.95 # ============================================================================= <>: [0.7, 0.75, 0.8, 0.85, 0.9, 0.95] 53 | loss_weight: 1.0 # ============================================================================= << abl2-weight >> : [1.0, 1.5, 2.0, 2.5, 3.0, 4.0] 54 | #------ 2)cutmix augs ------# 55 | use_cutmix: True 56 | use_cutmix_adaptive: True 57 | use_cutmix_trigger_prob: 1.0 58 | 59 | # # # # # # # # # # # # # # 60 | # 3. output files, and loss 61 | # # # # # # # # # # # # # # 62 | saver: 63 | snapshot_dir: checkpoints 64 | pretrain: '' 65 | use_tb: False 66 | 67 | criterion: 68 | type: CELoss 69 | kwargs: 70 | use_weight: False 71 | 72 | # # # # # # # # # # # # # # 73 | # 4. models 74 | # # # # # # # # # # # # # # 75 | net: # Required. 76 | num_classes: 21 77 | sync_bn: True 78 | ema_decay: 0.999 79 | encoder: 80 | # type: augseg.models.resnet.resnet50 81 | # pretrain: ./pretrained/resnet50.pth 82 | type: augseg.models.resnet.resnet101 83 | pretrain: ./pretrained/resnet101.pth 84 | kwargs: 85 | zero_init_residual: True 86 | multi_grid: True 87 | replace_stride_with_dilation: [False, False, True] # [output_stride = 16] 88 | # replace_stride_with_dilation: [False, True, True] # [output_stride = 8] 89 | decoder: 90 | type: augseg.models.decoder.dec_deeplabv3_plus 91 | kwargs: 92 | inner_planes: 256 93 | low_conv_planes: 48 94 | dilations: [6, 12, 18] # [output_stride = 16] 95 | # dilations: [12, 24, 36] # [output_stride = 8] 96 | -------------------------------------------------------------------------------- /training-logs/4 Semi-Voc-blender/r101-662/config_semi.yaml: -------------------------------------------------------------------------------- 1 | # # # # # # # # # # # # # # 2 | # 1. datasets 3 | # # # # # # # # # # # # # # 4 | dataset: # Required. 5 | type: pascal_semi 6 | train: 7 | data_root: ./data/VOC2012 8 | data_list: ./data/splitsall/pascal/662/labeled.txt 9 | batch_size: 8 10 | flip: True 11 | rand_resize: [0.5, 2.0] 12 | resize_base_size: 500 13 | crop: 14 | type: rand 15 | size: [513, 513] 16 | #---- 1) strong data augs ----# 17 | strong_aug: 18 | num_augs: 3 # !!!!!!!!!!!!!!!!!!!! key param1 (4): 1, 2, 3, 4 !!!!!!!!!!!!!!!!!!!! 19 | flag_use_random_num_sampling: True # !!!!!!!!!!!!!!!!!!!! key param1 (2): True, False !!!!!!!!!!!!!!!!!!!! 20 | val: 21 | data_root: ./data/VOC2012 22 | data_list: ./data/splitsall/pascal/val.txt 23 | batch_size: 1 24 | n_sup: 662 25 | workers: 4 26 | mean: [0.485, 0.456, 0.406] 27 | std: [0.229, 0.224, 0.225] 28 | ignore_label: 255 29 | 30 | # # # # # # # # # # # # # # 31 | # 2. training params 32 | # # # # # # # # # # # # # # 33 | trainer: # Required. 34 | epochs: 80 35 | sup_only_epoch: 0 # 0, 1 36 | evaluate_student: True 37 | optimizer: 38 | type: SGD 39 | kwargs: 40 | lr: 0.001 # 0.001:8*2gpus, 4*4gpus; 0.002:8*4gpus. 41 | momentum: 0.9 42 | weight_decay: 0.0001 43 | lr_scheduler: 44 | mode: poly 45 | kwargs: 46 | power: 0.9 47 | # # # # # # # # # # # # # # 48 | # unsupervised loss 49 | # # # # # # # # # # # # # # 50 | unsupervised: 51 | flag_extra_weak: False 52 | threshold: 0.95 # ============================================================================= <>: [0.7, 0.75, 0.8, 0.85, 0.9, 0.95] 53 | loss_weight: 1.0 # ============================================================================= << abl2-weight >> : [1.0, 1.5, 2.0, 2.5, 3.0, 4.0] 54 | #------ 2)cutmix augs ------# 55 | use_cutmix: True 56 | use_cutmix_adaptive: True 57 | use_cutmix_trigger_prob: 1.0 58 | 59 | # # # # # # # # # # # # # # 60 | # 3. output files, and loss 61 | # # # # # # # # # # # # # # 62 | saver: 63 | snapshot_dir: checkpoints 64 | pretrain: '' 65 | use_tb: False 66 | 67 | criterion: 68 | type: CELoss 69 | kwargs: 70 | use_weight: False 71 | 72 | # # # # # # # # # # # # # # 73 | # 4. models 74 | # # # # # # # # # # # # # # 75 | net: # Required. 76 | num_classes: 21 77 | sync_bn: True 78 | ema_decay: 0.999 79 | encoder: 80 | # type: augseg.models.resnet.resnet50 81 | # pretrain: ./pretrained/resnet50.pth 82 | type: augseg.models.resnet.resnet101 83 | pretrain: ./pretrained/resnet101.pth 84 | kwargs: 85 | zero_init_residual: True 86 | multi_grid: True 87 | replace_stride_with_dilation: [False, False, True] # [output_stride = 16] 88 | # replace_stride_with_dilation: [False, True, True] # [output_stride = 8] 89 | decoder: 90 | type: augseg.models.decoder.dec_deeplabv3_plus 91 | kwargs: 92 | inner_planes: 256 93 | low_conv_planes: 48 94 | dilations: [6, 12, 18] # [output_stride = 16] 95 | # dilations: [12, 24, 36] # [output_stride = 8] 96 | -------------------------------------------------------------------------------- /training-logs/4 Semi-Voc-blender/r50-1323/config_semi.yaml: -------------------------------------------------------------------------------- 1 | # # # # # # # # # # # # # # 2 | # 1. datasets 3 | # # # # # # # # # # # # # # 4 | dataset: # Required. 5 | type: pascal_semi 6 | train: 7 | data_root: ./data/VOC2012 8 | data_list: ./data/splitsall/pascal/1323/labeled.txt 9 | batch_size: 8 10 | flip: True 11 | rand_resize: [0.5, 2.0] 12 | resize_base_size: 500 13 | crop: 14 | type: rand 15 | size: [513, 513] 16 | #---- 1) strong data augs ----# 17 | strong_aug: 18 | num_augs: 3 # !!!!!!!!!!!!!!!!!!!! key param1 (4): 1, 2, 3, 4 !!!!!!!!!!!!!!!!!!!! 19 | flag_use_random_num_sampling: True # !!!!!!!!!!!!!!!!!!!! key param1 (2): True, False !!!!!!!!!!!!!!!!!!!! 20 | val: 21 | data_root: ./data/VOC2012 22 | data_list: ./data/splitsall/pascal/val.txt 23 | batch_size: 1 24 | n_sup: 1323 25 | workers: 4 26 | mean: [0.485, 0.456, 0.406] 27 | std: [0.229, 0.224, 0.225] 28 | ignore_label: 255 29 | 30 | # # # # # # # # # # # # # # 31 | # 2. training params 32 | # # # # # # # # # # # # # # 33 | trainer: # Required. 34 | epochs: 80 35 | sup_only_epoch: 0 # 0, 1 36 | evaluate_student: True 37 | optimizer: 38 | type: SGD 39 | kwargs: 40 | lr: 0.001 # 0.001:8*2gpus, 4*4gpus; 0.002:8*4gpus. 41 | momentum: 0.9 42 | weight_decay: 0.0001 43 | lr_scheduler: 44 | mode: poly 45 | kwargs: 46 | power: 0.9 47 | # # # # # # # # # # # # # # 48 | # unsupervised loss 49 | # # # # # # # # # # # # # # 50 | unsupervised: 51 | flag_extra_weak: False 52 | threshold: 0.95 # ============================================================================= <>: [0.7, 0.75, 0.8, 0.85, 0.9, 0.95] 53 | loss_weight: 1.0 # ============================================================================= << abl2-weight >> : [1.0, 1.5, 2.0, 2.5, 3.0, 4.0] 54 | #------ 2)cutmix augs ------# 55 | use_cutmix: True 56 | use_cutmix_adaptive: True 57 | use_cutmix_trigger_prob: 1.0 58 | 59 | # # # # # # # # # # # # # # 60 | # 3. output files, and loss 61 | # # # # # # # # # # # # # # 62 | saver: 63 | snapshot_dir: checkpoints 64 | pretrain: '' 65 | use_tb: False 66 | 67 | criterion: 68 | type: CELoss 69 | kwargs: 70 | use_weight: False 71 | 72 | # # # # # # # # # # # # # # 73 | # 4. models 74 | # # # # # # # # # # # # # # 75 | net: # Required. 