├── onnx2tflite ├── onnx2tflite.egg-info │ ├── dependency_links.txt │ ├── top_level.txt │ ├── requires.txt │ ├── SOURCES.txt │ └── PKG-INFO ├── onnx2tflite │ ├── __init__.py │ ├── utils │ │ ├── __init__.py │ │ ├── op_registry.py │ │ ├── dimension_utils.py │ │ ├── definitions.py │ │ └── graph_tools.py │ ├── layers │ │ ├── __init__.py │ │ ├── activations_layers.py │ │ ├── mathematics_layers.py │ │ ├── deformation_layers.py │ │ └── common_layers.py │ ├── components │ │ ├── __init__.py │ │ ├── dataloader.py │ │ ├── output_check.py │ │ ├── builder.py │ │ ├── onnx_loader.py │ │ └── builder1.py │ ├── __main__.py │ └── converter.py ├── build │ └── lib │ │ └── onnx2tflite │ │ ├── __init__.py │ │ ├── __main__.py │ │ └── converter.py ├── dist │ └── onnx2tflite-2.0-py3.8.egg ├── onnx2tflite.py ├── setup.py └── test │ ├── test_squeeze.py │ ├── test_reshape_transpose.py │ ├── test_concat.py │ └── test_torchvison.py ├── figs └── framework.png ├── pretrain ├── lolv1.onnx ├── lolv1.tflite ├── zrr_best_slim.pkl ├── lolv1_best_slim.pkl ├── uieb_best_slim.pkl ├── lolv2_real_best_slim.pkl └── mai25_isp_challenge_best_slim.pkl ├── datasets ├── lle │ ├── gt │ │ ├── 1.png │ │ └── 22.png │ └── input │ │ ├── 1.png │ │ └── 22.png ├── isp │ ├── gt │ │ ├── 10.png │ │ └── 15.png │ └── Input │ │ ├── 10.png │ │ └── 15.png └── uie │ ├── gt │ ├── 3_img_.png │ └── 8_img_.png │ └── input │ ├── 3_img_.png │ └── 8_img_.png ├── onnx_to_tf.py ├── tf_to_TFLite.py ├── config ├── isp.yaml └── lle.yaml ├── data ├── lledata.py ├── ispdata.py └── __init__.py ├── model ├── __init__.py ├── lle.py ├── isp.py ├── utils.py └── utils_IWO.py ├── logger.py ├── option.py ├── complexity.py ├── torch_to_onnx.py ├── test_TFLite_RGB.py ├── test_TFLite_ISP.py ├── README.md ├── loss.py ├── main.py └── LICENSE /onnx2tflite/onnx2tflite.egg-info/dependency_links.txt: -------------------------------------------------------------------------------- 1 | 2 | -------------------------------------------------------------------------------- /onnx2tflite/onnx2tflite.egg-info/top_level.txt: -------------------------------------------------------------------------------- 1 | onnx2tflite 2 | -------------------------------------------------------------------------------- /figs/framework.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/AVC2-UESTC/MobileIE/HEAD/figs/framework.png -------------------------------------------------------------------------------- /pretrain/lolv1.onnx: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/AVC2-UESTC/MobileIE/HEAD/pretrain/lolv1.onnx -------------------------------------------------------------------------------- /datasets/lle/gt/1.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/AVC2-UESTC/MobileIE/HEAD/datasets/lle/gt/1.png -------------------------------------------------------------------------------- /pretrain/lolv1.tflite: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/AVC2-UESTC/MobileIE/HEAD/pretrain/lolv1.tflite -------------------------------------------------------------------------------- /datasets/isp/gt/10.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/AVC2-UESTC/MobileIE/HEAD/datasets/isp/gt/10.png -------------------------------------------------------------------------------- /datasets/isp/gt/15.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/AVC2-UESTC/MobileIE/HEAD/datasets/isp/gt/15.png -------------------------------------------------------------------------------- /datasets/lle/gt/22.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/AVC2-UESTC/MobileIE/HEAD/datasets/lle/gt/22.png -------------------------------------------------------------------------------- /datasets/lle/input/1.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/AVC2-UESTC/MobileIE/HEAD/datasets/lle/input/1.png -------------------------------------------------------------------------------- /onnx2tflite/onnx2tflite/__init__.py: -------------------------------------------------------------------------------- 1 | __VERSION__ = "2.0" 2 | 3 | from .converter import onnx_converter -------------------------------------------------------------------------------- /datasets/isp/Input/10.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/AVC2-UESTC/MobileIE/HEAD/datasets/isp/Input/10.png -------------------------------------------------------------------------------- /datasets/isp/Input/15.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/AVC2-UESTC/MobileIE/HEAD/datasets/isp/Input/15.png -------------------------------------------------------------------------------- /datasets/lle/input/22.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/AVC2-UESTC/MobileIE/HEAD/datasets/lle/input/22.png -------------------------------------------------------------------------------- /datasets/uie/gt/3_img_.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/AVC2-UESTC/MobileIE/HEAD/datasets/uie/gt/3_img_.png -------------------------------------------------------------------------------- /datasets/uie/gt/8_img_.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/AVC2-UESTC/MobileIE/HEAD/datasets/uie/gt/8_img_.png -------------------------------------------------------------------------------- /pretrain/zrr_best_slim.pkl: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/AVC2-UESTC/MobileIE/HEAD/pretrain/zrr_best_slim.pkl -------------------------------------------------------------------------------- /datasets/uie/input/3_img_.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/AVC2-UESTC/MobileIE/HEAD/datasets/uie/input/3_img_.png -------------------------------------------------------------------------------- /datasets/uie/input/8_img_.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/AVC2-UESTC/MobileIE/HEAD/datasets/uie/input/8_img_.png -------------------------------------------------------------------------------- /onnx2tflite/build/lib/onnx2tflite/__init__.py: -------------------------------------------------------------------------------- 1 | __VERSION__ = "2.0" 2 | 3 | from .converter import onnx_converter -------------------------------------------------------------------------------- /pretrain/lolv1_best_slim.pkl: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/AVC2-UESTC/MobileIE/HEAD/pretrain/lolv1_best_slim.pkl -------------------------------------------------------------------------------- /pretrain/uieb_best_slim.pkl: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/AVC2-UESTC/MobileIE/HEAD/pretrain/uieb_best_slim.pkl -------------------------------------------------------------------------------- /pretrain/lolv2_real_best_slim.pkl: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/AVC2-UESTC/MobileIE/HEAD/pretrain/lolv2_real_best_slim.pkl -------------------------------------------------------------------------------- /onnx2tflite/dist/onnx2tflite-2.0-py3.8.egg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/AVC2-UESTC/MobileIE/HEAD/onnx2tflite/dist/onnx2tflite-2.0-py3.8.egg -------------------------------------------------------------------------------- /onnx2tflite/onnx2tflite/utils/__init__.py: -------------------------------------------------------------------------------- 1 | from .dimension_utils import * 2 | from .op_registry import OPERATOR 3 | from .definitions import * -------------------------------------------------------------------------------- /pretrain/mai25_isp_challenge_best_slim.pkl: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/AVC2-UESTC/MobileIE/HEAD/pretrain/mai25_isp_challenge_best_slim.pkl -------------------------------------------------------------------------------- /onnx2tflite/onnx2tflite.egg-info/requires.txt: -------------------------------------------------------------------------------- 1 | onnx 2 | onnxruntime 3 | onnx-simplifier 4 | numpy<=1.24 5 | tensorflow<2.13,>=2.5 6 | opencv-python 7 | -------------------------------------------------------------------------------- /onnx2tflite/onnx2tflite/layers/__init__.py: -------------------------------------------------------------------------------- 1 | from .conv_layers import * 2 | from .common_layers import * 3 | from .activations_layers import * 4 | from .mathematics_layers import * 5 | from .deformation_layers import * -------------------------------------------------------------------------------- /onnx2tflite/onnx2tflite/components/__init__.py: -------------------------------------------------------------------------------- 1 | from .output_check import get_elements_error 2 | from .onnx_loader import load_onnx_modelproto 3 | from .builder import keras_builder, tflite_builder 4 | 5 | __all__ = ['load_onnx_modelproto', 'keras_builder', 'tflite_builder', 'get_elements_error'] -------------------------------------------------------------------------------- /onnx2tflite/onnx2tflite.egg-info/SOURCES.txt: -------------------------------------------------------------------------------- 1 | setup.py 2 | onnx2tflite/__init__.py 3 | onnx2tflite/__main__.py 4 | onnx2tflite/converter.py 5 | onnx2tflite.egg-info/PKG-INFO 6 | onnx2tflite.egg-info/SOURCES.txt 7 | onnx2tflite.egg-info/dependency_links.txt 8 | onnx2tflite.egg-info/requires.txt 9 | onnx2tflite.egg-info/top_level.txt 10 | test/test_concat.py 11 | test/test_reshape_transpose.py 12 | test/test_squeeze.py 13 | test/test_torchvison.py -------------------------------------------------------------------------------- /onnx_to_tf.py: -------------------------------------------------------------------------------- 1 | import onnx 2 | from onnx_tf.backend import prepare 3 | import os 4 | os.environ['CUDA_VISIBLE_DEVICES'] = '0' 5 | 6 | onnx_model_path = './LLE.onnx' 7 | onnx_model = onnx.load(onnx_model_path) 8 | 9 | onnx.checker.check_model(onnx_model) 10 | print("ONNX to TensorFlow") 11 | 12 | try: 13 | tf_rep = prepare(onnx_model) 14 | tf_model_path = 'lle_tf' 15 | tf_rep.export_graph(tf_model_path) 16 | print(f"Success, and save to {tf_model_path}") 17 | except Exception as e: 18 | print(f"ERROR: {e}") 19 | 20 | -------------------------------------------------------------------------------- /tf_to_TFLite.py: -------------------------------------------------------------------------------- 1 | import tensorflow as tf 2 | import os 3 | 4 | os.environ['CUDA_VISIBLE_DEVICES'] = '0' 5 | 6 | saved_model_dir = "./lle_tf" 7 | converter = tf.lite.TFLiteConverter.from_saved_model(saved_model_dir) 8 | converter.target_spec.supported_ops = [tf.lite.OpsSet.TFLITE_BUILTINS, tf.lite.OpsSet.SELECT_TF_OPS] 9 | 10 | tflite_model = converter.convert() 11 | 12 | tflite_model_path = "LLE.tflite" 13 | with open(tflite_model_path, 'wb') as f: 14 | f.write(tflite_model) 15 | 16 | print(f"TFLite save to {tflite_model_path}") 17 | -------------------------------------------------------------------------------- /onnx2tflite/onnx2tflite.py: -------------------------------------------------------------------------------- 1 | import os 2 | os.environ['CUDA_VISIBLE_DEVICES'] = '2' 3 | from onnx2tflite.converter import onnx_converter 4 | onnx_path = "/data2/yanhailong/IR-Based/ICCV2025/MobileIE/LLE.onnx" 5 | 6 | onnx_converter( 7 | onnx_model_path = onnx_path, 8 | need_simplify = True, 9 | output_path = "//data2/yanhailong/IR-Based/ICCV2025/MobileIE/", 10 | target_formats = ['tflite'], # or ['keras'], ['keras', 'tflite'] 11 | weight_quant = False, 12 | int8_model = False, 13 | int8_mean = None, 14 | int8_std = None, 15 | image_root = None 16 | ) -------------------------------------------------------------------------------- /onnx2tflite/setup.py: -------------------------------------------------------------------------------- 1 | import os 2 | from setuptools import setup, find_packages 3 | abs_path = os.path.dirname(os.path.abspath(__file__)) 4 | 5 | setup( 6 | name="onnx2tflite", 7 | version="2.0", 8 | author="MPolaris", 9 | description="onnx to keras/tensorflow lite", 10 | long_description=open(os.path.join(abs_path, "readme.md")).read(), 11 | long_description_content_type='text/markdown', 12 | packages=find_packages(include=['onnx2tflite']), 13 | license="Apache-2.0", 14 | platforms=["Windows", "linux"], 15 | install_requires=open(os.path.join(abs_path, "requirements.txt")).read().splitlines() 16 | ) -------------------------------------------------------------------------------- /config/isp.yaml: -------------------------------------------------------------------------------- 1 | exp_name: isp 2 | 3 | train: 4 | warmup: false 5 | warmup_epoch: 10 6 | lr_warmup: 1e-6 7 | train_inp: ./isp/train/huawei_raw 8 | train_gt: ./isp/train/canon 9 | valid_inp: ./isp/test/huawei_raw 10 | valid_gt: ./isp/test/canon 11 | batch_size: 6 12 | epoch: 1000 13 | lr: 1e-3 14 | num_workers: 20 15 | save_every: 20 16 | save_slim: true 17 | 18 | test: 19 | test_inp: ./isp/test/huawei_raw 20 | test_gt: ./isp/test/canon 21 | num_workers: 0 22 | save: false 23 | 24 | demo: 25 | demo_inp: ./traindata/isp/test/huawei_raw 26 | num_workers: 0 27 | 28 | model: 29 | type: original # [original, re-parameterized] 30 | pretrained: false 31 | need_slim: false #true 32 | rep_scale: 4 33 | channels: 12 34 | -------------------------------------------------------------------------------- /config/lle.yaml: -------------------------------------------------------------------------------- 1 | exp_name: lle 2 | 3 | train: 4 | warmup: True 5 | warmup_epoch: 10 6 | lr_warmup: 1e-6 7 | train_inp: ./lowlight/LOLdataset/our485/low 8 | train_gt: ./lowlight/LOLdataset/our485/high 9 | valid_inp: ./lowlight/LOLdataset/eval15/low 10 | valid_gt: ./lowlight/LOLdataset/eval15/high 11 | batch_size: 4 12 | epoch: 2000 13 | lr: 1e-3 14 | num_workers: 0 15 | save_every: 20 16 | save_slim: true 17 | 18 | test: 19 | test_inp: ./lowlight/LOLdataset/eval15/low 20 | test_gt: ./lowlight/LOLdataset/eval15/high 21 | num_workers: 0 22 | save: false 23 | 24 | demo: 25 | demo_inp: ./lowlight/LOLdataset/eval15/low 26 | num_workers: 0 27 | 28 | model: 29 | type: original # [original, re-parameterized]: 30 | pretrained: false 31 | need_slim: false # true 32 | rep_scale: 4 33 | channels: 12 34 | -------------------------------------------------------------------------------- /data/lledata.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import numpy as np 3 | import os 4 | from PIL import Image 5 | 6 | 7 | class LLEData(torch.utils.data.Dataset): 8 | def __init__(self, opt, inp_path, gt_path=None): 9 | super(LLEData, self).