76 | num_classes: 21 77 | sync_bn: True 78 | ema_decay: 0.999 79 | encoder: 80 | type: augseg.models.resnet.resnet50 81 | pretrain: ./pretrained/resnet50.pth 82 | # type: augseg.models.resnet.resnet101 83 | # pretrain: ./pretrained/resnet101.pth 84 | kwargs: 85 | zero_init_residual: True 86 | multi_grid: True 87 | replace_stride_with_dilation: [False, False, True] # [output_stride = 16] 88 | # replace_stride_with_dilation: [False, True, True] # [output_stride = 8] 89 | decoder: 90 | type: augseg.models.decoder.dec_deeplabv3_plus 91 | kwargs: 92 | inner_planes: 256 93 | low_conv_planes: 48 94 | dilations: [6, 12, 18] # [output_stride = 16] 95 | # dilations: [12, 24, 36] # [output_stride = 8] 96 | -------------------------------------------------------------------------------- /training-logs/4 Semi-Voc-blender/r50-2646/config_semi.yaml: -------------------------------------------------------------------------------- 1 | # # # # # # # # # # # # # # 2 | # 1. datasets 3 | # # # # # # # # # # # # # # 4 | dataset: # Required. 5 | type: pascal_semi 6 | train: 7 | data_root: ./data/VOC2012 8 | data_list: ./data/splitsall/pascal/2646/labeled.txt 9 | batch_size: 8 10 | flip: True 11 | rand_resize: [0.5, 2.0] 12 | resize_base_size: 500 13 | crop: 14 | type: rand 15 | size: [513, 513] 16 | #---- 1) strong data augs ----# 17 | strong_aug: 18 | num_augs: 3 # !!!!!!!!!!!!!!!!!!!! key param1 (4): 1, 2, 3, 4 !!!!!!!!!!!!!!!!!!!! 19 | flag_use_random_num_sampling: True # !!!!!!!!!!!!!!!!!!!! key param1 (2): True, False !!!!!!!!!!!!!!!!!!!! 20 | val: 21 | data_root: ./data/VOC2012 22 | data_list: ./data/splitsall/pascal/val.txt 23 | batch_size: 1 24 | n_sup: 2646 25 | workers: 4 26 | mean: [0.485, 0.456, 0.406] 27 | std: [0.229, 0.224, 0.225] 28 | ignore_label: 255 29 | 30 | # # # # # # # # # # # # # # 31 | # 2. training params 32 | # # # # # # # # # # # # # # 33 | trainer: # Required. 34 | epochs: 80 35 | sup_only_epoch: 0 # 0, 1 36 | evaluate_student: True 37 | optimizer: 38 | type: SGD 39 | kwargs: 40 | lr: 0.001 # 0.001:8*2gpus, 4*4gpus; 0.002:8*4gpus. 41 | momentum: 0.9 42 | weight_decay: 0.0001 43 | lr_scheduler: 44 | mode: poly 45 | kwargs: 46 | power: 0.9 47 | # # # # # # # # # # # # # # 48 | # unsupervised loss 49 | # # # # # # # # # # # # # # 50 | unsupervised: 51 | flag_extra_weak: False 52 | threshold: 0.95 # ============================================================================= <>: [0.7, 0.75, 0.8, 0.85, 0.9, 0.95] 53 | loss_weight: 1.0 # ============================================================================= << abl2-weight >> : [1.0, 1.5, 2.0, 2.5, 3.0, 4.0] 54 | #------ 2)cutmix augs ------# 55 | use_cutmix: True 56 | use_cutmix_adaptive: True 57 | use_cutmix_trigger_prob: 1.0 58 | 59 | # # # # # # # # # # # # # # 60 | # 3. output files, and loss 61 | # # # # # # # # # # # # # # 62 | saver: 63 | snapshot_dir: checkpoints 64 | pretrain: '' 65 | use_tb: False 66 | 67 | criterion: 68 | type: CELoss 69 | kwargs: 70 | use_weight: False 71 | 72 | # # # # # # # # # # # # # # 73 | # 4. models 74 | # # # # # # # # # # # # # # 75 | net: # Required. 76 | num_classes: 21 77 | sync_bn: True 78 | ema_decay: 0.999 79 | encoder: 80 | type: augseg.models.resnet.resnet50 81 | pretrain: ./pretrained/resnet50.pth 82 | # type: augseg.models.resnet.resnet101 83 | # pretrain: ./pretrained/resnet101.pth 84 | kwargs: 85 | zero_init_residual: True 86 | multi_grid: True 87 | replace_stride_with_dilation: [False, False, True] # [output_stride = 16] 88 | # replace_stride_with_dilation: [False, True, True] # [output_stride = 8] 89 | decoder: 90 | type: augseg.models.decoder.dec_deeplabv3_plus 91 | kwargs: 92 | inner_planes: 256 93 | low_conv_planes: 48 94 | dilations: [6, 12, 18] # [output_stride = 16] 95 | # dilations: [12, 24, 36] # [output_stride = 8] 96 | -------------------------------------------------------------------------------- /training-logs/4 Semi-Voc-blender/r50-662/config_semi.yaml: -------------------------------------------------------------------------------- 1 | # # # # # # # # # # # # # # 2 | # 1. datasets 3 | # # # # # # # # # # # # # # 4 | dataset: # Required. 5 | type: pascal_semi 6 | train: 7 | data_root: ./data/VOC2012 8 | data_list: ./data/splitsall/pascal/662/labeled.txt 9 | batch_size: 8 10 | flip: True 11 | rand_resize: [0.5, 2.0] 12 | resize_base_size: 500 13 | crop: 14 | type: rand 15 | size: [513, 513] 16 | #---- 1) strong data augs ----# 17 | strong_aug: 18 | num_augs: 3 # !!!!!!!!!!!!!!!!!!!! key param1 (4): 1, 2, 3, 4 !!!!!!!!!!!!!!!!!!!! 19 | flag_use_random_num_sampling: True # !!!!!!!!!!!!!!!!!!!! key param1 (2): True, False !!!!!!!!!!!!!!!!!!!! 20 | val: 21 | data_root: ./data/VOC2012 22 | data_list: ./data/splitsall/pascal/val.txt 23 | batch_size: 1 24 | n_sup: 662 25 | workers: 4 26 | mean: [0.485, 0.456, 0.406] 27 | std: [0.229, 0.224, 0.225] 28 | ignore_label: 255 29 | 30 | # # # # # # # # # # # # # # 31 | # 2. training params 32 | # # # # # # # # # # # # # # 33 | trainer: # Required. 34 | epochs: 80 35 | sup_only_epoch: 0 # 0, 1 36 | evaluate_student: True 37 | optimizer: 38 | type: SGD 39 | kwargs: 40 | lr: 0.001 # 0.001:8*2gpus, 4*4gpus; 0.002:8*4gpus. 41 | momentum: 0.9 42 | weight_decay: 0.0001 43 | lr_scheduler: 44 | mode: poly 45 | kwargs: 46 | power: 0.9 47 | # # # # # # # # # # # # # # 48 | # unsupervised loss 49 | # # # # # # # # # # # # # # 50 | unsupervised: 51 | flag_extra_weak: False 52 | threshold: 0.95 # ============================================================================= <>: [0.7, 0.75, 0.8, 0.85, 0.9, 0.95] 53 | loss_weight: 1.0 # ============================================================================= << abl2-weight >> : [1.0, 1.5, 2.0, 2.5, 3.0, 4.0] 54 | #------ 2)cutmix augs ------# 55 | use_cutmix: True 56 | use_cutmix_adaptive: True 57 | use_cutmix_trigger_prob: 1.0 58 | 59 | # # # # # # # # # # # # # # 60 | # 3. output files, and loss 61 | # # # # # # # # # # # # # # 62 | saver: 63 | snapshot_dir: checkpoints 64 | pretrain: '' 65 | use_tb: False 66 | 67 | criterion: 68 | type: CELoss 69 | kwargs: 70 | use_weight: False 71 | 72 | # # # # # # # # # # # # # # 73 | # 4. models 74 | # # # # # # # # # # # # # # 75 | net: # Required. 76 | num_classes: 21 77 | sync_bn: True 78 | ema_decay: 0.999 79 | encoder: 80 | type: augseg.models.resnet.resnet50 81 | pretrain: ./pretrained/resnet50.pth 82 | # type: augseg.models.resnet.resnet101 83 | # pretrain: ./pretrained/resnet101.pth 84 | kwargs: 85 | zero_init_residual: True 86 | multi_grid: True 87 | replace_stride_with_dilation: [False, False, True] # [output_stride = 16] 88 | # replace_stride_with_dilation: [False, True, True] # [output_stride = 8] 89 | decoder: 90 | type: augseg.models.decoder.dec_deeplabv3_plus 91 | kwargs: 92 | inner_planes: 256 93 | low_conv_planes: 48 94 | dilations: [6, 12, 18] # [output_stride = 16] 95 | # dilations: [12, 24, 36] # [output_stride = 8] 96 | -------------------------------------------------------------------------------- /training-logs/5 Semi-Voc-blender-split-u2pl/r101-1323/config_semi.yaml: -------------------------------------------------------------------------------- 1 | # # # # # # # # # # # # # # 2 | # 1. datasets 3 | # # # # # # # # # # # # # # 4 | dataset: # Required. 