__init__() 10 | self.img_li = [path for path in os.listdir(inp_path)] 11 | self.inp_path = inp_path 12 | self.gt_path = gt_path 13 | self.opt = opt 14 | 15 | def __getitem__(self, index): 16 | inp = Image.open(os.path.join(self.inp_path, self.img_li[index])) 17 | inp = np.array(inp).transpose([2, 0, 1]) 18 | inp = inp.astype(np.float32) / 255 19 | 20 | inp = torch.Tensor(np.array(inp)) 21 | inp = inp.to(self.opt.device) 22 | 23 | if self.gt_path: # gt_path -> train/test not demo 24 | gt = Image.open(os.path.join(self.gt_path, self.img_li[index])) 25 | gt = np.array(gt).transpose([2, 0, 1]) 26 | gt = gt.astype(np.float32) / 255 27 | 28 | gt = torch.Tensor(np.array(gt)) 29 | gt = gt.to(self.opt.device) 30 | 31 | return inp, gt, self.img_li[index].split('.')[0] 32 | return inp, self.img_li[index].split('.')[0] 33 | 34 | def __len__(self): 35 | return len(self.img_li) 36 | -------------------------------------------------------------------------------- /onnx2tflite/onnx2tflite/utils/op_registry.py: -------------------------------------------------------------------------------- 1 | class Registry(object): 2 | def __init__(self, name) -> None: 3 | self._name = name 4 | self._operator_dict = dict() 5 | 6 | def __len__(self): 7 | return len(self._operator_dict) 8 | 9 | @property 10 | def name(self): 11 | return self._name 12 | 13 | @property 14 | def operator_dict(self): 15 | return self._operator_dict 16 | 17 | def get(self, key): 18 | return self._operator_dict.get(key, None) 19 | 20 | def _register_operator(self, op_class, op_name=None): 21 | if (not isinstance(op_name, str)) or op_name is None: 22 | op_name = op_class.__name__ 23 | 24 | if self._operator_dict.get(op_name, None): 25 | raise KeyError(f'{op_name} is already registered in {self._name}') 26 | 27 | self._operator_dict[op_name] = op_class 28 | 29 | def register_operator(self, name=None, op_class=None): 30 | if op_class is not None: 31 | self._register_operator(op_class, name) 32 | return op_class 33 | 34 | def _register(cls): 35 | self._register_operator(cls, name) 36 | return cls 37 | 38 | return _register 39 | 40 | OPERATOR = Registry("TensorflowOP") -------------------------------------------------------------------------------- /onnx2tflite/test/test_squeeze.py: -------------------------------------------------------------------------------- 1 | import os 2 | import pytest 3 | 4 | import torch 5 | import torch.nn as nn 6 | from onnx2tflite import onnx_converter 7 | 8 | MODEL_ROOT = "./unit_test" 9 | os.makedirs(MODEL_ROOT, exist_ok=True) 10 | 11 | @pytest.mark.filterwarnings('ignore::UserWarning') 12 | @pytest.mark.filterwarnings('ignore::DeprecationWarning') 13 | def test_squeeze(): 14 | class Squeeze(nn.Module): 15 | def __init__(self, *args, **kwargs) -> None: 16 | super().__init__(*args, **kwargs) 17 | 18 | def forward(self, x): 19 | x = torch.unsqueeze(x, dim=1) 20 | # x = torch.tile(x, dims=(2,1,1)) 21 | x = torch.squeeze(x, dim=1) 22 | 23 | return x 24 | 25 | model = Squeeze() 26 | x = torch.randn(1,1,1,2) 27 | 28 | onnx_model_path = os.path.join(MODEL_ROOT, "test_squeeze.onnx") 29 | torch.onnx.export(model, x, onnx_model_path, opset_version=11) 30 | 31 | res = onnx_converter( 32 | onnx_model_path = onnx_model_path, 33 | need_simplify = True, 34 | output_path = MODEL_ROOT, 35 | target_formats = ['tflite'], 36 | native_groupconv=False, 37 | fp16_model=False, 38 | int8_model=False, 39 | ) 40 | 41 | assert res['tflite_error'] < 1e-3 -------------------------------------------------------------------------------- /model/__init__.py: -------------------------------------------------------------------------------- 1 | import torch 2 | from importlib import import_module 3 | from .lle import MobileIELLENet, MobileIELLENetS 4 | from .isp import MobileIEISPNet, MobileIEISPNetS 5 | 6 | __all__ = { 7 | 'MobileIELLENet', 8 | 'MobileIELLENetS', 9 | 'MobileIEISPNet', 10 | 'MobileIEISPNetS', 11 | 'import_model' 12 | } 13 | 14 | def import_model(opt): 15 | model_name = 'MobileIE'+opt.model_task.upper() 16 | kwargs = {'channels': opt.config['model']['channels']} 17 | 18 | if opt.config['model']['type'] == 're-parameterized': 19 | model_name += 'NetS' 20 | elif opt.config['model']['type'] == 'original': 21 | model_name += 'Net' 22 | kwargs['rep_scale'] = opt.config['model']['rep_scale'] 23 | else: 24 | raise ValueError('unknown model type, please choose from [original, re-parameterized]') 25 | 26 | model = getattr(import_module('model'), model_name)(**kwargs) 27 | model = model.to(opt.device) 28 | 29 | if opt.config['model']['pretrained']: 30 | #model.load_state_dict(torch.load(opt.config['model']['pretrained'])) 31 | model.load_state_dict(torch.load(opt.config['model']['pretrained']), strict=False) 32 | 33 | if opt.config['model']['type'] == 'original' and opt.config['model']['need_slim'] is True: 34 | model = model.slim().to(opt.device) 35 | return model 36 | -------------------------------------------------------------------------------- /onnx2tflite/test/test_reshape_transpose.py: -------------------------------------------------------------------------------- 1 | import os 2 | import pytest 3 | 4 | import torch 5 | import torch.nn as nn 6 | from onnx2tflite import onnx_converter 7 | 8 | MODEL_ROOT = "./unit_test" 9 | os.makedirs(MODEL_ROOT, exist_ok=True) 10 | 11 | @pytest.mark.filterwarnings('ignore::UserWarning') 12 | @pytest.mark.filterwarnings('ignore::DeprecationWarning') 13 | def test_reshape_trans(): 14 | class test1(nn.Module): 15 | def __init__(self, *args, **kwargs) -> None: 16 | super().__init__(*args, **kwargs) 17 | self.conv1 = nn.Conv2d(3, 3, 3, 2, 1) 18 | self.conv2 = nn.Conv2d(3, 3, 3, 2, 1) 19 | 20 | def forward(self, x): 21 | x = torch.reshape(x, (1, 3, 32, 16)) 22 | # x = torch.transpose(x, (0, 1, 3, 2)) 23 | x = torch.transpose(x, 3, 2) 24 | x = self.conv1(x) 25 | x = self.conv2(x) 26 | return x 27 | 28 | model = test1() 29 | x = torch.randn(1, 3*32*16) 30 | 31 | onnx_model_path = os.path.join(MODEL_ROOT, "test_reshape_trans.onnx") 32 | torch.onnx.export(model, x, onnx_model_path, opset_version=11) 33 | 34 | res = onnx_converter( 35 | onnx_model_path = onnx_model_path, 36 | need_simplify = True, 37 | output_path = MODEL_ROOT, 38 | target_formats = ['tflite'], 39 | native_groupconv=False, 40 | fp16_model=False, 41 | int8_model = False, 42 | ) 43 | 44 | assert res['tflite_error'] < 1e-3 -------------------------------------------------------------------------------- /data/ispdata.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import numpy as np 3 | import os 4 | from PIL import Image 5 | 6 | 7 | class ISPData(torch.utils.data.Dataset): 8 | def __init__(self, opt, raw_path, rgb_path=None): 9 | super(ISPData, self).__init__() 10 | self.img_li = [path for path in os.listdir(raw_path)] 11 | self.raw_path = raw_path 12 | self.rgb_path = rgb_path 13 | self.opt = opt 14 | 15 | def __getitem__(self, index): 16 | raw = Image.open(os.path.join(self.raw_path, self.img_li[index])) 17 | raw = np.array(raw) 18 | raw = self.bayer2rggb(raw) 19 | raw = raw.astype(np.float32) / 4095 20 | 21 | raw = torch.Tensor(np.array(raw)) 22 | raw = raw.to(self.opt.device) 23 | 24 | if self.rgb_path: # gt_path -> train/test not demo 25 | rgb = Image.open(os.path.join(self.rgb_path, self.img_li[index])) 26 | rgb = np.array(rgb).transpose([2, 0, 1]) 27 | rgb = rgb.astype(np.float32) / 255 28 | 29 | rgb = torch.Tensor(np.array(rgb)) 30 | rgb = rgb.to(self.opt.device) 31 | 32 | return raw, rgb, self.img_li[index].split('.')[0] 33 | 34 | return raw, self.img_li[index].split('.')[0] 35 | 36 | def __len__(self): 37 | return len(self.img_li) 38 | 39 | def bayer2rggb(self, img_bayer): 40 | h, w = img_bayer.shape 41 | img_bayer = img_bayer.reshape(h // 2, 2, w // 2, 2) 42 | img_bayer = img_bayer.transpose([1, 3, 0, 2]).reshape([-1, h // 2, w // 2]) 43 | return img_bayer 44 | -------------------------------------------------------------------------------- /logger.py: -------------------------------------------------------------------------------- 1 | import logging 2 | 3 | 4 | class Logger: 5 | 6 | def __init__( 7 | self, 8 | opt, 9 | logging_level=logging.INFO, 10 | file_level=logging.INFO, 11 | stream_level=logging.INFO 12 | ): 13 | self.opt = opt 14 | self.log_path = opt.log_path 15 | self.logging_level = logging_level 16 | 17 | self.file_level = file_level 18 | self.stream_level = stream_level 19 | 20 | self.logger = logging.getLogger('logger.log') 21 | self.logger.setLevel(self.logging_level) 22 | 23 | self.configure() 24 | 25 | def configure(self): 26 | log_format = logging.Formatter('%(asctime)s %(levelname)s: %(message)s') 27 | 28 | stream_handler = logging.StreamHandler() 29 | stream_handler.setLevel(self.stream_level) 30 | stream_handler.setFormatter(log_format) 31 | 32 | file_handler = logging.FileHandler(self.log_path) 33 | file_handler.setLevel(self.file_level) 34 | file_handler.setFormatter(log_format) 35 | 36 | self.logger.addHandler(file_handler) 37 | self.logger.addHandler(stream_handler) 38 | 39 | def debug(self, message): 40 | self.logger.debug(message) 41 | 42 | def info(self, message): 43 | self.logger.info(message) 44 | 45 | def warn(self, message): 46 | self.logger.warning(message) 47 | 48 | def error(self, message): 49 | self.logger.error(message) 50 | 51 | def critical(self, message): 52 | self.logger.critical(message) 53 | -------------------------------------------------------------------------------- /onnx2tflite/test/test_concat.py: -------------------------------------------------------------------------------- 1 | import os 2 | import pytest 3 | 4 | import torch 5 | import torch.nn as nn 6 | from onnx2tflite import onnx_converter 7 | 8 | MODEL_ROOT = "./unit_test" 9 | os.makedirs(MODEL_ROOT, exist_ok=True) 10 | 11 | @pytest.mark.filterwarnings('ignore::UserWarning') 12 | @pytest.mark.filterwarnings('ignore::DeprecationWarning') 13 | def test_concat(): 14 | class Concat(nn.Module): 15 | def __init__(self, *args, **kwargs) -> None: 16 | super().__init__(*args, **kwargs) 17 | self.conv1 = nn.Conv2d(3, 3, 3, 2, 1) 18 | # self.conv2 = nn.Conv2d(3, 3, 3, 2, 1) 19 | self._const = torch.randn(1,2,16,8) 20 | 21 | def forward(self, x1, x2, x3): 22 | x1 = torch.reshape(x1, (1, 3, 16, 8)) 23 | # x = torch.transpose(x, (0, 1, 3, 2)) 24 | x2 = torch.transpose(x2, 3, 2) 25 | x3 = self.conv1(x3) 26 | x = torch.concat([x1,x2,x3,self._const], dim=1) 27 | return x 28 | 29 | model = Concat() 30 | x1 = torch.randn(1,3*16*8) 31 | x2 = torch.randn(1,3,8,16) 32 | x3 = torch.randn(1,3,32,16) 33 | 34 | onnx_model_path = os.path.join(MODEL_ROOT, "test_concat.onnx") 35 | torch.onnx.export(model, (x1,x2,x3), onnx_model_path, opset_version=11) 36 | 37 | res = onnx_converter( 38 | onnx_model_path = onnx_model_path, 39 | need_simplify = True, 40 | output_path = MODEL_ROOT, 41 | target_formats = ['tflite'], 42 | native_groupconv=False, 43 | fp16_model=False, 44 | int8_model=False, 45 | ) 46 | 47 | assert res['tflite_error'] < 1e-3 -------------------------------------------------------------------------------- /onnx2tflite/onnx2tflite/utils/dimension_utils.py: -------------------------------------------------------------------------------- 1 | import tensorflow as tf 2 | ''' 3 | shape and axis transform utils func. 4 | ''' 5 | def channel_to_last_dimension(axis): 6 | ''' 7 | make channel first to channel last 8 | ''' 9 | if axis == 0: 10 | axis = 0 11 | elif axis == 1: 12 | axis = -1 13 | else: 14 | axis -= 1 15 | return axis 16 | 17 | def shape_NCD_to_NDC_format(shape): 18 | ''' 19 | make shape format from channel first to channel last 20 | ''' 21 | if len(shape) <= 2: 22 | return tuple(shape) 23 | new_shape = [shape[0], *shape[2:], shape[1]] 24 | return tuple(new_shape) 25 | 26 | def shape_NDC_to_NCD_format(shape): 27 | ''' 28 | make shape format from channel last to channel first 29 | ''' 30 | if len(shape) <= 2: 31 | return tuple(shape) 32 | new_shape = [shape[0], shape[-1], *shape[1:-1]] 33 | return tuple(new_shape) 34 | 35 | def tensor_NCD_to_NDC_format(tensor): 36 | ''' 37 | make tensor format from channel first to channel last 38 | ''' 39 | if(len(tensor.shape) > 2): 40 | shape = [i for i in range(len(tensor.shape))] 41 | shape = shape_NCD_to_NDC_format(shape) 42 | tensor = tf.transpose(tensor, perm=shape) 43 | return tensor 44 | 45 | def tensor_NDC_to_NCD_format(tensor): 46 | ''' 47 | make tensor format from channel last to channel first 48 | ''' 49 | if(len(tensor.shape) > 2): 50 | shape = [i for i in range(len(tensor.shape))] 51 | shape = shape_NDC_to_NCD_format(shape) 52 | tensor = tf.transpose(tensor, perm=shape) 53 | return tensor 54 | 55 | def intfloat_to_list(x:int or float, lens:int): 56 | if isinstance(x, (int, float)): 57 | return [x]*lens 58 | else: 59 | return x -------------------------------------------------------------------------------- /onnx2tflite/onnx2tflite/utils/definitions.py: -------------------------------------------------------------------------------- 1 | import tensorflow as tf 2 | from abc import ABC 3 | from enum import Enum, unique 4 | 5 | @unique 6 | class Layout(Enum): 7 | Default = 0 8 | Channel_First = 1 << 0# for onnx format 9 | Channel_Last = 1 << 1 # for tensorflow format 10 | Channel_None = 1 << 2 # no channel 11 | 12 | class Node_Layout: 13 | def __init__(self, name:str, pre:list=[], nxt:list=[]) -> None: 14 | self.name = name 15 | self.pre = pre 16 | self.nxt = nxt 17 | self.layout = Layout.Default 18 | 19 | class BaseOP(ABC): 20 | def __init__(self, tensor_graph, const_weights, node_attributes, node_inputs, node_outputs, layout_dict) -> None: 21 | pass 22 | 23 | onnx2tf_type = { 24 | 1: tf.float32, # ONNX_FLOAT 25 | 2: tf.uint8, # ONNX_UINT8 26 | 3: tf.int8, # ONNX_INT8 27 | 4: tf.uint16, # ONNX_UINT16 28 | 5: tf.int16, # ONNX_INT16 29 | 6: tf.int32, # ONNX_INT32 30 | 7: tf.int64, # ONNX_INT64 31 | 8: tf.string, # ONNX_STRING 32 | 9: tf.bool, # ONNX_BOOL 33 | 10: tf.float16, # ONNX_FLOAT16 34 | 11: tf.float64, # ONNX_DOUBLE 35 | 12: tf.uint32, # ONNX_UINT32 36 | 13: tf.uint64, # ONNX_UINT64 37 | 14: tf.complex64, # ONNX_COMPLEX64 38 | 15: tf.complex128 # ONNX_COMPLEX128 39 | } 40 | 41 | np2tf_type = { 42 | "int32": tf.int32, 43 | "int64": tf.int64, 44 | "float32": tf.float32, 45 | "float64": tf.float64, 46 | "bool": tf.bool, 47 | "uint8": tf.uint8, 48 | "int8": tf.int8, 49 | "int16": tf.int16, 50 | "uint16": tf.uint16, 51 | "uint32": tf.uint32, 52 | "uint64": tf.uint64, 53 | "complex64": tf.complex64, 54 | "complex128": tf.complex128 55 | } 56 | 57 | FORCE_CHANNEL_LAST_OP = ["Conv", "ConvTranspose", "DepthToSpace", "Pad", "AveragePool", "MaxPool", "Upsample", "Resize", "Gemm"] 58 | FORCE_CHANNEL_FIRST_OP = ["Reshape", "Transpose", "ScatterND", "MatMul"] 59 | 60 | -------------------------------------------------------------------------------- /onnx2tflite/test/test_torchvison.py: -------------------------------------------------------------------------------- 1 | ''' 2 | unit test for torchvision models 3 | ''' 4 | import os 5 | import pytest 6 | 7 | import torch 8 | import torchvision 9 | from onnx2tflite import onnx_converter 10 | 11 | MODEL_ROOT = "./unit_test" 12 | os.makedirs(MODEL_ROOT, exist_ok=True) 13 | 14 | @pytest.mark.filterwarnings('ignore::UserWarning') 15 | @pytest.mark.filterwarnings('ignore::DeprecationWarning') 16 | def test_resnet(): 17 | model = torchvision.models.resnet18(False) 18 | onnx_model_path = os.path.join(MODEL_ROOT, "resnet18.onnx") 19 | torch.onnx.export(model, torch.randn(1, 3, 224, 224), onnx_model_path, opset_version=13) 20 | error = onnx_converter(onnx_model_path, need_simplify = True, output_path = MODEL_ROOT, target_formats = ['tflite'])['tflite_error'] 21 | assert error < 1e-3 22 | 23 | @pytest.mark.filterwarnings('ignore::UserWarning') 24 | @pytest.mark.filterwarnings('ignore::DeprecationWarning') 25 | def test_mobilenet(): 26 | model = torchvision.models.mobilenet_v2(False) 27 | onnx_model_path = os.path.join(MODEL_ROOT, "mobilenet_v2.onnx") 28 | torch.onnx.export(model, torch.randn(1, 3, 224, 224), onnx_model_path, opset_version=13) 29 | error = onnx_converter(onnx_model_path, need_simplify = True, output_path = MODEL_ROOT, target_formats = ['tflite'])['tflite_error'] 30 | assert error < 1e-3 31 | 32 | @pytest.mark.filterwarnings('ignore::UserWarning') 33 | @pytest.mark.filterwarnings('ignore::DeprecationWarning') 34 | def test_deeplabv3(): 35 | model = torchvision.models.segmentation.deeplabv3_resnet50(False) 36 | onnx_model_path = os.path.join(MODEL_ROOT, "deeplabv3_resnet50.onnx") 37 | torch.onnx.export(model, torch.randn(1, 3, 512, 1024), onnx_model_path, opset_version=13) 38 | error = onnx_converter(onnx_model_path, need_simplify = True, output_path = MODEL_ROOT, target_formats = ['tflite'])['tflite_error'] 39 | assert error < 1e-3 40 | 41 | @pytest.mark.filterwarnings('ignore::UserWarning') 42 | @pytest.mark.filterwarnings('ignore::DeprecationWarning') 43 | def test_vit(): 44 | model = torchvision.models.vit_b_16(False) 45 | onnx_model_path = os.path.join(MODEL_ROOT, "vit_b_16.onnx") 46 | torch.onnx.export(model, torch.randn(1, 3, 224, 224), onnx_model_path, opset_version=13) 47 | error = onnx_converter(onnx_model_path, need_simplify = True, output_path = MODEL_ROOT, target_formats = ['tflite'])['tflite_error'] 48 | assert error < 1e-3 -------------------------------------------------------------------------------- /option.py: -------------------------------------------------------------------------------- 1 | import os 2 | import argparse 3 | import yaml 4 | from datetime import datetime 5 | 6 | 7 | def get_option(): 8 | parser = argparse.ArgumentParser() 9 | parser.add_argument( 10 | '-task', 11 | default='train', 12 | type=str, 13 | choices=['train', 'test', 'demo'], 14 | help='choose the task for running the model' 15 | ) 16 | parser.add_argument( 17 | '-model_task', 18 | default='isp', 19 | type=str, 20 | choices=['isp', 'lle', 'sr'], 21 | help='the model of the task' 22 | ) 23 | parser.add_argument( 24 | '-device', 25 | default='cuda', 26 | type=str, 27 | help='choose the device to run the model' 28 | ) 29 | opt = parser.parse_args() 30 | opt = opt_format(opt) 31 | return opt 32 | 33 | 34 | def load_yaml(path): 35 | with open(path, 'r') as f: 36 | model_config = yaml.load(f, Loader=yaml.FullLoader) 37 | return model_config 38 | 39 | 40 | def save_yaml(path, file_dict): 41 | with open(path, 'w') as f: 42 | f.write(yaml.dump(file_dict, allow_unicode=True)) 43 | 44 | 45 | def opt_format(opt): 46 | opt.root = os.getcwd() 47 | opt.config = r'{}/config/{}.yaml'.format(opt.root, opt.model_task) 48 | opt.config = load_yaml(opt.config) 49 | 50 | proper_time = str(datetime.now()).split('.')[0].replace(':', '-') 51 | 52 | opt.config['exp_name'] = '{}_{}'.format(opt.task, opt.config['exp_name']) 53 | 54 | opt.experiments = r'{}/experiments/{}'.format(opt.root, '{} {}'.format(proper_time, opt.config['exp_name'])) 55 | if not os.path.exists(opt.experiments): 56 | os.mkdir(opt.experiments) 57 | 58 | config_path = r'{}/config.yaml'.format(opt.experiments) 59 | save_yaml(config_path, opt.config) 60 | 61 | if opt.task == 'demo' or (opt.task == 'test' and opt.config['test']['save'] != False): 62 | opt.save_image = True 63 | opt.save_image_dir = r'{}/{}'.format(opt.experiments, 'images') 64 | if not os.path.exists(opt.save_image_dir): 65 | os.mkdir(opt.save_image_dir) 66 | 67 | opt.log_path = r'{}/logger.log'.format(opt.experiments) 68 | 69 | if opt.task == 'train': 70 | opt.save_model = True 71 | opt.save_model_dir = r'{}/{}'.format(opt.experiments, 'models') 72 | if not os.path.exists(opt.save_model_dir): 73 | os.mkdir(opt.save_model_dir) 74 | 75 | return opt 76 | -------------------------------------------------------------------------------- /complexity.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import time 3 | from thop import profile, clever_format 4 | from model import lle 5 | import os 6 | os.environ['CUDA_VISIBLE_DEVICES'] = '0' 7 | 8 | width = 640 9 | height = 480 10 | 11 | 12 | def compute_FLOPs_and_model_size(model, width, height): 13 | input = torch.randn(1, 3, width, height).cuda() 14 | macs, params = profile(model, inputs=(input,), verbose=False) 15 | return macs, params 16 | 17 | @torch.no_grad() 18 | def compute_fps_and_inference_time(model, shape, epoch=100, warmup=10, device=None): 19 | total_time = 0.0 20 | 21 | if not device: 22 | device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu") 23 | model = model.to(device) 24 | 25 | model.eval() # Switch to evaluation mode 26 | 27 | # Warm-up iterations 28 | for _ in range(warmup): 29 | data = torch.randn(shape).to(device) 30 | model(data) 31 | 32 | # Actual timing iterations 33 | for _ in range(epoch): 34 | data = torch.randn(shape).to(device) 35 | 36 | start = time.time() 37 | outputs = model(data) 38 | torch.cuda.synchronize() # Ensure CUDA has finished all tasks 39 | end = time.time() 40 | 41 | total_time += (end - start) 42 | 43 | avg_inference_time = total_time / epoch 44 | fps = epoch / total_time 45 | 46 | return fps, avg_inference_time 47 | 48 | def test_model_flops(width, height): 49 | model = lle.MobileIES(channels=12) 50 | model.cuda() 51 | 52 | FLOPs, params = compute_FLOPs_and_model_size(model, width, height) 53 | 54 | model_size = params * 4.0 / 1024 / 1024 55 | flops, params = clever_format([FLOPs, params], "%.3f") 56 | 57 | print('Number of parameters: {}'.format(params)) 58 | print('Size of model: {:.2f} MB'.format(model_size)) 59 | print('Computational complexity: {} FLOPs'.format(flops)) 60 | 61 | def test_fps_and_inference_time(width, height): 62 | model = lle.MobileIES(channels=12) 63 | model.cuda() 64 | 65 | device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu') 66 | fps, avg_inference_time = compute_fps_and_inference_time(model, (1, 3, width, height), device=device) 67 | print('device: {} - fps: {:.3f}, average inference time per frame: {:.6f} seconds'.format(device.type, fps, avg_inference_time)) 68 | 69 | if __name__ == '__main__': 70 | test_model_flops(width, height) 71 | test_fps_and_inference_time(width, height) 72 | 73 | -------------------------------------------------------------------------------- /onnx2tflite/onnx2tflite/__main__.py: -------------------------------------------------------------------------------- 1 | import argparse 2 | from .converter import onnx_converter 3 | 4 | def parse_opt(): 5 | parser = argparse.ArgumentParser() 6 | parser.add_argument('--weights', type=str, required=True, help='onnx model path') 7 | parser.add_argument('--outpath', type=str, default=None, help='tflite model save path') 8 | parser.add_argument('--input-node-names', nargs="+", default=None, help='which inputs is you want, support middle layers, None will using onnx orignal inputs') 9 | parser.add_argument('--output-node-names', nargs="+", default=None, help='which outputs is you want, support middle layers, None will using onnx orignal outputs') 10 | parser.add_argument('--nosimplify', default=False, action='store_true', help='do not simplify model') 11 | parser.add_argument("--native-groupconv", default=False, action='store_true', help='using native method for groupconv, only support for tflite version >= 2.9') 12 | parser.add_argument('--weigthquant', default=False, action='store_true', help='weight only int8 quant') 13 | parser.add_argument('--fp16', default=False, action='store_true', help='fp16 quant, include input output') 14 | parser.add_argument('--int8', default=False, action='store_true', help='int8 quant, include input output') 15 | parser.add_argument('--imgroot', type=str, default=None, help='when int8=True, imgroot should give for calculating running_mean and running_norm') 16 | parser.add_argument('--int8mean', type=float, nargs='+', default=[123.675, 116.28, 103.53], help='int8 image preprocesses mean, float or list') 17 | parser.add_argument('--int8std', type=float, nargs='+', default=[58.395, 57.12, 57.375], help='int8 image preprocesses std, float or list') 18 | parser.add_argument('--formats', nargs='+', default=['keras', 'tflite'], help='available formats are (h5, tflite)') 19 | opt = parser.parse_args() 20 | return opt 21 | 22 | def run(): 23 | opt = parse_opt() 24 | onnx_converter( 25 | onnx_model_path = opt.weights, 26 | need_simplify = not opt.nosimplify, 27 | input_node_names = opt.input_node_names, 28 | output_node_names = opt.output_node_names, 29 | output_path = opt.outpath, 30 | target_formats = opt.formats, 31 | native_groupconv = opt.native_groupconv, 32 | weight_quant=opt.weigthquant, 33 | fp16_model=opt.fp16, 34 | int8_model=opt.int8, 35 | int8_mean=opt.int8mean, 36 | int8_std=opt.int8std, 37 | image_root=opt.imgroot 38 | ) 39 | 40 | if __name__ == "__main__": 41 | run() -------------------------------------------------------------------------------- /onnx2tflite/build/lib/onnx2tflite/__main__.py: -------------------------------------------------------------------------------- 1 | import argparse 2 | from .converter import onnx_converter 3 | 4 | def parse_opt(): 5 | parser = argparse.ArgumentParser() 6 | parser.add_argument('--weights', type=str, required=True, help='onnx model path') 7 | parser.add_argument('--outpath', type=str, default=None, help='tflite model save path') 8 | parser.add_argument('--input-node-names', nargs="+", default=None, help='which inputs is you want, support middle layers, None will using onnx orignal inputs') 9 | parser.add_argument('--output-node-names', nargs="+", default=None, help='which outputs is you want, support middle layers, None will using onnx orignal outputs') 10 | parser.add_argument('--nosimplify', default=False, action='store_true', help='do not simplify model') 11 | parser.add_argument("--native-groupconv", default=False, action='store_true', help='using native method for groupconv, only support for tflite version >= 2.9') 12 | parser.add_argument('--weigthquant', default=False, action='store_true', help='weight only int8 quant') 13 | parser.add_argument('--fp16', default=False, action='store_true', help='fp16 quant, include input output') 14 | parser.add_argument('--int8', default=False, action='store_true', help='int8 quant, include input output') 15 | parser.add_argument('--imgroot', type=str, default=None, help='when int8=True, imgroot should give for calculating running_mean and running_norm') 16 | parser.add_argument('--int8mean', type=float, nargs='+', default=[123.675, 116.28, 103.53], help='int8 image preprocesses mean, float or list') 17 | parser.add_argument('--int8std', type=float, nargs='+', default=[58.395, 57.12, 57.375], help='int8 image preprocesses std, float or list') 18 | parser.add_argument('--formats', nargs='+', default=['keras', 'tflite'], help='available formats are (h5, tflite)') 19 | opt = parser.parse_args() 20 | return opt 21 | 22 | def run(): 23 | opt = parse_opt() 24 | onnx_converter( 25 | onnx_model_path = opt.weights, 26 | need_simplify = not opt.nosimplify, 27 | input_node_names = opt.input_node_names, 28 | output_node_names = opt.output_node_names, 29 | output_path = opt.outpath, 30 | target_formats = opt.formats, 31 | native_groupconv = opt.native_groupconv, 32 | weight_quant=opt.weigthquant, 33 | fp16_model=opt.fp16, 34 | int8_model=opt.int8, 35 | int8_mean=opt.int8mean, 36 | int8_std=opt.int8std, 37 | image_root=opt.imgroot 38 | ) 39 | 40 | if __name__ == "__main__": 41 | run() -------------------------------------------------------------------------------- /torch_to_onnx.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import torch.nn as nn 3 | 4 | class MobileIENetS(nn.Module): 5 | def __init__(self, channels): 6 | super(MobileIENetS, self).__init__() 7 | self.head = FST( 8 | nn.Sequential( 9 | nn.Conv2d(3, channels, 5, 1, 2), 10 | nn.PReLU(channels), 11 | nn.Conv2d(channels, channels, 3, 1, 1) 12 | ), 13 | channels 14 | ) 15 | self.body = FST( 16 | nn.Conv2d(channels, channels, 3, 1, 1), 17 | channels 18 | ) 19 | self.att = nn.Sequential( 20 | nn.AdaptiveAvgPool2d(1), 21 | nn.Conv2d(channels, channels, 1), 22 | nn.Sigmoid() 23 | ) 24 | self.att1 = nn.Sequential( 25 | nn.Conv2d(1, channels, 1, 1), 26 | nn.Sigmoid() 27 | ) 28 | self.tail = nn.Conv2d(channels, 3, 3, 1, 1) 29 | 30 | def forward(self, x): 31 | x0 = self.head(x) 32 | x1 = self.body(x0) 33 | x2 = self.att(x1) 34 | max_out, _ = torch.max(x2 * x1, dim=1, keepdim=True) 35 | x3 = self.att1(max_out) 36 | x4 = torch.mul(x3, x2) * x1 37 | return self.tail(x4) 38 | 39 | class FST(nn.Module): 40 | def __init__(self, block1, channels): 41 | super(FST, self).__init__() 42 | self.block1 = block1 43 | self.weight1 = nn.Parameter(torch.randn(1)) 44 | self.weight2 = nn.Parameter(torch.randn(1)) 45 | self.bias = nn.Parameter(torch.randn((1, channels, 1, 1))) 46 | 47 | def forward(self, x): 48 | x1 = self.block1(x) 49 | weighted_block1 = self.weight1 * x1 50 | weighted_block2 = self.weight2 * x1 51 | return weighted_block1 * weighted_block2 + self.bias 52 | 53 | def export_onnx(pretrained_model_path): 54 | model = MobileIENetS(12) 55 | 56 | checkpoint = torch.load(pretrained_model_path) 57 | model.load_state_dict(checkpoint) 58 | model.eval() 59 | 60 | dummy_input = torch.randn(1, 3, 400, 600) 61 | 62 | torch.onnx.export( 63 | model, 64 | dummy_input, 65 | "LLE.onnx", 66 | opset_version=12, 67 | export_params=True, 68 | do_constant_folding=True, 69 | input_names=['input'], 70 | output_names=['output'], 71 | dynamic_axes=None 72 | ) 73 | print("ONNX Success.") 