5 | type: pascal_semi 6 | train: 7 | data_root: ./data/VOC2012 8 | data_list: ./data/splitsall/pascal_u2pl/1323/labeled.txt 9 | batch_size: 8 10 | flip: True 11 | rand_resize: [0.5, 2.0] 12 | resize_base_size: 500 13 | crop: 14 | type: rand 15 | size: [513, 513] 16 | #---- 1) strong data augs ----# 17 | strong_aug: 18 | num_augs: 3 # !!!!!!!!!!!!!!!!!!!! key param1 (4): 1, 2, 3, 4 !!!!!!!!!!!!!!!!!!!! 19 | flag_use_random_num_sampling: True # !!!!!!!!!!!!!!!!!!!! key param1 (2): True, False !!!!!!!!!!!!!!!!!!!! 20 | val: 21 | data_root: ./data/VOC2012 22 | data_list: ./data/splitsall/pascal_u2pl/val.txt 23 | batch_size: 1 24 | n_sup: 1323 25 | workers: 4 26 | mean: [0.485, 0.456, 0.406] 27 | std: [0.229, 0.224, 0.225] 28 | ignore_label: 255 29 | 30 | # # # # # # # # # # # # # # 31 | # 2. training params 32 | # # # # # # # # # # # # # # 33 | trainer: # Required. 34 | epochs: 80 35 | sup_only_epoch: 0 # 0, 1 36 | evaluate_student: True 37 | optimizer: 38 | type: SGD 39 | kwargs: 40 | lr: 0.001 # 0.001:8*2gpus, 4*4gpus; 0.002:8*4gpus. 41 | momentum: 0.9 42 | weight_decay: 0.0001 43 | lr_scheduler: 44 | mode: poly 45 | kwargs: 46 | power: 0.9 47 | # # # # # # # # # # # # # # 48 | # unsupervised loss 49 | # # # # # # # # # # # # # # 50 | unsupervised: 51 | flag_extra_weak: False 52 | threshold: 0.95 # ============================================================================= <>: [0.7, 0.75, 0.8, 0.85, 0.9, 0.95] 53 | loss_weight: 1.0 # ============================================================================= << abl2-weight >> : [1.0, 1.5, 2.0, 2.5, 3.0, 4.0] 54 | #------ 2)cutmix augs ------# 55 | use_cutmix: True 56 | use_cutmix_adaptive: True 57 | use_cutmix_trigger_prob: 1.0 58 | 59 | # # # # # # # # # # # # # # 60 | # 3. output files, and loss 61 | # # # # # # # # # # # # # # 62 | saver: 63 | snapshot_dir: checkpoints 64 | pretrain: '' 65 | use_tb: False 66 | 67 | criterion: 68 | type: CELoss 69 | kwargs: 70 | use_weight: False 71 | 72 | # # # # # # # # # # # # # # 73 | # 4. models 74 | # # # # # # # # # # # # # # 75 | net: # Required. 76 | num_classes: 21 77 | sync_bn: True 78 | ema_decay: 0.999 79 | encoder: 80 | # type: augseg.models.resnet.resnet50 81 | # pretrain: ./pretrained/resnet50.pth 82 | type: augseg.models.resnet.resnet101 83 | pretrain: ./pretrained/resnet101.pth 84 | kwargs: 85 | zero_init_residual: True 86 | multi_grid: True 87 | replace_stride_with_dilation: [False, False, True] # [output_stride = 16] 88 | # replace_stride_with_dilation: [False, True, True] # [output_stride = 8] 89 | decoder: 90 | type: augseg.models.decoder.dec_deeplabv3_plus 91 | kwargs: 92 | inner_planes: 256 93 | low_conv_planes: 48 94 | dilations: [6, 12, 18] # [output_stride = 16] 95 | # dilations: [12, 24, 36] # [output_stride = 8] 96 | -------------------------------------------------------------------------------- /training-logs/5 Semi-Voc-blender-split-u2pl/r101-2646/config_semi.yaml: -------------------------------------------------------------------------------- 1 | # # # # # # # # # # # # # # 2 | # 1. datasets 3 | # # # # # # # # # # # # # # 4 | dataset: # Required. 5 | type: pascal_semi 6 | train: 7 | data_root: ./data/VOC2012 8 | data_list: ./data/splitsall/pascal_u2pl/2646/labeled.txt 9 | batch_size: 8 10 | flip: True 11 | rand_resize: [0.5, 2.0] 12 | resize_base_size: 500 13 | crop: 14 | type: rand 15 | size: [513, 513] 16 | #---- 1) strong data augs ----# 17 | strong_aug: 18 | num_augs: 3 # !!!!!!!!!!!!!!!!!!!! key param1 (4): 1, 2, 3, 4 !!!!!!!!!!!!!!!!!!!! 19 | flag_use_random_num_sampling: True # !!!!!!!!!!!!!!!!!!!! key param1 (2): True, False !!!!!!!!!!!!!!!!!!!! 20 | val: 21 | data_root: ./data/VOC2012 22 | data_list: ./data/splitsall/pascal_u2pl/val.txt 23 | batch_size: 1 24 | n_sup: 2646 25 | workers: 4 26 | mean: [0.485, 0.456, 0.406] 27 | std: [0.229, 0.224, 0.225] 28 | ignore_label: 255 29 | 30 | # # # # # # # # # # # # # # 31 | # 2. training params 32 | # # # # # # # # # # # # # # 33 | trainer: # Required. 34 | epochs: 80 35 | sup_only_epoch: 0 # 0, 1 36 | evaluate_student: True 37 | optimizer: 38 | type: SGD 39 | kwargs: 40 | lr: 0.001 # 0.001:8*2gpus, 4*4gpus; 0.002:8*4gpus. 41 | momentum: 0.9 42 | weight_decay: 0.0001 43 | lr_scheduler: 44 | mode: poly 45 | kwargs: 46 | power: 0.9 47 | # # # # # # # # # # # # # # 48 | # unsupervised loss 49 | # # # # # # # # # # # # # # 50 | unsupervised: 51 | flag_extra_weak: False 52 | threshold: 0.95 # ============================================================================= <>: [0.7, 0.75, 0.8, 0.85, 0.9, 0.95] 53 | loss_weight: 1.0 # ============================================================================= << abl2-weight >> : [1.0, 1.5, 2.0, 2.5, 3.0, 4.0] 54 | #------ 2)cutmix augs ------# 55 | use_cutmix: True 56 | use_cutmix_adaptive: True 57 | use_cutmix_trigger_prob: 1.0 58 | 59 | # # # # # # # # # # # # # # 60 | # 3. output files, and loss 61 | # # # # # # # # # # # # # # 62 | saver: 63 | snapshot_dir: checkpoints 64 | pretrain: '' 65 | use_tb: False 66 | 67 | criterion: 68 | type: CELoss 69 | kwargs: 70 | use_weight: False 71 | 72 | # # # # # # # # # # # # # # 73 | # 4. models 74 | # # # # # # # # # # # # # # 75 | net: # Required. 76 | num_classes: 21 77 | sync_bn: True 78 | ema_decay: 0.999 79 | encoder: 80 | # type: augseg.models.resnet.resnet50 81 | # pretrain: ./pretrained/resnet50.pth 82 | type: augseg.models.resnet.resnet101 83 | pretrain: ./pretrained/resnet101.pth 84 | kwargs: 85 | zero_init_residual: True 86 | multi_grid: True 87 | replace_stride_with_dilation: [False, False, True] # [output_stride = 16] 88 | # replace_stride_with_dilation: [False, True, True] # [output_stride = 8] 89 | decoder: 90 | type: augseg.models.decoder.dec_deeplabv3_plus 91 | kwargs: 92 | inner_planes: 256 93 | low_conv_planes: 48 94 | dilations: [6, 12, 18] # [output_stride = 16] 95 | # dilations: [12, 24, 36] # [output_stride = 8] 96 | -------------------------------------------------------------------------------- /training-logs/5 Semi-Voc-blender-split-u2pl/r101-662/config_semi.yaml: -------------------------------------------------------------------------------- 1 | # # # # # # # # # # # # # # 2 | # 1. datasets 3 | # # # # # # # # # # # # # # 4 | dataset: # Required. 5 | type: pascal_semi 6 | train: 7 | data_root: ./data/VOC2012 8 | data_list: ./data/splitsall/pascal_u2pl/662/labeled.txt 9 | batch_size: 8 10 | flip: True 11 | rand_resize: [0.5, 2.0] 12 | resize_base_size: 500 13 | crop: 14 | type: rand 15 | size: [513, 513] 16 | #---- 1) strong data augs ----# 17 | strong_aug: 18 | num_augs: 3 # !!!!!!!!!!!!!!!!!!!! key param1 (4): 1, 2, 3, 4 !!!!!!!!!!!!!!!!!!!! 19 | flag_use_random_num_sampling: True # !!!!!!!!!!!!!!!!!!!! key param1 (2): True, False !!!!!!!!!!!!!!!!!!!! 20 | val: 21 | data_root: ./data/VOC2012 22 | data_list: ./data/splitsall/pascal_u2pl/val.txt 23 | batch_size: 1 24 | n_sup: 662 25 | workers: 4 26 | mean: [0.485, 0.456, 0.406] 27 | std: [0.229, 0.224, 0.225] 28 | ignore_label: 255 29 | 30 | # # # # # # # # # # # # # # 31 | # 2. training params 32 | # # # # # # # # # # # # # # 33 | trainer: # Required. 34 | epochs: 80 35 | sup_only_epoch: 0 # 0, 1 36 | evaluate_student: True 37 | optimizer: 38 | type: SGD 39 | kwargs: 40 | lr: 0.001 # 0.001:8*2gpus, 4*4gpus; 0.002:8*4gpus. 41 | momentum: 0.9 42 | weight_decay: 0.0001 43 | lr_scheduler: 44 | mode: poly 45 | kwargs: 46 | power: 0.9 47 | # # # # # # # # # # # # # # 48 | # unsupervised loss 49 | # # # # # # # # # # # # # # 50 | unsupervised: 51 | flag_extra_weak: False 52 | threshold: 0.95 # ============================================================================= <>: [0.7, 0.75, 0.8, 0.85, 0.9, 0.95] 53 | loss_weight: 1.0 # ============================================================================= << abl2-weight >> : [1.0, 1.5, 2.0, 2.5, 3.0, 4.0] 54 | #------ 2)cutmix augs ------# 55 | use_cutmix: True 56 | use_cutmix_adaptive: True 57 | use_cutmix_trigger_prob: 1.0 58 | 59 | # # # # # # # # # # # # # # 60 | # 3. output files, and loss 61 | # # # # # # # # # # # # # # 62 | saver: 63 | snapshot_dir: checkpoints 64 | pretrain: '' 65 | use_tb: False 66 | 67 | criterion: 68 | type: CELoss 69 | kwargs: 70 | use_weight: False 71 | 72 | # # # # # # # # # # # # # # 73 | # 4. models 74 | # # # # # # # # # # # # # # 75 | net: # Required. 76 | num_classes: 21 77 | sync_bn: True 78 | ema_decay: 0.999 79 | encoder: 80 | # type: augseg.models.resnet.resnet50 81 | # pretrain: ./pretrained/resnet50.pth 82 | type: augseg.models.resnet.resnet101 83 | pretrain: ./pretrained/resnet101.pth 84 | kwargs: 85 | zero_init_residual: True 86 | multi_grid: True 87 | replace_stride_with_dilation: [False, False, True] # [output_stride = 16] 88 | # replace_stride_with_dilation: [False, True, True] # [output_stride = 8] 89 | decoder: 90 | type: augseg.models.decoder.dec_deeplabv3_plus 91 | kwargs: 92 | inner_planes: 256 93 | low_conv_planes: 48 94 | dilations: [6, 12, 18] # [output_stride = 16] 95 | # dilations: [12, 24, 36] # [output_stride = 8] 96 | -------------------------------------------------------------------------------- /training-logs/5 Semi-Voc-blender-split-u2pl/r50-1323/config_semi.yaml: -------------------------------------------------------------------------------- 1 | # # # # # # # # # # # # # # 2 | # 1. datasets 3 | # # # # # # # # # # # # # # 4 | dataset: # Required. 5 | type: pascal_semi 6 | train: 7 | data_root: ./data/VOC2012 8 | data_list: ./data/splitsall/pascal_u2pl/1323/labeled.txt 9 | batch_size: 8 10 | flip: True 11 | rand_resize: [0.5, 2.0] 12 | resize_base_size: 500 13 | crop: 14 | type: rand 15 | size: [513, 513] 16 | #---- 1) strong data augs ----# 17 | strong_aug: 18 | num_augs: 3 # !!!!!!!!!!!!!!!!!!!! key param1 (4): 1, 2, 3, 4 !!!!!!!!!!!!!!!!!!!! 19 | flag_use_random_num_sampling: True # !!!!!!!!!!!!!!!!!!!! key param1 (2): True, False !!!!!!!!!!!!!!!!!!!! 20 | val: 21 | data_root: ./data/VOC2012 22 | data_list: ./data/splitsall/pascal_u2pl/val.txt 23 | batch_size: 1 24 | n_sup: 1323 25 | workers: 4 26 | mean: [0.485, 0.456, 0.406] 27 | std: [0.229, 0.224, 0.225] 28 | ignore_label: 255 29 | 30 | # # # # # # # # # # # # # # 31 | # 2. training params 32 | # # # # # # # # # # # # # # 33 | trainer: # Required. 34 | epochs: 80 35 | sup_only_epoch: 0 # 0, 1 36 | evaluate_student: True 37 | optimizer: 38 | type: SGD 39 | kwargs: 40 | lr: 0.001 # 0.001:8*2gpus, 4*4gpus; 0.002:8*4gpus. 41 | momentum: 0.9 42 | weight_decay: 0.0001 43 | lr_scheduler: 44 | mode: poly 45 | kwargs: 46 | power: 0.9 47 | # # # # # # # # # # # # # # 48 | # unsupervised loss 49 | # # # # # # # # # # # # # # 50 | unsupervised: 51 | flag_extra_weak: False 52 | threshold: 0.95 # ============================================================================= <>: [0.7, 0.75, 0.8, 0.85, 0.9, 0.95] 53 | loss_weight: 1.0 # ============================================================================= << abl2-weight >> : [1.0, 1.5, 2.0, 2.5, 3.0, 4.0] 54 | #------ 2)cutmix augs ------# 55 | use_cutmix: True 56 | use_cutmix_adaptive: True 57 | use_cutmix_trigger_prob: 1.0 58 | 59 | # # # # # # # # # # # # # # 60 | # 3. output files, and loss 61 | # # # # # # # # # # # # # # 62 | saver: 63 | snapshot_dir: checkpoints 64 | pretrain: '' 65 | use_tb: False 66 | 67 | criterion: 68 | type: CELoss 69 | kwargs: 70 | use_weight: False 71 | 72 | # # # # # # # # # # # # # # 73 | # 4. models 74 | # # # # # # # # # # # # # # 75 | net: # Required. 76 | num_classes: 21 77 | sync_bn: True 78 | ema_decay: 0.999 79 | encoder: 80 | type: augseg.models.resnet.resnet50 81 | pretrain: ./pretrained/resnet50.pth 82 | # type: augseg.models.resnet.resnet101 83 | # pretrain: ./pretrained/resnet101.pth 84 | kwargs: 85 | zero_init_residual: True 86 | multi_grid: True 87 | replace_stride_with_dilation: [False, False, True] # [output_stride = 16] 88 | # replace_stride_with_dilation: [False, True, True] # [output_stride = 8] 89 | decoder: 90 | type: augseg.models.decoder.dec_deeplabv3_plus 91 | kwargs: 92 | inner_planes: 256 93 | low_conv_planes: 48 94 | dilations: [6, 12, 18] # [output_stride = 16] 95 | # dilations: [12, 24, 36] # [output_stride = 8] 96 | -------------------------------------------------------------------------------- /training-logs/5 Semi-Voc-blender-split-u2pl/r50-2646/config_semi.yaml: -------------------------------------------------------------------------------- 1 | # # # # # # # # # # # # # # 2 | # 1. datasets 3 | # # # # # # # # # # # # # # 4 | dataset: # Required. 5 | type: pascal_semi 6 | train: 7 | data_root: ./data/VOC2012 8 | data_list: ./data/splitsall/pascal_u2pl/2646/labeled.txt 9 | batch_size: 8 10 | flip: True 11 | rand_resize: [0.5, 2.0] 12 | resize_base_size: 500 13 | crop: 14 | type: rand 15 | size: [513, 513] 16 | #---- 1) strong data augs ----# 17 | strong_aug: 18 | num_augs: 3 # !!!!!!!!!!!!!!!!!!!! key param1 (4): 1, 2, 3, 4 !!!!!!!!!!!!!!!!!!!! 19 | flag_use_random_num_sampling: True # !!!!!!!!!!!!!!!!!!!! key param1 (2): True, False !!!!!!!!!!!!!!!!!!!! 20 | val: 21 | data_root: ./data/VOC2012 22 | data_list: ./data/splitsall/pascal_u2pl/val.txt 23 | batch_size: 1 24 | n_sup: 2646 25 | workers: 4 26 | mean: [0.485, 0.456, 0.406] 27 | std: [0.229, 0.224, 0.225] 28 | ignore_label: 255 29 | 30 | # # # # # # # # # # # # # # 31 | # 2. training params 32 | # # # # # # # # # # # # # # 33 | trainer: # Required. 