74 | 75 | if __name__ == "__main__": 76 | pretrained_model_path = r'./pretrain/lolv1_best_slim.pkl' 77 | export_onnx(pretrained_model_path) 78 | -------------------------------------------------------------------------------- /data/__init__.py: -------------------------------------------------------------------------------- 1 | from torch.utils import data 2 | from importlib import import_module 3 | import torchvision.transforms as transforms 4 | 5 | from .ispdata import ISPData 6 | from .lledata import LLEData 7 | 8 | __all__ = { 9 | 'ISPData', 10 | 'LLEData', 11 | 'import_loader' 12 | } 13 | 14 | 15 | def import_loader(opt): 16 | dataset_name = opt.model_task.upper()+'Data' 17 | dataset = getattr(import_module('data'), dataset_name) 18 | 19 | if opt.task == 'train': 20 | train_inp_path = opt.config['train']['train_inp'] 21 | train_gt_path = opt.config['train']['train_gt'] 22 | valid_inp_path = opt.config['train']['valid_inp'] 23 | valid_gt_path = opt.config['train']['valid_gt'] 24 | 25 | 26 | train_data = dataset(opt, train_inp_path, train_gt_path) 27 | #train_data = dataset(opt, train_inp_path, train_gt_path, transform=train_transform) 28 | 29 | if opt.model_task == 'sr': 30 | valid_data = dataset(opt, valid_inp_path, valid_gt_path, 'valid') 31 | else: 32 | valid_data = dataset(opt, valid_inp_path, valid_gt_path) 33 | 34 | train_loader = data.DataLoader( 35 | train_data, 36 | batch_size=opt.config['train']['batch_size'], 37 | shuffle=True, 38 | num_workers=opt.config['train']['num_workers'], 39 | drop_last=True, 40 | ) 41 | valid_loader = data.DataLoader( 42 | valid_data, 43 | batch_size=1, 44 | shuffle=False, 45 | num_workers=opt.config['train']['num_workers'], 46 | drop_last=False, 47 | ) 48 | return train_loader, valid_loader 49 | 50 | elif opt.task == 'test': 51 | inp_test_path = opt.config['test']['test_inp'] 52 | gt_test_path = opt.config['test']['test_gt'] 53 | 54 | test_data = dataset(opt, inp_test_path, gt_test_path) 55 | test_loader = data.DataLoader( 56 | test_data, 57 | batch_size=1, 58 | shuffle=False, 59 | num_workers=opt.config['test']['num_workers'], 60 | drop_last=False, 61 | ) 62 | return test_loader 63 | 64 | elif opt.task == 'demo': 65 | inp_demo_path = opt.config['demo']['demo_inp'] 66 | demo_data = dataset(opt, inp_demo_path) 67 | demo_loader = data.DataLoader( 68 | demo_data, 69 | batch_size=1, 70 | shuffle=False, 71 | num_workers=opt.config['demo']['num_workers'], 72 | drop_last=False, 73 | ) 74 | return demo_loader 75 | 76 | else: 77 | raise ValueError('unknown task, please choose from [train, test, demo]') 78 | -------------------------------------------------------------------------------- /test_TFLite_RGB.py: -------------------------------------------------------------------------------- 1 | import numpy as np 2 | from tensorflow import lite 3 | import os 4 | from PIL import Image 5 | 6 | def load_model(model_path): 7 | interpreter = lite.Interpreter(model_path=model_path) 8 | interpreter.allocate_tensors() 9 | return interpreter 10 | 11 | def print_model_structure(interpreter): 12 | print("Model Structure:") 13 | input_details = interpreter.get_input_details() 14 | output_details = interpreter.get_output_details() 15 | 16 | print("\nInput details:") 17 | for input_tensor in input_details: 18 | print(f"Name: {input_tensor['name']}, Shape: {input_tensor['shape']}, Type: {input_tensor['dtype']}") 19 | 20 | print("\nOutput details:") 21 | for output_tensor in output_details: 22 | print(f"Name: {output_tensor['name']}, Shape: {output_tensor['shape']}, Type: {output_tensor['dtype']}") 23 | 24 | def preprocess_image(image_path): 25 | #target_size = (1024, 1024) 26 | img = Image.open(image_path).convert("RGB") 27 | #img = img.resize(target_size, Image.BICUBIC) 28 | img_array = np.array(img).astype(np.float32) / 255.0 29 | img_array = np.expand_dims(img_array, axis=0) 30 | print(img_array.shape) 31 | return img_array 32 | 33 | def inference_and_save_results(interpreter, input_image_folder, output_image_folder): 34 | input_details = interpreter.get_input_details() 35 | output_details = interpreter.get_output_details() 36 | 37 | if not os.path.exists(output_image_folder): 38 | os.makedirs(output_image_folder) 39 | 40 | for image_name in os.listdir(input_image_folder): 41 | image_path = os.path.join(input_image_folder, image_name) 42 | if not image_path.lower().endswith(('png', 'jpg', 'jpeg')): 43 | continue 44 | 45 | img_array = preprocess_image(image_path) 46 | 47 | interpreter.set_tensor(input_details[0]['index'], img_array) 48 | interpreter.invoke() 49 | output_img = interpreter.get_tensor(output_details[0]['index']) 50 | 51 | output_img = np.clip(output_img, 0., 1.) 52 | output_img = np.squeeze(output_img) 53 | output_img = (output_img * 255).astype(np.uint8) 54 | 55 | pil_img = Image.fromarray(output_img) 56 | output_image_path = os.path.join(output_image_folder, image_name) 57 | pil_img.save(output_image_path) 58 | 59 | def main(): 60 | model_path = './LLE.tflite' 61 | input_image_folder = './lowlight/LOLdataset/eval15/low' 62 | output_image_folder = './experiments/results' 63 | 64 | interpreter = load_model(model_path) 65 | print_model_structure(interpreter) 66 | inference_and_save_results(interpreter, input_image_folder, output_image_folder) 67 | 68 | if __name__ == "__main__": 69 | main() 70 | -------------------------------------------------------------------------------- /onnx2tflite/onnx2tflite/components/dataloader.py: -------------------------------------------------------------------------------- 1 | import os 2 | import cv2 3 | import logging 4 | import numpy as np 5 | 6 | LOG = logging.getLogger("Quantization DataLoder :") 7 | 8 | class RandomLoader(object): 9 | def __init__(self, target_size): 10 | self.target_size = target_size 11 | LOG.warning(f"Generate quantization data from random, it's will lead to accuracy problem!") 12 | 13 | def __iter__(self): 14 | self.index = 0 15 | return self 16 | 17 | def __next__(self): 18 | if self.index > 5: 19 | raise StopIteration() 20 | self.index += 1 21 | return [np.random.randn(*self.target_size).astype(np.float32)] 22 | 23 | class ImageLoader(object): 24 | ''' 25 | generate data for quantization from image datas. 26 | img_quan_data = (img - mean)/std, it's important for accuracy of model. 27 | ''' 28 | VALID_FORMAT = ['.jpg', '.png', '.jpeg'] 29 | 30 | def __init__(self, img_root, target_size, mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375]) -> None: 31 | assert os.path.exists(img_root), F"{img_root} is not exists, please check!" 32 | self.fns = os.listdir(img_root) 33 | self.fns = list(filter(lambda fn: os.path.splitext(fn)[-1].lower() in self.VALID_FORMAT, self.fns)) 34 | self.nums = len(self.fns) 35 | assert self.nums > 0, f"No images detected in {img_root}." 36 | if self.nums > 100: 37 | LOG.warning(f"{self.nums} images detected, the number of recommended images is less than 100.") 38 | else: 39 | LOG.info(f"{self.nums} images detected.") 40 | self.fns = [os.path.join(img_root, fn) for fn in self.fns] 41 | 42 | self.batch, self.size = target_size[0], target_size[1:-1] 43 | if isinstance(mean, list): 44 | mean = np.array(mean, dtype=np.float32) 45 | if isinstance(std, list): 46 | std = np.array(std, dtype=np.float32) 47 | self.mean, self.std = mean, std 48 | 49 | def __iter__(self): 50 | self.index = 0 51 | return self 52 | 53 | def __next__(self): 54 | if self.index >= self.nums: 55 | raise StopIteration() 56 | 57 | _input = cv2.imread(self.fns[self.index]) 58 | _input = cv2.resize(_input, self.size)[:, :, ::-1]#BGR->RGB 59 | _input = _input.astype(np.float32) 60 | 61 | if self.mean is not None: 62 | _input = (_input - self.mean) 63 | if self.std is not None: 64 | _input = _input/self.std 65 | 66 | _input = np.expand_dims(_input, axis=0) 67 | if self.batch > 1: 68 | _input = np.repeat(_input, self.batch, axis=0).astype(np.float32) 69 | 70 | self.index += 1 71 | return [_input] 72 | -------------------------------------------------------------------------------- /onnx2tflite/onnx2tflite/utils/graph_tools.py: -------------------------------------------------------------------------------- 1 | from onnx import numpy_helper 2 | import tensorflow as tf 3 | from tensorflow import keras 4 | from .definitions import * 5 | 6 | # copy from https://github.com/gmalivenko/onnx2keras 7 | def decode_node_attribute(node)->dict: 8 | """ 9 | Parse ONNX attributes to Python dictionary 10 | :param args: ONNX attributes object 11 | :return: Python dictionary 12 | """ 13 | def onnx_attribute_to_dict(onnx_attr): 14 | """ 15 | Parse ONNX attribute 16 | :param onnx_attr: ONNX attribute 17 | :return: Python data type 18 | """ 19 | if onnx_attr.HasField('t'): 20 | return numpy_helper.to_array(getattr(onnx_attr, 't')) 21 | 22 | for attr_type in ['f', 'i']: 23 | if onnx_attr.HasField(attr_type): 24 | return getattr(onnx_attr, attr_type) 25 | 26 | # s need to be decode, bytes to string 27 | if onnx_attr.HasField('s'): 28 | return getattr(onnx_attr, 's').decode() 29 | 30 | for attr_type in ['floats', 'ints', 'strings']: 31 | if getattr(onnx_attr, attr_type): 32 | return list(getattr(onnx_attr, attr_type)) 33 | return {arg.name: onnx_attribute_to_dict(arg) for arg in node.attribute} 34 | 35 | def build_tf_inputs(model_graph, layout_dict:dict): 36 | inputs_name = [] 37 | for inp in model_graph.input: 38 | input_shape = [x.dim_value for x in inp.type.tensor_type.shape.dim] 39 | if input_shape == []: 40 | continue 41 | inputs_name.append(inp.name) 42 | layout_dict[inp.name] = Layout.Default 43 | if len(input_shape) < 3: 44 | layout_dict[inp.name] = Layout.Channel_None 45 | 46 | _inputs_name = inputs_name.copy() 47 | for node in model_graph.node: 48 | op_name, node_inputs = node.op_type, node.input 49 | # output_layout = Layout.Default 50 | for ninp in node_inputs: 51 | if ninp in _inputs_name and op_name in FORCE_CHANNEL_LAST_OP and layout_dict[ninp] == Layout.Default: 52 | layout_dict[ninp] = Layout.Channel_Last 53 | _inputs_name.remove(ninp) 54 | if ninp in _inputs_name and op_name in FORCE_CHANNEL_FIRST_OP and layout_dict[ninp] == Layout.Default: 55 | layout_dict[ninp] = Layout.Channel_First 56 | _inputs_name.remove(ninp) 57 | # output_layout = output_layout | node_dict[ninp] 58 | 59 | if len(_inputs_name) == 0: 60 | break 61 | 62 | input_nodes = {} 63 | for inp in model_graph.input: 64 | input_shape = [x.dim_value for x in inp.type.tensor_type.shape.dim] 65 | if input_shape == []: 66 | continue 67 | batch_size = 1 if input_shape[0] <= 0 else input_shape[0] 68 | input_shape = input_shape[1:] 69 | if layout_dict[inp.name] == Layout.Channel_Last: 70 | input_shape = input_shape[1:] + input_shape[0:1] 71 | 72 | input_nodes[inp.name] = keras.Input(shape=input_shape, batch_size=batch_size, dtype=onnx2tf_type.get(inp.type.tensor_type.elem_type)) 73 | 74 | return input_nodes 75 | -------------------------------------------------------------------------------- /test_TFLite_ISP.py: -------------------------------------------------------------------------------- 1 | import numpy as np 2 | from tensorflow import lite 3 | import os 4 | import cv2 5 | from PIL import Image 6 | 7 | def load_model(model_path): 8 | interpreter = lite.Interpreter(model_path=model_path) 9 | interpreter.allocate_tensors() 10 | return interpreter 11 | 12 | def print_model_structure(interpreter): 13 | print("Model Structure:") 14 | input_details = interpreter.get_input_details() 15 | output_details = interpreter.get_output_details() 16 | 17 | print("\nInput details:") 18 | for input_tensor in input_details: 19 | print(f"Name: {input_tensor['name']}, Shape: {input_tensor['shape']}, Type: {input_tensor['dtype']}") 20 | 21 | print("\nOutput details:") 22 | for output_tensor in output_details: 23 | print(f"Name: {output_tensor['name']}, Shape: {output_tensor['shape']}, Type: {output_tensor['dtype']}") 24 | 25 | def bayer2rggb(img_bayer): 26 | h, w = img_bayer.shape 27 | img_bayer = img_bayer.reshape(h // 2, 2, w // 2, 2) 28 | img_bayer = img_bayer.transpose([1, 3, 0, 2]).reshape([4, h // 2, w // 2]) # [4, h//2, w//2] 29 | return img_bayer 30 | 31 | def inference_and_save_results(interpreter, input_image_folder, output_image_folder): 32 | input_details = interpreter.get_input_details() 33 | output_details = interpreter.get_output_details() 34 | 35 | if not os.path.exists(output_image_folder): 36 | os.makedirs(output_image_folder) 37 | 38 | for image_name in os.listdir(input_image_folder): 39 | image_path = os.path.join(input_image_folder, image_name) 40 | if not image_path.lower().endswith(('png', 'jpg', 'jpeg')): 41 | continue 42 | 43 | img = Image.open(image_path) 44 | img_array = np.array(img) 45 | img_array = bayer2rggb(img_array) # Convert Bayer pattern to RGGB 46 | img_array = img_array.astype(np.float32) / 4095.0 # Normalize to [0, 1] 47 | 48 | # Convert to shape [1, 128, 128, 4] for the model input 49 | img_array = np.transpose(img_array, (1, 2, 0)) 50 | img_array = np.expand_dims(img_array, axis=0) 51 | 52 | interpreter.set_tensor(input_details[0]['index'], img_array) 53 | interpreter.invoke() 54 | output_img = interpreter.get_tensor(output_details[0]['index']) 55 | output_img = np.clip(output_img, 0., 1.) 56 | output_img = np.squeeze(output_img) 57 | print(output_img.shape) 58 | 59 | #output_img = output_img.transpose(1, 2, 0) 60 | output_img = (output_img * 255).astype(np.uint8) 61 | pil_img = Image.fromarray(output_img) 62 | 63 | # Save the output image using PIL (Image.save) 64 | output_image_path = os.path.join(output_image_folder, image_name) 65 | pil_img.save(output_image_path) # This automatically saves as RGB 66 | 67 | def main(): 68 | model_path = './ISP.tflite' # TFLite model path 69 | input_image_folder = './ISP/Input' # Input the Image folder path 70 | output_image_folder = './Output' # Output the Image folder path 71 | 72 | interpreter = load_model(model_path) 73 | print_model_structure(interpreter) 74 | inference_and_save_results(interpreter, input_image_folder, output_image_folder) 75 | 76 | if __name__ == "__main__": 77 | main() 78 | -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 |

[ICCV 2025] MobileIE: An Extremely Lightweight and Effective ConvNet for Real-Time Image Enhancement on Mobile Devices

2 | 3 |
4 |
5 | Hailong Yan1  6 | Ao Li1  7 | Xiangtao Zhang1  8 | Zhe Liu1  9 | Zenglin Shi2  10 | Ce Zhu1  11 | Le Zhang1,†  12 |
13 | 1 UESTC   2 Hefei University of Technology
14 | † Corresponding author.