34 | epochs: 80 35 | sup_only_epoch: 0 # 0, 1 36 | evaluate_student: True 37 | optimizer: 38 | type: SGD 39 | kwargs: 40 | lr: 0.001 # 0.001:8*2gpus, 4*4gpus; 0.002:8*4gpus. 41 | momentum: 0.9 42 | weight_decay: 0.0001 43 | lr_scheduler: 44 | mode: poly 45 | kwargs: 46 | power: 0.9 47 | # # # # # # # # # # # # # # 48 | # unsupervised loss 49 | # # # # # # # # # # # # # # 50 | unsupervised: 51 | flag_extra_weak: False 52 | threshold: 0.95 # ============================================================================= <>: [0.7, 0.75, 0.8, 0.85, 0.9, 0.95] 53 | loss_weight: 1.0 # ============================================================================= << abl2-weight >> : [1.0, 1.5, 2.0, 2.5, 3.0, 4.0] 54 | #------ 2)cutmix augs ------# 55 | use_cutmix: True 56 | use_cutmix_adaptive: True 57 | use_cutmix_trigger_prob: 1.0 58 | 59 | # # # # # # # # # # # # # # 60 | # 3. output files, and loss 61 | # # # # # # # # # # # # # # 62 | saver: 63 | snapshot_dir: checkpoints 64 | pretrain: '' 65 | use_tb: False 66 | 67 | criterion: 68 | type: CELoss 69 | kwargs: 70 | use_weight: False 71 | 72 | # # # # # # # # # # # # # # 73 | # 4. models 74 | # # # # # # # # # # # # # # 75 | net: # Required. 76 | num_classes: 21 77 | sync_bn: True 78 | ema_decay: 0.999 79 | encoder: 80 | type: augseg.models.resnet.resnet50 81 | pretrain: ./pretrained/resnet50.pth 82 | # type: augseg.models.resnet.resnet101 83 | # pretrain: ./pretrained/resnet101.pth 84 | kwargs: 85 | zero_init_residual: True 86 | multi_grid: True 87 | replace_stride_with_dilation: [False, False, True] # [output_stride = 16] 88 | # replace_stride_with_dilation: [False, True, True] # [output_stride = 8] 89 | decoder: 90 | type: augseg.models.decoder.dec_deeplabv3_plus 91 | kwargs: 92 | inner_planes: 256 93 | low_conv_planes: 48 94 | dilations: [6, 12, 18] # [output_stride = 16] 95 | # dilations: [12, 24, 36] # [output_stride = 8] 96 | -------------------------------------------------------------------------------- /training-logs/5 Semi-Voc-blender-split-u2pl/r50-662/config_semi.yaml: -------------------------------------------------------------------------------- 1 | # # # # # # # # # # # # # # 2 | # 1. datasets 3 | # # # # # # # # # # # # # # 4 | dataset: # Required. 5 | type: pascal_semi 6 | train: 7 | data_root: ./data/VOC2012 8 | data_list: ./data/splitsall/pascal_u2pl/662/labeled.txt 9 | batch_size: 8 10 | flip: True 11 | rand_resize: [0.5, 2.0] 12 | resize_base_size: 500 13 | crop: 14 | type: rand 15 | size: [513, 513] 16 | #---- 1) strong data augs ----# 17 | strong_aug: 18 | num_augs: 3 # !!!!!!!!!!!!!!!!!!!! key param1 (4): 1, 2, 3, 4 !!!!!!!!!!!!!!!!!!!! 19 | flag_use_random_num_sampling: True # !!!!!!!!!!!!!!!!!!!! key param1 (2): True, False !!!!!!!!!!!!!!!!!!!! 20 | val: 21 | data_root: ./data/VOC2012 22 | data_list: ./data/splitsall/pascal_u2pl/val.txt 23 | batch_size: 1 24 | n_sup: 662 25 | workers: 4 26 | mean: [0.485, 0.456, 0.406] 27 | std: [0.229, 0.224, 0.225] 28 | ignore_label: 255 29 | 30 | # # # # # # # # # # # # # # 31 | # 2. training params 32 | # # # # # # # # # # # # # # 33 | trainer: # Required. 34 | epochs: 80 35 | sup_only_epoch: 0 # 0, 1 36 | evaluate_student: True 37 | optimizer: 38 | type: SGD 39 | kwargs: 40 | lr: 0.001 # 0.001:8*2gpus, 4*4gpus; 0.002:8*4gpus. 41 | momentum: 0.9 42 | weight_decay: 0.0001 43 | lr_scheduler: 44 | mode: poly 45 | kwargs: 46 | power: 0.9 47 | # # # # # # # # # # # # # # 48 | # unsupervised loss 49 | # # # # # # # # # # # # # # 50 | unsupervised: 51 | flag_extra_weak: False 52 | threshold: 0.95 # ============================================================================= <>: [0.7, 0.75, 0.8, 0.85, 0.9, 0.95] 53 | loss_weight: 1.0 # ============================================================================= << abl2-weight >> : [1.0, 1.5, 2.0, 2.5, 3.0, 4.0] 54 | #------ 2)cutmix augs ------# 55 | use_cutmix: True 56 | use_cutmix_adaptive: True 57 | use_cutmix_trigger_prob: 1.0 58 | 59 | # # # # # # # # # # # # # # 60 | # 3. output files, and loss 61 | # # # # # # # # # # # # # # 62 | saver: 63 | snapshot_dir: checkpoints 64 | pretrain: '' 65 | use_tb: False 66 | 67 | criterion: 68 | type: CELoss 69 | kwargs: 70 | use_weight: False 71 | 72 | # # # # # # # # # # # # # # 73 | # 4. models 74 | # # # # # # # # # # # # # # 75 | net: # Required. 76 | num_classes: 21 77 | sync_bn: True 78 | ema_decay: 0.999 79 | encoder: 80 | type: augseg.models.resnet.resnet50 81 | pretrain: ./pretrained/resnet50.pth 82 | # type: augseg.models.resnet.resnet101 83 | # pretrain: ./pretrained/resnet101.pth 84 | kwargs: 85 | zero_init_residual: True 86 | multi_grid: True 87 | replace_stride_with_dilation: [False, False, True] # [output_stride = 16] 88 | # replace_stride_with_dilation: [False, True, True] # [output_stride = 8] 89 | decoder: 90 | type: augseg.models.decoder.dec_deeplabv3_plus 91 | kwargs: 92 | inner_planes: 256 93 | low_conv_planes: 48 94 | dilations: [6, 12, 18] # [output_stride = 16] 95 | # dilations: [12, 24, 36] # [output_stride = 8] 96 | -------------------------------------------------------------------------------- /training-logs/6 Semi-citys/R101-1488/config_semi.yaml: -------------------------------------------------------------------------------- 1 | # # # # # # # # # # # # # # 2 | # 1. datasets 3 | # # # # # # # # # # # # # # 4 | dataset: # Required. 5 | type: cityscapes_semi 6 | train: 7 | data_root: ./data/cityscapes 8 | data_list: ./data/splitsall/cityscapes/1488/labeled.txt 9 | batch_size: 4 10 | flip: True 11 | rand_resize: [0.5, 2.0] 12 | resize_base_size: [1024, 2048] 13 | crop: 14 | type: rand 15 | size: [800, 800] 16 | #---- 1) strong data augs ----# 17 | strong_aug: 18 | num_augs: 3 19 | flag_use_random_num_sampling: True # flag_use_wide_range 20 | 21 | val: 22 | data_root: ./data/cityscapes 23 | data_list: ./data/splitsall/cityscapes/val.txt 24 | batch_size: 4 25 | crop: 26 | size: [800, 800] 27 | n_sup: 1488 28 | workers: 4 29 | mean: [0.485, 0.456, 0.406] 30 | std: [0.229, 0.224, 0.225] 31 | ignore_label: 255 32 | 33 | # # # # # # # # # # # # # # 34 | # 2. training params 35 | # # # # # # # # # # # # # # 36 | trainer: # Required. 37 | epochs: 240 38 | sup_only_epoch: 0 # 0, 1 39 | evaluate_student: True 40 | optimizer: 41 | type: SGD 42 | kwargs: 43 | lr: 0.01 # 4GPUs 44 | momentum: 0.9 45 | weight_decay: 0.0005 46 | lr_scheduler: 47 | mode: poly 48 | kwargs: 49 | power: 0.9 50 | unsupervised: 51 | flag_extra_weak: False 52 | threshold: -0.1 # ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ params1: 0.7, 0.75, 0.8, 0.85, 0.9, 0.95 53 | loss_weight: 1.0 # ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ params2: 0.7, 1.0, 1.5, 2.0, 3.0, 4.0 54 | #------ 2)cutmix augs ------# 55 | use_cutmix: True 56 | use_cutmix_adaptive: True 57 | use_cutmix_trigger_prob: 1.0 # wide range, but trigger by 100% 58 | 59 | # # # # # # # # # # # # # # 60 | # 3. output files, and loss 61 | # # # # # # # # # # # # # # 62 | saver: 63 | snapshot_dir: checkpoints 64 | pretrain: '' 65 | use_tb: False 66 | 67 | criterion: 68 | type: ohem 69 | kwargs: 70 | thresh: 0.7 71 | min_kept: 100000 72 | 73 | # # # # # # # # # # # # # # 74 | # 4. models 75 | # # # # # # # # # # # # # # 76 | net: # Required. 