15 | 16 |

17 | 📄 arXiv Paper   18 |

19 |
20 | 21 |
22 | Abstract: Recent advancements in deep neural networks have driven significant progress in image enhancement (IE). However, deploying deep learning models on resource-constrained platforms, such as mobile devices, remains challenging due to high computation and memory demands. To address these challenges and facilitate real-time IE on mobile, we introduce an extremely lightweight Convolutional Neural Network (CNN) framework with around 4K parameters. Our approach integrates reparameterization with an Incremental Weight Optimization strategy to ensure efficiency. Additionally, we enhance performance with a Feature Self-Transform module and a Hierarchical Dual-Path Attention mechanism, optimized with a Local Variance-Weighted loss. With this efficient framework, we are the first to achieve real-time IE inference at up to 1,100 frames per second (FPS) while delivering competitive image quality, achieving the best trade-off between speed and performance across multiple IE tasks. 23 |
24 | 25 |

26 | 27 |

28 | 29 | --- 30 | 31 | 32 | ### Preparation 33 | 34 | 1. Replace the dataset path in the config file. 35 | 2. If you want to train the model, set the type in config to "original" and need_slims to "false". 36 | 3. If you want to test the pretrain model, set the type in config to "re-parameterized", need_slims to "true", and load the re-parameterized pre-trained model. You can also run inference with TFLite model by executing "test_TFLite_RGB.py/test_TFLite_ISP.py". 37 | 4. You can use the TFLite model and import it into AI Benchmark (https://ai-benchmark.com/) to obtain the inference speed on mobile devices. 38 | 5. If you want to perform UIE task, replace the dataset path in config/lle.yaml with your underwater image dataset. 39 | 40 | ### Train 41 | 42 | ```bash 43 | python main.py -task train -model_task lle/isp -device cuda 44 | ``` 45 | 46 | ### Test 47 | 48 | ```bash 49 | python main.py -task test -model_task lle/isp -device cuda 50 | ``` 51 | 52 | ### Demo 53 | 54 | ```bash 55 | python main.py -task demo -model_task lle/isp -device cuda 56 | ``` 57 | 58 | ### Contact 59 | If you have any questions, please contact me by e-mail (yanhailong@std.uestc.edu.cn; yhl00825@163.com). 60 | 61 | ### Citation 62 | 63 | If you find the code helpful in your research or work, please cite the following paper: 64 | 65 | ``` 66 | @InProceedings{yan2025mobileie, 67 | author = {Yan, Hailong and Li, Ao and Zhang, Xiangtao and Liu, Zhe and Shi, Zenglin and Zhu, Ce and Zhang, Le}, 68 | title = {MobileIE: An Extremely Lightweight and Effective ConvNet for Real-Time Image Enhancement on Mobile Devices}, 69 | booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision}, 70 | month = {October}, 71 | year = {2025}, 72 | } 73 | ``` 74 | -------------------------------------------------------------------------------- /loss.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import torch.nn as nn 3 | from option import get_option 4 | 5 | class CharbonnierLoss(nn.Module): 6 | def __init__(self, eps=1e-6): 7 | super(CharbonnierLoss, self).__init__() 8 | self.eps2 = eps ** 2 9 | 10 | def forward(self, inp, target): 11 | return ((nn.functional.mse_loss(inp, target, reduction='none') + self.eps2) ** .5).mean() 12 | ##################################################################################################### 13 | class OutlierAwareLoss(nn.Module): 14 | def __init__(self,): 15 | super(OutlierAwareLoss, self).__init__() 16 | 17 | def forward(self, out, lab): 18 | delta = out - lab 19 | var = delta.std((2, 3), keepdims=True) / (2 ** .5) 20 | avg = delta.mean((2, 3), True) 21 | weight = torch.tanh((delta - avg).abs() / (var + 1e-6)).detach() 22 | loss = (delta.abs() * weight).mean() 23 | return loss 24 | 25 | ##################################################################################################### 26 | class LossWarmup(nn.Module): 27 | def __init__(self): 28 | super(LossWarmup, self).__init__() 29 | self.loss_cb = CharbonnierLoss(1e-8) 30 | self.loss_cs = nn.CosineSimilarity() 31 | 32 | def forward(self, inp, gt, warmup1, warmup2): 33 | loss = self.loss_cb(warmup2, inp) + \ 34 | (self.loss_cb(warmup1, gt) + (1 - self.loss_cs(warmup1.clip(0, 1), gt)).mean()) 35 | 36 | return loss 37 | 38 | 39 | class LossLLE(nn.Module): 40 | def __init__(self): 41 | super(LossLLE, self).__init__() 42 | self.loss_cs = nn.CosineSimilarity() 43 | self.loss_oa = OutlierAwareLoss() 44 | self.psnr = PSNRLoss() 45 | 46 | def forward(self, out, gt): 47 | loss = (self.loss_oa(out, gt) + (1 - self.loss_cs(out.clip(0, 1), gt)).mean()) + 2 * self.psnr(out, gt) 48 | return loss 49 | 50 | class LossISP(nn.Module): 51 | def __init__(self): 52 | super(LossISP, self).__init__() 53 | self.loss_cs = nn.CosineSimilarity() 54 | self.loss_oa = OutlierAwareLoss() 55 | self.psnr = PSNRLoss() 56 | 57 | def forward(self, out, gt): 58 | loss = (self.loss_oa(out, gt) + (1 - self.loss_cs(out.clip(0, 1), gt)).mean()) + 2 * self.psnr(out, gt) 59 | return loss 60 | 61 | def import_loss(training_task): 62 | if training_task == 'isp': 63 | return LossISP() 64 | elif training_task == 'lle': 65 | return LossLLE() 66 | elif training_task == 'warmup': 67 | return LossWarmup() 68 | else: 69 | raise ValueError('unknown training task, please choose from [isp, lle, warmup].') 70 | 71 | class PSNRLoss(nn.Module): 72 | 73 | def __init__(self, loss_weight=1.0, reduction='mean', toY=False): 74 | super(PSNRLoss, self).__init__() 75 | assert reduction == 'mean' 76 | self.loss_weight = loss_weight 77 | self.toY = toY 78 | self.coef = torch.tensor([65.481, 128.553, 24.966]).reshape(1, 3, 1, 1) 79 | self.first = True 80 | 81 | def forward(self, pred, target): 82 | assert len(pred.size()) == 4 83 | if self.toY: 84 | if self.first: 85 | self.coef = self.coef.to(pred.device) 86 | self.first = False 87 | 88 | pred = (pred * self.coef).sum(dim=1).unsqueeze(dim=1) + 16. 89 | target = (target * self.coef).sum(dim=1).unsqueeze(dim=1) + 16. 90 | 91 | pred, target = pred / 255., target / 255. 92 | pass 93 | assert len(pred.size()) == 4 94 | imdff=pred-target 95 | rmse=((imdff**2).mean(dim=(1,2,3))+1e-8).sqrt() 96 | loss=20*torch.log10(1/rmse).mean() 97 | loss=(50.0-loss)/100.0 98 | return loss 99 | -------------------------------------------------------------------------------- /onnx2tflite/onnx2tflite/components/output_check.py: -------------------------------------------------------------------------------- 1 | import os 2 | import numpy as np 3 | import tensorflow as tf 4 | import onnxruntime as ort 5 | from onnx2tflite.utils.definitions import Layout 6 | from onnx2tflite.utils.dimension_utils import tensor_NDC_to_NCD_format 7 | 8 | def tflite_run(model_path:str) -> np.ndarray: 9 | ''' 10 | tflite runtime 11 | ''' 12 | tflite_runtime = tf.lite.Interpreter(model_path, num_threads=4) 13 | tflite_runtime.allocate_tensors() 14 | input_details, output_details = tflite_runtime.get_input_details(), tflite_runtime.get_output_details() 15 | for i in range(len(input_details)): 16 | tflite_runtime.set_tensor(input_details[i]['index'], np.ones(input_details[i]['shape'], dtype=np.float32)) 17 | tflite_runtime.invoke() 18 | 19 | # only compare one output is ok. 20 | tflite_output = tflite_runtime.get_tensor(output_details[0]['index']) 21 | return tflite_output 22 | 23 | def keras_run(model_path:str) -> np.ndarray: 24 | ''' 25 | keras runtime 26 | ''' 27 | keras_runtime = tf.keras.models.load_model(model_path) 28 | _input = [] 29 | for inp in keras_runtime.inputs: 30 | _input.append(np.ones(list(inp.shape), dtype=np.float32)) 31 | 32 | keras_output = keras_runtime.predict(_input) 33 | # only compare one output is ok. 34 | if isinstance(keras_output, list): 35 | keras_output = keras_output[0] 36 | return keras_output 37 | 38 | 39 | def get_elements_error(onnx_proto, keras_model_path:str, tflite_model_path:str, input_layout:dict, output_layout:dict) -> dict: 40 | ''' 41 | use ones input arr to check model. 42 | more carefully check is up to youself custom code. 43 | ''' 44 | result = {} 45 | # test onnx 46 | onnx_runtime = ort.InferenceSession(onnx_proto.SerializeToString()) 47 | onnx_inputs = {} 48 | for inp in onnx_runtime.get_inputs(): 49 | shape = inp.shape 50 | if isinstance(shape[0], str) or shape[0] < 1: 51 | shape[0] = 1 52 | onnx_inputs[inp.name] = np.ones(shape, dtype=np.float32) 53 | if len(shape) > 2: 54 | _transpose_index = [i for i in range(len(shape))] 55 | _transpose_index = _transpose_index[0:1] + _transpose_index[2:] + _transpose_index[1:2] 56 | onnx_outputs = onnx_runtime.run([], onnx_inputs) 57 | 58 | channel_last = False 59 | for oup in onnx_proto.graph.output: 60 | channel_last = output_layout[oup.name] == Layout.Channel_Last 61 | break 62 | 63 | if keras_model_path is not None: 64 | # test keras model 65 | keras_output = keras_run(keras_model_path) 66 | if channel_last: 67 | keras_output = tensor_NDC_to_NCD_format(keras_output) 68 | # get max error 69 | keras_max_error = 1000 70 | for onnx_output in onnx_outputs: 71 | if onnx_output.shape != keras_output.shape: 72 | continue 73 | diff = np.abs(onnx_output - keras_output) 74 | max_diff = np.max(diff) 75 | keras_max_error = min(keras_max_error, max_diff) 76 | result['keras'] = keras_max_error 77 | 78 | if tflite_model_path is not None: 79 | # test tflite 80 | tflite_output = tflite_run(tflite_model_path) 81 | if channel_last: 82 | tflite_output = tensor_NDC_to_NCD_format(tflite_output) 83 | # get max error 84 | tflite_max_error = 1000 85 | for onnx_output in onnx_outputs: 86 | if onnx_output.shape != tflite_output.shape: 87 | continue 88 | diff = np.abs(onnx_output - tflite_output) 89 | max_diff = np.max(diff) 90 | tflite_max_error = min(tflite_max_error, max_diff) 91 | result['tflite'] = tflite_max_error 92 | 93 | return result -------------------------------------------------------------------------------- /model/lle.py: -------------------------------------------------------------------------------- 1 | import torch.nn as nn 2 | import torch 3 | from .utils import ( 4 | MBRConv5, 5 | MBRConv3, 6 | MBRConv1, 7 | DropBlock, 8 | FST, 9 | FSTS, 10 | ) 11 | 12 | class MobileIELLENet(nn.Module): 13 | def __init__(self, channels, rep_scale=4): 14 | super(MobileIELLENet, self).__init__() 15 | self.channels = channels 16 | self.head = FST( 17 | nn.Sequential( 18 | MBRConv5(3, channels, rep_scale=rep_scale), 19 | nn.PReLU(channels), 20 | MBRConv3(channels, channels, rep_scale=rep_scale) 21 | ), 22 | channels 23 | ) 24 | self.body = FST( 25 | MBRConv3(channels, channels, rep_scale=rep_scale), 26 | channels 27 | ) 28 | self.att = nn.Sequential( 29 | nn.AdaptiveAvgPool2d(1), 30 | MBRConv1(channels, channels, rep_scale=rep_scale), 31 | nn.Sigmoid() 32 | ) 33 | self.att1= nn.Sequential( 34 | MBRConv1(1, channels, rep_scale=rep_scale), 35 | nn.Sigmoid() 36 | ) 37 | self.tail = MBRConv3(channels, 3, rep_scale=rep_scale) 38 | self.tail_warm = MBRConv3(channels, 3, rep_scale=rep_scale) 39 | self.drop = DropBlock(3) 40 | 41 | def forward(self, x): 42 | x0 = self.head(x) 43 | x1 = self.body(x0) 44 | x2 = self.att(x1) 45 | max_out, _ = torch.max(x2 * x1 , dim=1, keepdim=True) 46 | x3 = self.att1(max_out) 47 | x4 = torch.mul(x2, x3) * x1 48 | return self.tail(x4) 49 | 50 | def forward_warm(self, x): 51 | x = self.drop(x) 52 | x = self.head(x) 53 | x = self.body(x) 54 | return self.tail(x), self.tail_warm(x) 55 | 56 | def slim(self): 57 | net_slim = MobileIELLENetS(self.channels) 58 | weight_slim = net_slim.state_dict() 59 | for name, mod in self.named_modules(): 60 | if isinstance(mod, MBRConv3) or isinstance(mod, MBRConv5) or isinstance(mod, MBRConv1): 61 | if '%s.weight' % name in weight_slim: 62 | w, b = mod.slim() 63 | weight_slim['%s.weight' % name] = w 64 | weight_slim['%s.bias' % name] = b 65 | elif isinstance(mod, FST): 66 | weight_slim['%s.bias' % name] = mod.bias 67 | weight_slim['%s.weight1' % name] = mod.weight1 68 | weight_slim['%s.weight2' % name] = mod.weight2 69 | elif isinstance(mod, nn.PReLU): 70 | weight_slim['%s.weight' % name] = mod.weight 71 | net_slim.load_state_dict(weight_slim) 72 | return net_slim 73 | 74 | class MobileIELLENetS(nn.Module): 75 | def __init__(self, channels): 76 | super(MobileIELLENetS, self).