77 | num_classes: 19 78 | sync_bn: True 79 | ema_decay: 0.996 80 | # aux_loss: 81 | # aux_plane: 1024 82 | # loss_weight: 0.4 83 | encoder: 84 | type: augseg.models.resnet.resnet101 85 | pretrain: ./pretrained/resnet101.pth 86 | # type: augseg.models.resnet.resnet50 87 | # pretrain: ./pretrained/resnet50.pth 88 | kwargs: 89 | zero_init_residual: True 90 | multi_grid: True 91 | replace_stride_with_dilation: [False, False, True] # [output_stride = 16] 92 | # replace_stride_with_dilation: [False, True, True] # [output_stride = 8] 93 | decoder: 94 | type: augseg.models.decoder.dec_deeplabv3_plus 95 | kwargs: 96 | inner_planes: 256 97 | low_conv_planes: 48 # 256 98 | dilations: [6, 12, 18] 99 | # dilations: [12, 24, 36] # [output_stride = 8] 100 | -------------------------------------------------------------------------------- /training-logs/6 Semi-citys/R101-186/config_semi.yaml: -------------------------------------------------------------------------------- 1 | # # # # # # # # # # # # # # 2 | # 1. datasets 3 | # # # # # # # # # # # # # # 4 | dataset: # Required. 5 | type: cityscapes_semi 6 | train: 7 | data_root: ./data/cityscapes 8 | data_list: ./data/splitsall/cityscapes/186/labeled.txt 9 | batch_size: 4 10 | flip: True 11 | rand_resize: [0.5, 2.0] 12 | resize_base_size: [1024, 2048] 13 | crop: 14 | type: rand 15 | size: [800, 800] 16 | #---- 1) strong data augs ----# 17 | strong_aug: 18 | num_augs: 3 19 | flag_use_random_num_sampling: True # flag_use_wide_range 20 | 21 | val: 22 | data_root: ./data/cityscapes 23 | data_list: ./data/splitsall/cityscapes/val.txt 24 | batch_size: 4 25 | crop: 26 | size: [800, 800] 27 | n_sup: 186 28 | workers: 4 29 | mean: [0.485, 0.456, 0.406] 30 | std: [0.229, 0.224, 0.225] 31 | ignore_label: 255 32 | 33 | # # # # # # # # # # # # # # 34 | # 2. training params 35 | # # # # # # # # # # # # # # 36 | trainer: # Required. 37 | epochs: 240 38 | sup_only_epoch: 0 # 0, 1 39 | evaluate_student: False 40 | optimizer: 41 | type: SGD 42 | kwargs: 43 | lr: 0.01 # 4GPUs 44 | momentum: 0.9 45 | weight_decay: 0.0005 46 | lr_scheduler: 47 | mode: poly 48 | kwargs: 49 | power: 0.9 50 | unsupervised: 51 | flag_extra_weak: False 52 | threshold: -0.1 # ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ params1: 0.7, 0.75, 0.8, 0.85, 0.9, 0.95 53 | loss_weight: 1.0 # ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ params2: 0.7, 1.0, 1.5, 2.0, 3.0, 4.0 54 | #------ 2)cutmix augs ------# 55 | use_cutmix: True 56 | use_cutmix_adaptive: True 57 | use_cutmix_trigger_prob: 1.0 # wide range, but trigger by 100% 58 | 59 | # # # # # # # # # # # # # # 60 | # 3. output files, and loss 61 | # # # # # # # # # # # # # # 62 | saver: 63 | snapshot_dir: checkpoints 64 | pretrain: '' 65 | use_tb: False 66 | 67 | criterion: 68 | type: ohem 69 | kwargs: 70 | thresh: 0.7 71 | min_kept: 100000 72 | 73 | # # # # # # # # # # # # # # 74 | # 4. models 75 | # # # # # # # # # # # # # # 76 | net: # Required. 77 | num_classes: 19 78 | sync_bn: True 79 | ema_decay: 0.996 80 | # aux_loss: 81 | # aux_plane: 1024 82 | # loss_weight: 0.4 83 | encoder: 84 | type: augseg.models.resnet.resnet101 85 | pretrain: ./pretrained/resnet101.pth 86 | # type: augseg.models.resnet.resnet50 87 | # pretrain: ./pretrained/resnet50.pth 88 | kwargs: 89 | zero_init_residual: True 90 | multi_grid: True 91 | replace_stride_with_dilation: [False, False, True] # [output_stride = 16] 92 | # replace_stride_with_dilation: [False, True, True] # [output_stride = 8] 93 | decoder: 94 | type: augseg.models.decoder.dec_deeplabv3_plus 95 | kwargs: 96 | inner_planes: 256 97 | low_conv_planes: 48 # 256 98 | dilations: [6, 12, 18] 99 | # dilations: [12, 24, 36] # [output_stride = 8] 100 | -------------------------------------------------------------------------------- /training-logs/6 Semi-citys/R101-372/config_semi.yaml: -------------------------------------------------------------------------------- 1 | # # # # # # # # # # # # # # 2 | # 1. datasets 3 | # # # # # # # # # # # # # # 4 | dataset: # Required. 5 | type: cityscapes_semi 6 | train: 7 | data_root: ./data/cityscapes 8 | data_list: ./data/splitsall/cityscapes/372/labeled.txt 9 | batch_size: 4 10 | flip: True 11 | rand_resize: [0.5, 2.0] 12 | resize_base_size: [1024, 2048] 13 | crop: 14 | type: rand 15 | size: [800, 800] 16 | #---- 1) strong data augs ----# 17 | strong_aug: 18 | num_augs: 3 19 | flag_use_random_num_sampling: True 20 | val: 21 | data_root: ./data/cityscapes 22 | data_list: ./data/splitsall/cityscapes/val.txt 23 | batch_size: 4 24 | crop: 25 | size: [800, 800] 26 | n_sup: 372 27 | workers: 4 28 | mean: [0.485, 0.456, 0.406] 29 | std: [0.229, 0.224, 0.225] 30 | ignore_label: 255 31 | 32 | # # # # # # # # # # # # # # 33 | # 2. training params 34 | # # # # # # # # # # # # # # 35 | trainer: # Required. 36 | epochs: 240 37 | sup_only_epoch: 0 # 0, 1 38 | evaluate_student: False 39 | optimizer: 40 | type: SGD 41 | kwargs: 42 | lr: 0.01 # 4GPUs 43 | momentum: 0.9 44 | weight_decay: 0.0005 45 | lr_scheduler: 46 | mode: poly 47 | kwargs: 48 | power: 0.9 49 | unsupervised: 50 | flag_extra_weak: False 51 | threshold: -0.7 # ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ params1: 0.7, 0.75, 0.8, 0.85, 0.9, 0.95 52 | loss_weight: 1.0 # ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ params2: 0.7, 1.0, 1.5, 2.0, 3.0, 4.0 53 | #------ 2)cutmix augs ------# 54 | use_cutmix: True 55 | use_cutmix_adaptive: True 56 | use_cutmix_trigger_prob: 1.0 # wide range, but trigger by 100% 57 | 58 | # # # # # # # # # # # # # # 59 | # 3. output files, and loss 60 | # # # # # # # # # # # # # # 61 | saver: 62 | snapshot_dir: checkpoints 63 | pretrain: '' 64 | use_tb: False 65 | 66 | criterion: 67 | type: ohem 68 | kwargs: 69 | thresh: 0.7 70 | min_kept: 100000 71 | 72 | # # # # # # # # # # # # # # 73 | # 4. models 74 | # # # # # # # # # # # # # # 75 | net: # Required. 76 | num_classes: 19 77 | sync_bn: True 78 | ema_decay: 0.996 79 | # aux_loss: 80 | # aux_plane: 1024 81 | # loss_weight: 0.4 82 | encoder: 83 | type: augseg.models.resnet.resnet101 84 | pretrain: ./pretrained/resnet101.pth 85 | # type: augseg.models.resnet.resnet50 86 | # pretrain: ./pretrained/resnet50.pth 87 | kwargs: 88 | zero_init_residual: True 89 | multi_grid: True 90 | replace_stride_with_dilation: [False, False, True] # [output_stride = 16] 91 | # replace_stride_with_dilation: [False, True, True] # [output_stride = 8] 92 | decoder: 93 | type: augseg.models.decoder.dec_deeplabv3_plus 94 | kwargs: 95 | inner_planes: 256 96 | low_conv_planes: 48 # 256 97 | dilations: [6, 12, 18] 98 | # dilations: [12, 24, 36] # [output_stride = 8] 99 | -------------------------------------------------------------------------------- /training-logs/6 Semi-citys/R101-744/config_semi.yaml: -------------------------------------------------------------------------------- 1 | # # # # # # # # # # # # # # 2 | # 1. datasets 3 | # # # # # # # # # # # # # # 4 | dataset: # Required. 