__init__() 77 | self.head = FSTS( 78 | nn.Sequential( 79 | nn.Conv2d(3, channels, 5, 1, 2), 80 | nn.PReLU(channels), 81 | nn.Conv2d(channels, channels, 3, 1, 1) 82 | ), 83 | channels 84 | ) 85 | self.body = FSTS( 86 | nn.Conv2d(channels, channels, 3, 1, 1), 87 | channels 88 | ) 89 | self.att = nn.Sequential( 90 | nn.AdaptiveAvgPool2d(1), 91 | nn.Conv2d(channels, channels, 1), 92 | nn.Sigmoid() 93 | ) 94 | self.att1 = nn.Sequential( 95 | nn.Conv2d(1, channels, 1, 1), 96 | nn.Sigmoid() 97 | ) 98 | self.tail = nn.Conv2d(channels, 3, 3, 1, 1) 99 | 100 | def forward(self, x): 101 | x0 = self.head(x) 102 | x1 = self.body(x0) 103 | x2 = self.att(x1) 104 | max_out, _ = torch.max(x2 * x1, dim=1, keepdim=True) 105 | x3 = self.att1(max_out) 106 | x4 = torch.mul(x3, x2) * x1 107 | return self.tail(x4) 108 | 109 | 110 | -------------------------------------------------------------------------------- /model/isp.py: -------------------------------------------------------------------------------- 1 | import torch.nn as nn 2 | import torch 3 | from .utils import ( 4 | MBRConv5, 5 | MBRConv3, 6 | MBRConv1, 7 | DropBlock, 8 | FST, 9 | FSTS, 10 | ) 11 | 12 | class MobileIEISPNet(nn.Module): 13 | def __init__(self, channels, rep_scale=4): 14 | super(MobileIEISPNet, self).__init__() 15 | self.channels = channels 16 | self.head = FST( 17 | nn.Sequential( 18 | MBRConv5(4, channels, rep_scale=rep_scale), 19 | nn.PReLU(channels), 20 | MBRConv3(channels, channels, rep_scale=rep_scale) 21 | ), 22 | channels 23 | ) 24 | self.body = FST( 25 | MBRConv3(channels, channels, rep_scale=rep_scale), 26 | channels 27 | ) 28 | self.att = nn.Sequential( 29 | nn.AdaptiveAvgPool2d(1), 30 | MBRConv1(channels, channels, rep_scale=rep_scale), 31 | nn.Sigmoid() 32 | ) 33 | self.att1= nn.Sequential( 34 | MBRConv1(1, channels, rep_scale=rep_scale), 35 | nn.Sigmoid() 36 | ) 37 | self.tail = nn.Sequential(nn.PixelShuffle(2), MBRConv3(3, 3, rep_scale=rep_scale)) 38 | self.tail_warm = MBRConv3(channels, 4, rep_scale=rep_scale) 39 | self.drop = DropBlock(3) 40 | 41 | def forward(self, x): 42 | x0 = self.head(x) 43 | x1 = self.body(x0) 44 | x2 = self.att(x1) 45 | max_out, _ = torch.max(x2 * x1, dim=1, keepdim=True) 46 | x3 = self.att1(max_out) 47 | x4 = torch.mul(x3, x2) * x1 48 | return self.tail(x4) 49 | 50 | def forward_warm(self, x): 51 | x = self.drop(x) 52 | x = self.head(x) 53 | x = self.body(x) 54 | return self.tail(x), self.tail_warm(x) 55 | 56 | def slim(self): 57 | net_slim = MobileIEISPNetS(self.channels) 58 | weight_slim = net_slim.state_dict() 59 | for name, mod in self.named_modules(): 60 | if isinstance(mod, MBRConv3) or isinstance(mod, MBRConv5) or isinstance(mod, MBRConv1): 61 | if '%s.weight' % name in weight_slim: 62 | w, b = mod.slim() 63 | weight_slim['%s.weight' % name] = w 64 | weight_slim['%s.bias' % name] = b 65 | elif isinstance(mod, FST): 66 | weight_slim['%s.bias' % name] = mod.bias 67 | weight_slim['%s.weight1' % name] = mod.weight1 68 | weight_slim['%s.weight2' % name] = mod.weight2 69 | elif isinstance(mod, nn.PReLU): 70 | weight_slim['%s.weight' % name] = mod.weight 71 | net_slim.load_state_dict(weight_slim) 72 | return net_slim 73 | 74 | class MobileIEISPNetS(nn.Module): 75 | def __init__(self, channels): 76 | super(MobileIEISPNetS, self).__init__() 77 | self.head = FSTS( 78 | nn.Sequential( 79 | nn.Conv2d(4, channels, 5, 1, 2), 80 | nn.PReLU(channels), 81 | nn.Conv2d(channels, channels, 3, 1, 1) 82 | ), 83 | channels 84 | ) 85 | self.body = FSTS( 86 | nn.Conv2d(channels, channels, 3, 1, 1), 87 | channels 88 | ) 89 | self.att = nn.Sequential( 90 | nn.AdaptiveAvgPool2d(1), 91 | nn.Conv2d(channels, channels, 1), 92 | nn.Sigmoid() 93 | ) 94 | self.att1 = nn.Sequential( 95 | nn.Conv2d(1, channels, 1, 1), 96 | nn.Sigmoid() 97 | ) 98 | self.tail = nn.Sequential(nn.PixelShuffle(2), nn.Conv2d(3, 3, 3, 1, 1)) 99 | 100 | def forward(self, x): 101 | x0 = self.head(x) 102 | x1 = self.body(x0) 103 | x2 = self.att(x1) 104 | max_out, _ = torch.max(x2 * x1, dim=1, keepdim=True) 105 | x3 = self.att1(max_out) 106 | x4 = torch.mul(x3, x2) * x1 107 | return self.tail(x4) 108 | -------------------------------------------------------------------------------- /onnx2tflite/onnx2tflite/components/builder.py: -------------------------------------------------------------------------------- 1 | import os 2 | os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2' 3 | 4 | import tensorflow as tf 5 | from tensorflow import keras 6 | from onnx import numpy_helper 7 | from .dataloader import RandomLoader, ImageLoader 8 | 9 | from onnx2tflite.utils import OPERATOR 10 | from onnx2tflite.layers import conv_layers 11 | from onnx2tflite.utils.definitions import * 12 | from onnx2tflite.utils.graph_tools import build_tf_inputs, decode_node_attribute 13 | 14 | def keras_builder(onnx_model, native_groupconv:bool=False): 15 | 16 | conv_layers.USE_NATIVE_GROUP_CONV = native_groupconv 17 | 18 | model_graph = onnx_model.graph 19 | layout_dict, tf_tensor = {}, {} 20 | 21 | ''' 22 | init onnx model's build-in tensors 23 | ''' 24 | onnx_weights = dict() 25 | for initializer in model_graph.initializer: 26 | onnx_weights[initializer.name] = numpy_helper.to_array(initializer) 27 | 28 | ''' 29 | build input nodes 30 | ''' 31 | input_nodes = build_tf_inputs(model_graph, layout_dict) 32 | tf_tensor.update(input_nodes) 33 | 34 | ''' 35 | build model inline node by iterate onnx nodes. 36 | ''' 37 | for node in model_graph.node: 38 | op_name, node_inputs, node_outputs = node.op_type, node.input, node.output 39 | op_attr = decode_node_attribute(node) 40 | 41 | tf_operator = OPERATOR.get(op_name) 42 | if tf_operator is None: 43 | raise KeyError(f"{op_name} not implemented yet") 44 | 45 | _inputs = None 46 | if len(node_inputs) > 0: 47 | _inputs = tf_tensor[node_inputs[0]] if node_inputs[0] in tf_tensor else onnx_weights[node_inputs[0]] 48 | 49 | # init layout 50 | for index in range(len(node_outputs)): 51 | layout_dict[node_outputs[index]] = layout_dict.get(node_inputs[0], Layout.Default) 52 | 53 | res = tf_operator(tf_tensor, onnx_weights, node_inputs, op_attr, node_outputs, layout_dict)(_inputs) 54 | if isinstance(res, list): 55 | for index in range(len(node_outputs)): 56 | tf_tensor[node_outputs[index]] = res[index] 57 | else: 58 | tf_tensor[node_outputs[0]] = res 59 | 60 | ''' 61 | build keras model 62 | ''' 63 | input_nodes = [tf_tensor[x.name] for x in model_graph.input] 64 | outputs_nodes = [tf_tensor[x.name] for x in model_graph.output] 65 | keras_model = keras.Model(inputs=input_nodes, outputs=outputs_nodes) 66 | keras_model.trainable = False 67 | # keras_model.summary() 68 | # print(layout_dict) 69 | input_layout, output_layout = {}, {} 70 | for inp in model_graph.input: 71 | input_layout[inp.name] = layout_dict[inp.name] 72 | for oup in model_graph.output: 73 | output_layout[oup.name] = layout_dict[oup.name] 74 | return keras_model, input_layout, output_layout 75 | 76 | def tflite_builder(keras_model, weight_quant:bool=False, fp16_model=False, int8_model:bool=False, image_root:str=None, 77 | int8_mean:list or float = [123.675, 116.28, 103.53], int8_std:list or float = [58.395, 57.12, 57.375]): 78 | converter = tf.lite.TFLiteConverter.from_keras_model(keras_model) 79 | converter.target_spec.supported_ops = [tf.lite.OpsSet.TFLITE_BUILTINS, tf.lite.OpsSet.SELECT_TF_OPS] 80 | if weight_quant or int8_model or fp16_model: 81 | converter.experimental_new_converter = True 82 | converter.optimizations = [tf.lite.Optimize.DEFAULT] 83 | 84 | if fp16_model: 85 | converter.target_spec.supported_types = [tf.float16] 86 | converter.inference_input_type = tf.float32 87 | converter.inference_output_type = tf.float32 88 | elif int8_model: 89 | assert len(keras_model.inputs) == 1, f"help want, only support single input model." 90 | shape = list(keras_model.inputs[0].shape) 91 | dataset = RandomLoader(shape) if image_root is None else ImageLoader(image_root, shape, int8_mean, int8_std) 92 | converter.representative_dataset = lambda: dataset 93 | converter.target_spec.supported_ops = [tf.lite.OpsSet.TFLITE_BUILTINS_INT8, tf.lite.OpsSet.SELECT_TF_OPS] 94 | converter.target_spec.supported_types = [] 95 | converter.inference_input_type = tf.uint8 96 | converter.inference_output_type = tf.uint8 97 | converter.experimental_new_converter = True 98 | 99 | tflite_model = converter.convert() 100 | return tflite_model -------------------------------------------------------------------------------- /onnx2tflite/onnx2tflite/components/onnx_loader.py: -------------------------------------------------------------------------------- 1 | import os 2 | import onnx 3 | import logging 4 | from onnxsim import simplify # ONNX 模型简化工具 5 | 6 | LOG = logging.getLogger("onnx_loader running:") 7 | LOG.setLevel(logging.INFO) 8 | 9 | def clean_model_input(model_proto): 10 | """ 11 | 清理 ONNX 模型的输入，删除 ONNX 计算图中冗余的输入节点。 12 | 13 | 逻辑： 14 | - 遍历 ONNX 计算图中的 `graph.input` 15 | - 如果某个 `input` 也出现在 `initializer` 中，则说明它是一个冗余输入（即它的值已经在 `initializer` 中存储） 16 | - 从 `graph.input` 中移除这些冗余输入 17 | 18 | 参数： 19 | - model_proto (onnx.ModelProto): 需要清理的 ONNX 模型 20 | """ 21 | inputs = model_proto.graph.input # 获取 ONNX 计算图中的输入 22 | name_to_input = {} # 创建输入名称到输入对象的映射 23 | for input in inputs: 24 | name_to_input[input.name] = input 25 | 26 | names = [] 27 | for initializer in model_proto.graph.initializer: # 遍历所有初始化参数 28 | if initializer.name in name_to_input: # 如果初始化参数的名字在输入列表中 29 | inputs.remove(name_to_input[initializer.name]) # 删除该输入 30 | names.append(initializer.name) 31 | 32 | if len(names) > 0: 33 | LOG.warning(f"[{len(names)}] redundant input nodes are removed.\n \ 34 | nodes name : {','.join(names)}") 35 | 36 | def get_onnx_submodel(onnx_model_path:str, input_node_names:list=None, output_node_names:list=None): 37 | """ 38 | 截取 ONNX 子模型，即从原始 ONNX 模型中提取以 `input_node_names` 为输入、 39 | `output_node_names` 为输出的子图。 40 | 41 | 逻辑： 42 | - 载入 ONNX 模型 43 | - 确定输入节点和输出节点（如果未指定，则默认使用整个模型的输入/输出） 44 | - 使用 `onnx.utils.extract_model` 提取子模型并保存 45 | - 载入提取后的子模型并返回 46 | 47 | 参数： 48 | - onnx_model_path (str): ONNX 模型文件路径 49 | - input_node_names (list, optional): 指定子模型的输入节点名称 50 | - output_node_names (list, optional): 指定子模型的输出节点名称 51 | 52 | 返回： 53 | - model_proto (onnx.ModelProto): 提取的子模型 54 | """ 55 | model_proto = onnx.load(onnx_model_path) # 载入 ONNX 模型 56 | # 如果未指定输入节点，则默认使用 ONNX 模型的全部输入 57 | if input_node_names is None: 58 | input_node_names = [] 59 | for inp in model_proto.graph.input: 60 | input_node_names.append(inp.name) 61 | 62 | # 如果未指定输出节点，则默认使用 ONNX 模型的全部输出 63 | if output_node_names is None: 64 | output_node_names = [] 65 | for oup in model_proto.graph.output: 66 | output_node_names.append(oup.name) 67 | del model_proto # 释放原始模型的内存 68 | 69 | # 生成新模型的文件路径 70 | new_model_path = os.path.splitext(onnx_model_path)[0] + "_sub.onnx" 71 | # 提取子模型并保存 72 | onnx.utils.extract_model(onnx_model_path, new_model_path, input_node_names, output_node_names) 73 | # 载入提取后的子模型 74 | model_proto = onnx.load(new_model_path) 75 | return model_proto 76 | 77 | def get_proto(onnx_model_path:str, input_node_names:list=None, output_node_names:list=None): 78 | if input_node_names is None and output_node_names is None: 79 | return onnx.load(onnx_model_path) 80 | else: 81 | return get_onnx_submodel(onnx_model_path, input_node_names, output_node_names) 82 | 83 | def load_onnx_modelproto(onnx_model_path:str, input_node_names:list=None, output_node_names:list=None, need_simplify:bool=True): 84 | """ 85 | 载入 ONNX 模型，并根据需要进行简化和清理。 86 | 87 | 逻辑： 88 | - 检查 ONNX 模型文件是否存在 89 | - 载入完整的 ONNX 模型或子模型 90 | - 检测是否存在动态输入 91 | - 如果 `need_simplify=True`，则尝试使用 `onnx-simplifier` 进行模型优化 92 | - 移除 ONNX 计算图中冗余的输入 93 | 94 | 参数： 95 | - onnx_model_path (str): ONNX 模型文件路径 96 | - input_node_names (list, optional): 需要提取的输入节点 97 | - output_node_names (list, optional): 需要提取的输出节点 98 | - need_simplify (bool, optional): 是否对 ONNX 进行简化，默认启用 99 | 100 | 返回： 101 | - model_proto (onnx.ModelProto): 处理后的 ONNX 模型 102 | """ 103 | # 1. 检查 ONNX 文件是否存在 104 | if not os.path.exists(onnx_model_path): 105 | LOG.error(f"{onnx_model_path} is not exists.") 106 | raise FileExistsError(f"{onnx_model_path} is not exists.") 107 | # 2. 载入 ONNX 模型或子模型 108 | model_proto = get_proto(onnx_model_path, input_node_names, output_node_names) 109 | # 3. 检查是否存在动态输入（即输入形状中有未指定的维度） 110 | dynamic_input = False 111 | for inp in model_proto.graph.input: 112 | for x in inp.type.tensor_type.shape.dim: 113 | if x.dim_value <= 0: # 发现动态输入 114 | dynamic_input = True 115 | break 116 | # 4. 进行 ONNX 模型简化（如果启用） 117 | if need_simplify: 118 | success = False 119 | try: 120 | # 使用 `onnxsim` 进行模型优化，允许动态输入 121 | model_proto, success = simplify(model_proto, check_n=1, dynamic_input_shape=dynamic_input) 122 | except: 123 | success = False 124 | # 如果简化失败，记录警告信息 125 | if not success: 126 | LOG.warning(f"onnxsim is failed, maybe make convert fails.") 127 | 128 | model_proto = onnx.load(onnx_model_path) 129 | 130 | # 5. 