5 | type: cityscapes_semi 6 | train: 7 | data_root: ./data/cityscapes 8 | data_list: ./data/splitsall/cityscapes/744/labeled.txt 9 | batch_size: 4 10 | flip: True 11 | rand_resize: [0.5, 2.0] 12 | resize_base_size: [1024, 2048] 13 | crop: 14 | type: rand 15 | size: [800, 800] 16 | #---- 1) strong data augs ----# 17 | strong_aug: 18 | num_augs: 3 19 | flag_use_random_num_sampling: True # flag_use_wide_range 20 | 21 | val: 22 | data_root: ./data/cityscapes 23 | data_list: ./data/splitsall/cityscapes/val.txt 24 | batch_size: 4 25 | crop: 26 | size: [800, 800] 27 | n_sup: 744 28 | workers: 4 29 | mean: [0.485, 0.456, 0.406] 30 | std: [0.229, 0.224, 0.225] 31 | ignore_label: 255 32 | 33 | # # # # # # # # # # # # # # 34 | # 2. training params 35 | # # # # # # # # # # # # # # 36 | trainer: # Required. 37 | epochs: 240 38 | sup_only_epoch: 0 # 0, 1 39 | evaluate_student: True 40 | optimizer: 41 | type: SGD 42 | kwargs: 43 | lr: 0.01 # 4GPUs 44 | momentum: 0.9 45 | weight_decay: 0.0005 46 | lr_scheduler: 47 | mode: poly 48 | kwargs: 49 | power: 0.9 50 | unsupervised: 51 | flag_extra_weak: False 52 | threshold: -0.1 # ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ params1: 0.7, 0.75, 0.8, 0.85, 0.9, 0.95 53 | loss_weight: 1.0 # ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ params2: 0.7, 1.0, 1.5, 2.0, 3.0, 4.0 54 | #------ 2)cutmix augs ------# 55 | use_cutmix: True 56 | use_cutmix_adaptive: True 57 | use_cutmix_trigger_prob: 1.0 # wide range, but trigger by 100% 58 | 59 | # # # # # # # # # # # # # # 60 | # 3. output files, and loss 61 | # # # # # # # # # # # # # # 62 | saver: 63 | snapshot_dir: checkpoints 64 | pretrain: '' 65 | use_tb: False 66 | 67 | criterion: 68 | type: ohem 69 | kwargs: 70 | thresh: 0.7 71 | min_kept: 100000 72 | 73 | # # # # # # # # # # # # # # 74 | # 4. models 75 | # # # # # # # # # # # # # # 76 | net: # Required. 77 | num_classes: 19 78 | sync_bn: True 79 | ema_decay: 0.996 80 | # aux_loss: 81 | # aux_plane: 1024 82 | # loss_weight: 0.4 83 | encoder: 84 | type: augseg.models.resnet.resnet101 85 | pretrain: ./pretrained/resnet101.pth 86 | # type: augseg.models.resnet.resnet50 87 | # pretrain: ./pretrained/resnet50.pth 88 | kwargs: 89 | zero_init_residual: True 90 | multi_grid: True 91 | replace_stride_with_dilation: [False, False, True] # [output_stride = 16] 92 | # replace_stride_with_dilation: [False, True, True] # [output_stride = 8] 93 | decoder: 94 | type: augseg.models.decoder.dec_deeplabv3_plus 95 | kwargs: 96 | inner_planes: 256 97 | low_conv_planes: 48 # 256 98 | dilations: [6, 12, 18] 99 | # dilations: [12, 24, 36] # [output_stride = 8] 100 | -------------------------------------------------------------------------------- /training-logs/6 Semi-citys/R50-1488/config_semi.yaml: -------------------------------------------------------------------------------- 1 | # # # # # # # # # # # # # # 2 | # 1. datasets 3 | # # # # # # # # # # # # # # 4 | dataset: # Required. 5 | type: cityscapes_semi 6 | train: 7 | data_root: ./data/cityscapes 8 | data_list: ./data/splitsall/cityscapes/1488/labeled.txt 9 | batch_size: 4 10 | flip: True 11 | rand_resize: [0.5, 2.0] 12 | resize_base_size: [1024, 2048] 13 | crop: 14 | type: rand 15 | size: [800, 800] 16 | #---- 1) strong data augs ----# 17 | strong_aug: 18 | num_augs: 3 19 | flag_use_random_num_sampling: True 20 | val: 21 | data_root: ./data/cityscapes 22 | data_list: ./data/splitsall/cityscapes/val.txt 23 | batch_size: 4 24 | crop: 25 | size: [800, 800] 26 | n_sup: 1488 27 | workers: 4 28 | mean: [0.485, 0.456, 0.406] 29 | std: [0.229, 0.224, 0.225] 30 | ignore_label: 255 31 | 32 | # # # # # # # # # # # # # # 33 | # 2. training params 34 | # # # # # # # # # # # # # # 35 | trainer: # Required. 36 | epochs: 240 37 | sup_only_epoch: 0 # 0, 1 38 | evaluate_student: True 39 | optimizer: 40 | type: SGD 41 | kwargs: 42 | lr: 0.01 # 4GPUs 43 | momentum: 0.9 44 | weight_decay: 0.0005 45 | lr_scheduler: 46 | mode: poly 47 | kwargs: 48 | power: 0.9 49 | unsupervised: 50 | flag_extra_weak: False 51 | threshold: -0.7 # ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ params1: 0.7, 0.75, 0.8, 0.85, 0.9, 0.95 52 | loss_weight: 1.0 # ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ params2: 0.7, 1.0, 1.5, 2.0, 3.0, 4.0 53 | #------ 2)cutmix augs ------# 54 | use_cutmix: True 55 | use_cutmix_adaptive: True 56 | use_cutmix_trigger_prob: 1.0 # wide range, but trigger by 100% 57 | 58 | # # # # # # # # # # # # # # 59 | # 3. output files, and loss 60 | # # # # # # # # # # # # # # 61 | saver: 62 | snapshot_dir: checkpoints 63 | pretrain: '' 64 | use_tb: False 65 | 66 | criterion: 67 | type: ohem 68 | kwargs: 69 | thresh: 0.7 70 | min_kept: 100000 71 | 72 | # # # # # # # # # # # # # # 73 | # 4. models 74 | # # # # # # # # # # # # # # 75 | net: # Required. 76 | num_classes: 19 77 | sync_bn: True 78 | ema_decay: 0.996 79 | # aux_loss: 80 | # aux_plane: 1024 81 | # loss_weight: 0.4 82 | encoder: 83 | # type: augseg.models.resnet.resnet101 84 | # pretrain: ./pretrained/resnet101.pth 85 | type: augseg.models.resnet.resnet50 86 | pretrain: ./pretrained/resnet50.pth 87 | kwargs: 88 | zero_init_residual: True 89 | multi_grid: True 90 | replace_stride_with_dilation: [False, False, True] # [output_stride = 16] 91 | # replace_stride_with_dilation: [False, True, True] # [output_stride = 8] 92 | decoder: 93 | type: augseg.models.decoder.dec_deeplabv3_plus 94 | kwargs: 95 | inner_planes: 256 96 | low_conv_planes: 48 # 256 97 | dilations: [6, 12, 18] 98 | # dilations: [12, 24, 36] # [output_stride = 8] 99 | -------------------------------------------------------------------------------- /training-logs/6 Semi-citys/R50-186/config_semi.yaml: -------------------------------------------------------------------------------- 1 | # # # # # # # # # # # # # # 2 | # 1. datasets 3 | # # # # # # # # # # # # # # 4 | dataset: # Required. 5 | type: cityscapes_semi 6 | train: 7 | data_root: ./data/cityscapes 8 | data_list: ./data/splitsall/cityscapes/186/labeled.txt 9 | batch_size: 4 10 | flip: True 11 | rand_resize: [0.5, 2.0] 12 | resize_base_size: [1024, 2048] 13 | crop: 14 | type: rand 15 | size: [800, 800] 16 | #---- 1) strong data augs ----# 17 | strong_aug: 18 | num_augs: 3 19 | flag_use_random_num_sampling: True # flag_use_wide_range 20 | 21 | val: 22 | data_root: ./data/cityscapes 23 | data_list: ./data/splitsall/cityscapes/val.txt 24 | batch_size: 4 25 | crop: 26 | size: [800, 800] 27 | n_sup: 186 28 | workers: 4 29 | mean: [0.485, 0.456, 0.406] 30 | std: [0.229, 0.224, 0.225] 31 | ignore_label: 255 32 | 33 | # # # # # # # # # # # # # # 34 | # 2. training params 35 | # # # # # # # # # # # # # # 36 | trainer: # Required. 37 | epochs: 240 38 | sup_only_epoch: 0 # 0, 1 39 | evaluate_student: False 40 | optimizer: 41 | type: SGD 42 | kwargs: 43 | lr: 0.01 # 4GPUs 44 | momentum: 0.9 45 | weight_decay: 0.0005 46 | lr_scheduler: 47 | mode: poly 48 | kwargs: 49 | power: 0.9 50 | unsupervised: 51 | flag_extra_weak: False 52 | threshold: -0.1 # ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ params1: 0.7, 0.75, 0.8, 0.85, 0.9, 0.95 53 | loss_weight: 1.0 # ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ params2: 0.7, 1.0, 1.5, 2.0, 3.0, 4.0 54 | #------ 2)cutmix augs ------# 55 | use_cutmix: True 56 | use_cutmix_adaptive: True 57 | use_cutmix_trigger_prob: 1.0 # wide range, but trigger by 100% 58 | 59 | # # # # # # # # # # # # # # 60 | # 3. output files, and loss 61 | # # # # # # # # # # # # # # 62 | saver: 63 | snapshot_dir: checkpoints 64 | pretrain: '' 65 | use_tb: False 66 | 67 | criterion: 68 | type: ohem 69 | kwargs: 70 | thresh: 0.7 71 | min_kept: 100000 72 | 73 | # # # # # # # # # # # # # # 74 | # 4. models 75 | # # # # # # # # # # # # # # 76 | net: # Required. 