清理 ONNX 模型的冗余输入 131 | clean_model_input(model_proto) 132 | # 在返回 ONNX 之前，检查是否仍然存在动态输入 133 | 134 | ################################################################################## 135 | for inp in model_proto.graph.input: 136 | for x in inp.type.tensor_type.shape.dim: 137 | if x.dim_value <= 0: # 仍然是动态输入 138 | LOG.warning(f"ONNX 仍然包含动态输入: {inp.name}，维度: {[dim.dim_value for dim in inp.type.tensor_type.shape.dim]}") 139 | break # 只打印一次警告即可 140 | ################################################################################## 141 | 142 | return model_proto -------------------------------------------------------------------------------- /onnx2tflite/build/lib/onnx2tflite/converter.py: -------------------------------------------------------------------------------- 1 | import os 2 | import logging 3 | from .components import load_onnx_modelproto, keras_builder, tflite_builder, get_elements_error 4 | 5 | logging.basicConfig(level=logging.INFO) 6 | LOG = logging.getLogger("converter running:") 7 | 8 | def onnx_converter(onnx_model_path:str, output_path:str=None, 9 | input_node_names:list=None, output_node_names:list=None, 10 | need_simplify:bool=True, target_formats:list = ['keras', 'tflite'], 11 | native_groupconv:bool=False, 12 | weight_quant:bool=False, fp16_model:bool=False, int8_model:bool=False, image_root:str=None, 13 | int8_mean:list or float = [123.675, 116.28, 103.53], int8_std:list or float = [58.395, 57.12, 57.375])->float: 14 | """ 15 | Converts an ONNX model to various target formats with optional optimizations. 16 | 17 | Parameters: 18 | onnx_model_path (str): Path to the input ONNX model file. 19 | output_path (str, optional): Path to save the converted model(s). If None, the converted model(s) will be saved in the same directory as the input model. 20 | input_node_names (list, optional): List of input node names. If None, the default input nodes of the ONNX model are used. 21 | output_node_names (list, optional): List of output node names. If None, the default output nodes of the ONNX model are used. 22 | need_simplify (bool, optional): If True, the ONNX model will be simplified before conversion. Default is True. 23 | target_formats (list, optional): List of target formats to convert the ONNX model to. Default is ['keras', 'tflite']. 24 | native_groupconv (bool, optional): If True, retains native group convolution operations during conversion. Default is False. 25 | weight_quant (bool, optional): If True, applies weight quantization to the converted model. Default is False. 26 | fp16_model (bool, optional): If True, converts the model to use FP16 precision. Default is False. 27 | int8_model (bool, optional): If True, converts the model to use INT8 precision. Default is False. 28 | image_root (str, optional): Path to the root directory of images for calibration if INT8 quantization is enabled. Default is None. 29 | int8_mean (list or float, optional): Mean values for INT8 quantization. Default is [123.675, 116.28, 103.53]. 30 | int8_std (list or float, optional): Standard deviation values for INT8 quantization. Default is [58.395, 57.12, 57.375]. 31 | 32 | Returns: 33 | float: Error value. 34 | 35 | Note: 36 | - The function supports multiple target formats for conversion and allows for various optimizations such as simplification, quantization, and precision reduction. 37 | - When INT8 quantization is enabled, 'image_root', 'int8_mean', and 'int8_std' parameters are used for calibration. 38 | """ 39 | if not isinstance(target_formats, list) and 'keras' not in target_formats and 'tflite' not in target_formats: 40 | raise KeyError("'keras' or 'tflite' should in list") 41 | 42 | model_proto = load_onnx_modelproto(onnx_model_path, input_node_names, output_node_names, need_simplify) 43 | 44 | keras_model, input_layout, output_layout = keras_builder(model_proto, native_groupconv) 45 | 46 | if 'tflite' in target_formats: 47 | tflite_model = tflite_builder(keras_model, weight_quant, fp16_model, int8_model, image_root, int8_mean, int8_std) 48 | 49 | onnx_path, model_name = os.path.split(onnx_model_path) 50 | if output_path is None: 51 | output_path = onnx_path 52 | output_path = os.path.join(output_path, model_name.split('.')[0]) 53 | 54 | if fp16_model: 55 | output_path = output_path + "_fp16" 56 | elif int8_model: 57 | output_path = output_path + "_int8" 58 | 59 | keras_model_path = None 60 | if 'keras' in target_formats: 61 | keras_model_path = output_path + ".h5" 62 | keras_model.save(keras_model_path) 63 | LOG.info(f"keras model saved in {keras_model_path}") 64 | 65 | tflite_model_path = None 66 | if 'tflite' in target_formats: 67 | tflite_model_path = output_path + ".tflite" 68 | with open(tflite_model_path, "wb") as fp: 69 | fp.write(tflite_model) 70 | 71 | convert_result = {"keras":keras_model_path, "tflite":tflite_model_path, "keras_error":0, "tflite_error":0} 72 | # ignore quantization model 73 | if int8_model: 74 | return convert_result 75 | 76 | error_dict = {} 77 | try: 78 | error_dict = get_elements_error(model_proto, keras_model_path, tflite_model_path, input_layout, output_layout) 79 | keras_error, tflite_error = error_dict.get("keras", None), error_dict.get("tflite", None) 80 | if keras_error: 81 | if keras_error > 1e-2: 82 | LOG.error("h5 model elements' max error has reached {:^.4E}, but convert is done, please check {} carefully!".format(keras_error, keras_model_path)) 83 | elif keras_error > 1e-4: 84 | LOG.warning("h5 model elements' max error is {:^.4E}, pass, h5 saved in {}".format(keras_error, keras_model_path)) 85 | else: 86 | LOG.info("h5 model elements' max error is {:^.4E}, pass, h5 saved in {}".format(keras_error, keras_model_path)) 87 | if tflite_error: 88 | if tflite_error > 1e-2: 89 | LOG.error("tflite model elements' max error has reached {:^.4E}, but convert is done, please check {} carefully!".format(tflite_error, tflite_model_path)) 90 | elif tflite_error > 1e-4: 91 | LOG.warning("tflite model elements' max error is {:^.4E}, pass, tflite saved in {}".format(tflite_error, tflite_model_path)) 92 | else: 93 | LOG.info("tflite model elements' max error is {:^.4E}, pass, tflite saved in {}".format(tflite_error, tflite_model_path)) 94 | except: 95 | LOG.warning("convert is successed, but model running is failed, please check carefully!") 96 | 97 | convert_result["keras_error"] = error_dict.get("keras", None) 98 | convert_result["tflite_error"] = error_dict.get("tflite", None) 99 | return convert_result -------------------------------------------------------------------------------- /main.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import numpy as np 3 | import cv2 4 | from tqdm import tqdm 5 | from logger import Logger 6 | from option import get_option 7 | from data import import_loader 8 | from loss import import_loss 9 | from model import import_model 10 | import multiprocessing as mp 11 | import os 12 | os.environ['CUDA_VISIBLE_DEVICES'] = '5' 13 | 14 | def count_parameters(model): 15 | return sum(p.numel() for p in model.parameters() if p.requires_grad) 16 | 17 | def train(opt, logger): 18 | logger.info('task: {}, model task: {}'.format(opt.task, opt.model_task)) 19 | 20 | train_loader, valid_loader = import_loader(opt) 21 | lr = float(opt.config['train']['lr']) 22 | lr_warmup = float(opt.config['train']['lr_warmup']) 23 | 24 | loss_warmup = import_loss('warmup') 25 | loss_training = import_loss(opt.model_task) 26 | net = import_model(opt) 27 | # logger.info(net) 28 | num_params = count_parameters(net) 29 | print("Total number of parameters: ", num_params) 30 | 31 | net.train() 32 | # Phase Warming-up 33 | if opt.config['train']['warmup']: 34 | logger.info('start warming-up') 35 | 36 | optim_warm = torch.optim.Adam(net.parameters(), lr_warmup, weight_decay=0) 37 | epochs = opt.config['train']['warmup_epoch'] 38 | for epo in range(epochs): 39 | loss_li = [] 40 | for img_inp, img_gt, _ in tqdm(train_loader, ncols=80): 41 | optim_warm.zero_grad() 42 | warmup_out1, warmup_out2 = net.forward_warm(img_inp) 43 | loss = loss_warmup(img_inp, img_gt, warmup_out1, warmup_out2) 44 | loss.backward() 45 | optim_warm.step() 46 | loss_li.append(loss.item()) 47 | 48 | logger.info('epoch: {}, train_loss: {}'.format(epo+1, sum(loss_li)/len(loss_li))) 49 | torch.save(net.state_dict(), r'{}/model_pre.pkl'.format(opt.save_model_dir)) 50 | logger.info('warming-up phase done') 51 | 52 | # Phase Training 53 | best_psnr = 0 54 | epochs = int(opt.config['train']['epoch']) 55 | optim = torch.optim.Adam(net.parameters(), lr, weight_decay=0) 56 | lr_sch = torch.optim.lr_scheduler.CosineAnnealingWarmRestarts(optim, 50, 2, 1e-7) 57 | 58 | logger.info('start training') 59 | for epo in range(epochs): 60 | loss_li = [] 61 | test_psnr = [] 62 | net.train() 63 | for img_inp, img_gt, _ in tqdm(train_loader, ncols=80): 64 | out = net(img_inp) 65 | loss = loss_training(out, img_gt) 66 | optim.zero_grad() 67 | loss.backward() 68 | optim.step() 69 | loss_li.append(loss.item()) 70 | lr_sch.step() 71 | 72 | # Validation 73 | net.eval() 74 | for img_inp, img_gt, _ in tqdm(valid_loader, ncols=80): 75 | with torch.no_grad(): 76 | out = net(img_inp) 77 | mse = ((out - img_gt)**2).mean((2, 3)) 78 | psnr = (1 / mse).log10().mean() * 10 79 | test_psnr.append(psnr.item()) 80 | mean_psnr = sum(test_psnr)/len(test_psnr) 81 | 82 | if (epo+1) % int(opt.config['train']['save_every']) == 0: 83 | torch.save(net.state_dict(), r'{}/model_{}.pkl'.format(opt.save_model_dir, epo+1)) 84 | 85 | logger.info('epoch: {}, training loss: {}, validation psnr: {}'.format( 86 | epo+1, sum(loss_li) / len(loss_li), sum(test_psnr) / len(test_psnr) 87 | )) 88 | 89 | if mean_psnr > best_psnr: 90 | best_psnr = mean_psnr 91 | torch.save(net.state_dict(), r'{}/model_best.pkl'.format(opt.save_model_dir)) 92 | if opt.config['train']['save_slim']: 93 | net_slim = net.slim().to(opt.device) 94 | torch.save(net_slim.state_dict(), r'{}/model_best_slim.pkl'.format(opt.save_model_dir)) 95 | logger.info('best model saved and re-parameterized in epoch {}'.format(epo+1)) 96 | else: 97 | logger.info('best model saved in epoch in epoch {}'.format(epo+1)) 98 | 99 | logger.info('training done') 100 | 101 | 102 | def test(opt, logger): 103 | test_loader = import_loader(opt) 104 | net = import_model(opt) 105 | net.eval() 106 | psnr_list = [] 107 | logger.info('start testing') 108 | for (img_inp, img_gt, img_name) in test_loader: 109 | 110 | with torch.no_grad(): 111 | out = net(img_inp) 112 | mse = ((out - img_gt)**2).mean((2, 3)) 113 | psnr = (1 / mse).log10().mean() * 10 114 | 115 | if opt.config['test']['save']: 116 | out_img = (out.clip(0, 1)[0] * 255).permute([1, 2, 0]).cpu().numpy().astype(np.uint8)[..., ::-1] 117 | cv2.imwrite(r'{}/{}.png'.format(opt.save_image_dir, img_name[0]), out_img) 118 | 119 | psnr_list.append(psnr.item()) 120 | logger.info('image name: {}, test psnr: {}'.format(img_name[0], psnr)) 121 | 122 | logger.info('testing done, overall psnr: {}'.format(sum(psnr_list) / len(psnr_list))) 123 | 124 | 125 | def demo(opt, logger): 126 | demo_loader = import_loader(opt) 127 | net = import_model(opt) 128 | net.eval() 129 | logger.info('start demonstration') 130 | for img_inp, img_name in demo_loader: 131 | 132 | with torch.no_grad(): 133 | out = net(img_inp) 134 | out_img = (out.clip(0, 1)[0] * 255).permute([1, 2, 0]).cpu().numpy().astype(np.uint8)[..., ::-1] 135 | cv2.imwrite(r'{}/{}.png'.format(opt.save_image_dir, img_name[0]), out_img) 136 | logger.info('image name: {} output generated'.format(img_name[0])) 137 | logger.info('demonstration done') 138 | 139 | 140 | if __name__ == "__main__": 141 | mp.set_start_method('spawn') 142 | 143 | opt = get_option() 144 | logger = Logger(opt) 145 | 146 | if opt.task == 'train': 147 | train(opt, logger) 148 | elif opt.task == 'test': 149 | test(opt, logger) 150 | elif opt.task == 'demo': 151 | demo(opt, logger) 152 | else: 153 | raise ValueError('unknown task, please choose from [train, test, demo].') 154 | -------------------------------------------------------------------------------- /onnx2tflite/onnx2tflite/converter.py: -------------------------------------------------------------------------------- 1 | import os 2 | import logging 3 | from .components import load_onnx_modelproto, keras_builder, tflite_builder, get_elements_error 4 | 5 | logging.basicConfig(level=logging.INFO) 6 | LOG = logging.getLogger("converter running:") 7 | 8 | def onnx_converter(onnx_model_path:str, output_path:str=None, 9 | input_node_names:list=None, output_node_names:list=None, 10 | need_simplify:bool=True, target_formats:list = ['keras', 'tflite'], 11 | native_groupconv:bool=False, 12 | weight_quant:bool=False, fp16_model:bool=False, int8_model:bool=False, image_root:str=None, 13 | int8_mean:list or float = [123.675, 116.28, 103.53], int8_std:list or float = [58.395, 57.12, 57.375])->float: 14 | """ 15 | Converts an ONNX model to various target formats with optional optimizations. 