77 | num_classes: 19 78 | sync_bn: True 79 | ema_decay: 0.996 80 | # aux_loss: 81 | # aux_plane: 1024 82 | # loss_weight: 0.4 83 | encoder: 84 | # type: augseg.models.resnet.resnet101 85 | # pretrain: ./pretrained/resnet101.pth 86 | type: augseg.models.resnet.resnet50 87 | pretrain: ./pretrained/resnet50.pth 88 | kwargs: 89 | zero_init_residual: True 90 | multi_grid: True 91 | replace_stride_with_dilation: [False, False, True] # [output_stride = 16] 92 | # replace_stride_with_dilation: [False, True, True] # [output_stride = 8] 93 | decoder: 94 | type: augseg.models.decoder.dec_deeplabv3_plus 95 | kwargs: 96 | inner_planes: 256 97 | low_conv_planes: 48 # 256 98 | dilations: [6, 12, 18] 99 | # dilations: [12, 24, 36] # [output_stride = 8] 100 | -------------------------------------------------------------------------------- /training-logs/6 Semi-citys/R50-372/config_semi.yaml: -------------------------------------------------------------------------------- 1 | # # # # # # # # # # # # # # 2 | # 1. datasets 3 | # # # # # # # # # # # # # # 4 | dataset: # Required. 5 | type: cityscapes_semi 6 | train: 7 | data_root: ./data/cityscapes 8 | data_list: ./data/splitsall/cityscapes/372/labeled.txt 9 | batch_size: 4 10 | flip: True 11 | rand_resize: [0.5, 2.0] 12 | resize_base_size: [1024, 2048] 13 | crop: 14 | type: rand 15 | size: [800, 800] 16 | #---- 1) strong data augs ----# 17 | strong_aug: 18 | num_augs: 3 19 | flag_use_random_num_sampling: True 20 | val: 21 | data_root: ./data/cityscapes 22 | data_list: ./data/splitsall/cityscapes/val.txt 23 | batch_size: 4 24 | crop: 25 | size: [800, 800] 26 | n_sup: 372 27 | workers: 4 28 | mean: [0.485, 0.456, 0.406] 29 | std: [0.229, 0.224, 0.225] 30 | ignore_label: 255 31 | 32 | # # # # # # # # # # # # # # 33 | # 2. training params 34 | # # # # # # # # # # # # # # 35 | trainer: # Required. 36 | epochs: 240 37 | sup_only_epoch: 0 # 0, 1 38 | evaluate_student: True 39 | optimizer: 40 | type: SGD 41 | kwargs: 42 | lr: 0.01 # 4GPUs 43 | momentum: 0.9 44 | weight_decay: 0.0005 45 | lr_scheduler: 46 | mode: poly 47 | kwargs: 48 | power: 0.9 49 | unsupervised: 50 | flag_extra_weak: False 51 | threshold: -0.1 # ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ params1: 0.7, 0.75, 0.8, 0.85, 0.9, 0.95 52 | loss_weight: 1.0 # ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ params2: 0.7, 1.0, 1.5, 2.0, 3.0, 4.0 53 | #------ 2)cutmix augs ------# 54 | use_cutmix: True 55 | use_cutmix_adaptive: True 56 | use_cutmix_trigger_prob: 1.0 # wide range, but trigger by 100% 57 | 58 | # # # # # # # # # # # # # # 59 | # 3. output files, and loss 60 | # # # # # # # # # # # # # # 61 | saver: 62 | snapshot_dir: checkpoints 63 | pretrain: '' 64 | use_tb: False 65 | 66 | criterion: 67 | type: ohem 68 | kwargs: 69 | thresh: 0.7 70 | min_kept: 100000 71 | 72 | # # # # # # # # # # # # # # 73 | # 4. models 74 | # # # # # # # # # # # # # # 75 | net: # Required. 76 | num_classes: 19 77 | sync_bn: True 78 | ema_decay: 0.996 79 | # aux_loss: 80 | # aux_plane: 1024 81 | # loss_weight: 0.4 82 | encoder: 83 | # type: augseg.models.resnet.resnet101 84 | # pretrain: ./pretrained/resnet101.pth 85 | type: augseg.models.resnet.resnet50 86 | pretrain: ./pretrained/resnet50.pth 87 | kwargs: 88 | zero_init_residual: True 89 | multi_grid: True 90 | replace_stride_with_dilation: [False, False, True] # [output_stride = 16] 91 | # replace_stride_with_dilation: [False, True, True] # [output_stride = 8] 92 | decoder: 93 | type: augseg.models.decoder.dec_deeplabv3_plus 94 | kwargs: 95 | inner_planes: 256 96 | low_conv_planes: 48 # 256 97 | dilations: [6, 12, 18] 98 | # dilations: [12, 24, 36] # [output_stride = 8] 99 | -------------------------------------------------------------------------------- /training-logs/6 Semi-citys/R50-744/config_semi.yaml: -------------------------------------------------------------------------------- 1 | # # # # # # # # # # # # # # 2 | # 1. datasets 3 | # # # # # # # # # # # # # # 4 | dataset: # Required. 5 | type: cityscapes_semi 6 | train: 7 | data_root: ./data/cityscapes 8 | data_list: ./data/splitsall/cityscapes/744/labeled.txt 9 | batch_size: 4 10 | flip: True 11 | rand_resize: [0.5, 2.0] 12 | resize_base_size: [1024, 2048] 13 | crop: 14 | type: rand 15 | size: [800, 800] 16 | #---- 1) strong data augs ----# 17 | strong_aug: 18 | num_augs: 3 19 | flag_use_random_num_sampling: True 20 | val: 21 | data_root: ./data/cityscapes 22 | data_list: ./data/splitsall/cityscapes/val.txt 23 | batch_size: 4 24 | crop: 25 | size: [800, 800] 26 | n_sup: 744 27 | workers: 4 28 | mean: [0.485, 0.456, 0.406] 29 | std: [0.229, 0.224, 0.225] 30 | ignore_label: 255 31 | 32 | # # # # # # # # # # # # # # 33 | # 2. training params 34 | # # # # # # # # # # # # # # 35 | trainer: # Required. 36 | epochs: 240 37 | sup_only_epoch: 0 # 0, 1 38 | evaluate_student: False 39 | optimizer: 40 | type: SGD 41 | kwargs: 42 | lr: 0.01 # 4GPUs 43 | momentum: 0.9 44 | weight_decay: 0.0005 45 | lr_scheduler: 46 | mode: poly 47 | kwargs: 48 | power: 0.9 49 | unsupervised: 50 | flag_extra_weak: False 51 | threshold: -0.7 # ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ params1: 0.7, 0.75, 0.8, 0.85, 0.9, 0.95 52 | loss_weight: 1.0 # ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ params2: 0.7, 1.0, 1.5, 2.0, 3.0, 4.0 53 | #------ 2)cutmix augs ------# 54 | use_cutmix: True 55 | use_cutmix_adaptive: True 56 | use_cutmix_trigger_prob: 1.0 # wide range, but trigger by 100% 57 | 58 | # # # # # # # # # # # # # # 59 | # 3. output files, and loss 60 | # # # # # # # # # # # # # # 61 | saver: 62 | snapshot_dir: checkpoints 63 | pretrain: '' 64 | use_tb: False 65 | 66 | criterion: 67 | type: ohem 68 | kwargs: 69 | thresh: 0.7 70 | min_kept: 100000 71 | 72 | # # # # # # # # # # # # # # 73 | # 4. models 74 | # # # # # # # # # # # # # # 75 | net: # Required. 76 | num_classes: 19 77 | sync_bn: True 78 | ema_decay: 0.996 79 | # aux_loss: 80 | # aux_plane: 1024 81 | # loss_weight: 0.4 82 | encoder: 83 | # type: augseg.models.resnet.resnet101 84 | # pretrain: ./pretrained/resnet101.pth 85 | type: augseg.models.resnet.resnet50 86 | pretrain: ./pretrained/resnet50.pth 87 | kwargs: 88 | zero_init_residual: True 89 | multi_grid: True 90 | replace_stride_with_dilation: [False, False, True] # [output_stride = 16] 91 | # replace_stride_with_dilation: [False, True, True] # [output_stride = 8] 92 | decoder: 93 | type: augseg.models.decoder.dec_deeplabv3_plus 94 | kwargs: 95 | inner_planes: 256 96 | low_conv_planes: 48 # 256 97 | dilations: [6, 12, 18] 98 | # dilations: [12, 24, 36] # [output_stride = 8] 99 | --------------------------------------------------------------------------------