16 | 17 | Parameters: 18 | onnx_model_path (str): Path to the input ONNX model file. 19 | output_path (str, optional): Path to save the converted model(s). If None, the converted model(s) will be saved in the same directory as the input model. 20 | input_node_names (list, optional): List of input node names. If None, the default input nodes of the ONNX model are used. 21 | output_node_names (list, optional): List of output node names. If None, the default output nodes of the ONNX model are used. 22 | need_simplify (bool, optional): If True, the ONNX model will be simplified before conversion. Default is True. 23 | target_formats (list, optional): List of target formats to convert the ONNX model to. Default is ['keras', 'tflite']. 24 | native_groupconv (bool, optional): If True, retains native group convolution operations during conversion. Default is False. 25 | weight_quant (bool, optional): If True, applies weight quantization to the converted model. Default is False. 26 | fp16_model (bool, optional): If True, converts the model to use FP16 precision. Default is False. 27 | int8_model (bool, optional): If True, converts the model to use INT8 precision. Default is False. 28 | image_root (str, optional): Path to the root directory of images for calibration if INT8 quantization is enabled. Default is None. 29 | int8_mean (list or float, optional): Mean values for INT8 quantization. Default is [123.675, 116.28, 103.53]. 30 | int8_std (list or float, optional): Standard deviation values for INT8 quantization. Default is [58.395, 57.12, 57.375]. 31 | 32 | Returns: 33 | float: Error value. 34 | 35 | Note: 36 | - The function supports multiple target formats for conversion and allows for various optimizations such as simplification, quantization, and precision reduction. 37 | - When INT8 quantization is enabled, 'image_root', 'int8_mean', and 'int8_std' parameters are used for calibration. 38 | """ 39 | # 确保 target_formats 是列表类型，并包含 'keras' 或 'tflite' 40 | if not isinstance(target_formats, list) and 'keras' not in target_formats and 'tflite' not in target_formats: 41 | raise KeyError("'keras' or 'tflite' should in list") 42 | 43 | # 1. 加载 ONNX 模型并进行解析 44 | model_proto = load_onnx_modelproto(onnx_model_path, input_node_names, output_node_names, need_simplify) 45 | 46 | # 2. 将 ONNX 模型转换为 Keras 模型 47 | keras_model, input_layout, output_layout = keras_builder(model_proto, native_groupconv) 48 | 49 | # 3. 如果目标格式包括 'tflite'，则进一步转换为 TFLite 50 | if 'tflite' in target_formats: 51 | tflite_model = tflite_builder(keras_model, weight_quant, fp16_model, int8_model, image_root, int8_mean, int8_std) 52 | 53 | # 4. 处理输出路径 54 | onnx_path, model_name = os.path.split(onnx_model_path) # 获取 ONNX 模型所在目录 55 | if output_path is None: 56 | output_path = onnx_path # 若未提供 output_path，则默认保存在 ONNX 所在目录 57 | output_path = os.path.join(output_path, model_name.split('.')[0]) # 设置输出文件的基本路径 58 | 59 | # 5. 如果是 FP16 或 INT8 量化模型，修改输出路径 60 | if fp16_model: 61 | output_path = output_path + "_fp16" 62 | elif int8_model: 63 | output_path = output_path + "_int8" 64 | 65 | # 6. 处理 Keras 模型的保存 66 | keras_model_path = None 67 | if 'keras' in target_formats: 68 | keras_model_path = output_path + ".h5" 69 | keras_model.save(keras_model_path) 70 | LOG.info(f"keras model saved in {keras_model_path}") 71 | 72 | # 7. 处理 TFLite 模型的保存 73 | tflite_model_path = None 74 | if 'tflite' in target_formats: 75 | tflite_model_path = output_path + ".tflite" 76 | with open(tflite_model_path, "wb") as fp: 77 | fp.write(tflite_model) 78 | 79 | # 8. 记录转换结果 80 | convert_result = {"keras":keras_model_path, "tflite":tflite_model_path, "keras_error":0, "tflite_error":0} 81 | # ignore quantization model 82 | if int8_model: 83 | return convert_result 84 | 85 | error_dict = {} 86 | try: 87 | error_dict = get_elements_error(model_proto, keras_model_path, tflite_model_path, input_layout, output_layout) 88 | keras_error, tflite_error = error_dict.get("keras", None), error_dict.get("tflite", None) 89 | if keras_error: 90 | if keras_error > 1e-2: 91 | LOG.error("h5 model elements' max error has reached {:^.4E}, but convert is done, please check {} carefully!".format(keras_error, keras_model_path)) 92 | elif keras_error > 1e-4: 93 | LOG.warning("h5 model elements' max error is {:^.4E}, pass, h5 saved in {}".format(keras_error, keras_model_path)) 94 | else: 95 | LOG.info("h5 model elements' max error is {:^.4E}, pass, h5 saved in {}".format(keras_error, keras_model_path)) 96 | if tflite_error: 97 | if tflite_error > 1e-2: 98 | LOG.error("tflite model elements' max error has reached {:^.4E}, but convert is done, please check {} carefully!".format(tflite_error, tflite_model_path)) 99 | elif tflite_error > 1e-4: 100 | LOG.warning("tflite model elements' max error is {:^.4E}, pass, tflite saved in {}".format(tflite_error, tflite_model_path)) 101 | else: 102 | LOG.info("tflite model elements' max error is {:^.4E}, pass, tflite saved in {}".format(tflite_error, tflite_model_path)) 103 | except: 104 | LOG.warning("convert is successed, but model running is failed, please check carefully!") 105 | 106 | convert_result["keras_error"] = error_dict.get("keras", None) 107 | convert_result["tflite_error"] = error_dict.get("tflite", None) 108 | return convert_result -------------------------------------------------------------------------------- /onnx2tflite/onnx2tflite/layers/activations_layers.py: -------------------------------------------------------------------------------- 1 | import numpy as np 2 | import tensorflow as tf 3 | from tensorflow import keras 4 | 5 | from onnx2tflite.utils.definitions import Layout 6 | from onnx2tflite.utils import OPERATOR, channel_to_last_dimension, tensor_NCD_to_NDC_format 7 | 8 | @OPERATOR.register_operator("Relu") 9 | class TFRelu(): 10 | def __init__(self, *args, **kwargs) -> None: 11 | super().__init__() 12 | 13 | def __call__(self, inputs): 14 | return keras.activations.relu(inputs) 15 | 16 | @OPERATOR.register_operator("HardSigmoid") 17 | class TFHardSigmoid(): 18 | def __init__(self, tensor_grap, node_weights, node_inputs, node_attribute, *args, **kwargs) -> None: 19 | super().__init__() 20 | self.alpha = node_attribute.get("alpha", 0.2) 21 | self.beta = node_attribute.get("beta", 0.5) 22 | 23 | def __call__(self, inputs): 24 | return tf.clip_by_value(self.alpha*inputs+self.beta, 0, 1) 25 | 26 | @OPERATOR.register_operator("HardSwish") 27 | class TFHardSwish(): 28 | def __init__(self, *args, **kwargs) -> None: 29 | super().__init__() 30 | 31 | def __call__(self, inputs): 32 | return inputs*tf.clip_by_value(inputs/6+0.5, 0, 1) 33 | 34 | @OPERATOR.register_operator("Mish") 35 | class TFMish(): 36 | def __init__(self, *args, **kwargs) -> None: 37 | super().__init__() 38 | 39 | def __call__(self, inputs): 40 | return inputs*tf.tanh(tf.math.log(tf.math.exp(inputs)+1)) 41 | 42 | @OPERATOR.register_operator("Sigmoid") 43 | class TFSigmoid(): 44 | def __init__(self, *args, **kwargs) -> None: 45 | super().__init__() 46 | 47 | def __call__(self, inputs): 48 | return keras.activations.sigmoid(inputs) 49 | 50 | @OPERATOR.register_operator("LeakyRelu") 51 | class TFLeakyRelu(): 52 | def __init__(self, tensor_grap, node_weights, node_inputs, node_attribute, *args, **kwargs) -> None: 53 | super().__init__() 54 | self.alpha = node_attribute.get('alpha', 0.01) 55 | 56 | def __call__(self, inputs): 57 | return keras.activations.relu(inputs, alpha=self.alpha) 58 | 59 | @OPERATOR.register_operator("PRelu") 60 | class TFPRelu(): 61 | def __init__(self, tensor_grap, node_weights, node_inputs, node_attribute, node_outputs, layout_dict, *args, **kwargs) -> None: 62 | super().__init__() 63 | if 'slope' in node_attribute: 64 | self.slope = node_attribute['slope'] 65 | elif node_inputs[1] in node_weights: 66 | self.slope = node_weights[node_inputs[1]] 67 | else: 68 | self.slope = tensor_grap[node_inputs[1]] 69 | input_tensor_shape = tensor_grap[node_inputs[0]].shape 70 | channel_last = layout_dict[node_inputs[0]] == Layout.Channel_Last 71 | if isinstance(self.slope, np.ndarray): 72 | while self.slope.ndim < input_tensor_shape.ndims: 73 | self.slope = self.slope[np.newaxis, :] 74 | if channel_last: 75 | self.slope = tensor_NCD_to_NDC_format(self.slope) 76 | if self.slope.ndim > 1: 77 | # remove batchsize 78 | self.slope = self.slope[0] 79 | axes = [i for i in range(1, input_tensor_shape.ndims-1)] if channel_last else [i for i in range(2, input_tensor_shape.ndims)] 80 | self.PRelu = tf.keras.layers.PReLU(weights=[self.slope], shared_axes = axes) 81 | 82 | def __call__(self, inputs): 83 | return self.PRelu(inputs) 84 | 85 | @OPERATOR.register_operator("Sin") 86 | class TFSin(): 87 | def __init__(self, *args, **kwargs) -> None: 88 | super().__init__() 89 | 90 | def __call__(self, inputs): 91 | return tf.sin(inputs) 92 | 93 | @OPERATOR.register_operator("Sinh") 94 | class TFSinh(): 95 | def __init__(self, *args, **kwargs) -> None: 96 | super().__init__() 97 | 98 | def __call__(self, inputs): 99 | return tf.sinh(inputs) 100 | 101 | @OPERATOR.register_operator("Cos") 102 | class TFCos(): 103 | def __init__(self, *args, **kwargs) -> None: 104 | super().__init__() 105 | 106 | def __call__(self, inputs): 107 | return tf.cos(inputs) 108 | 109 | @OPERATOR.register_operator("Cosh") 110 | class TFCosh(): 111 | def __init__(self, *args, **kwargs) -> None: 112 | super().__init__() 113 | 114 | def __call__(self, inputs): 115 | return tf.cosh(inputs) 116 | 117 | @OPERATOR.register_operator("Tan") 118 | class TFTan(): 119 | def __init__(self, *args, **kwargs) -> None: 120 | super().__init__() 121 | 122 | def __call__(self, inputs): 123 | return tf.tan(inputs) 124 | 125 | @OPERATOR.register_operator("Tanh") 126 | class TFTanh(): 127 | def __init__(self, *args, **kwargs) -> None: 128 | super().__init__() 129 | 130 | def __call__(self, inputs): 131 | return tf.tanh(inputs) 132 | 133 | @OPERATOR.register_operator("Softmax") 134 | class TFSoftmax(): 135 | def __init__(self, tensor_grap, node_weights, node_inputs, node_attribute, node_outputs, layout_dict, *args, **kwargs) -> None: 136 | super().__init__() 137 | self.axis = node_attribute.get('axis', -1) 138 | if self.axis == -1: 139 | self.axis = len(tensor_grap[node_inputs[0]].shape.as_list()) - 1 140 | if layout_dict[node_inputs[0]] == Layout.Channel_Last: 141 | self.axis = channel_to_last_dimension(self.axis) 142 | 143 | def __call__(self, inputs): 144 | return keras.activations.softmax(inputs, axis=self.axis) 145 | 146 | @OPERATOR.register_operator("Softplus") 147 | class TFSoftplus(): 148 | def __init__(self, tensor_grap, node_weights, node_inputs, node_attribute, *args, **kwargs) -> None: 149 | super().__init__() 150 | 151 | def __call__(self, inputs): 152 | return keras.activations.softplus(inputs) 153 | 154 | @OPERATOR.register_operator("Softsign") 155 | class TFSoftsign(): 156 | def __init__(self, tensor_grap, node_weights, node_inputs, node_attribute, *args, **kwargs) -> None: 157 | super().__init__() 158 | 159 | def __call__(self, inputs): 160 | return keras.activations.softsign(inputs) 161 | 162 | @OPERATOR.register_operator("Selu") 163 | class TFSelu(): 164 | def __init__(self, tensor_grap, node_weights, node_inputs, node_attribute, *args, **kwargs) -> None: 165 | super().__init__() 166 | 167 | def __call__(self, inputs): 168 | return keras.activations.selu(inputs) 169 | 170 | @OPERATOR.register_operator("Elu") 171 | class TFElu(): 172 | def __init__(self, tensor_grap, node_weights, node_inputs, node_attribute, *args, **kwargs) -> None: 173 | super().__init__() 174 | 175 | def __call__(self, inputs): 176 | return keras.activations.elu(inputs) 177 | 178 | @OPERATOR.register_operator("Celu") 179 | class TFCelu(): 180 | def __init__(self, tensor_grap, node_weights, node_inputs, node_attribute, *args, **kwargs) -> None: 181 | super().__init__() 182 | self.alpha = node_attribute.get("alpha", 1.0) 183 | 184 | def __call__(self, inputs): 185 | return tf.maximum(inputs, 0) + tf.minimum(0, self.alpha*(tf.exp(inputs/self.alpha)-1)) -------------------------------------------------------------------------------- /onnx2tflite/onnx2tflite/components/builder1.py: -------------------------------------------------------------------------------- 1 | import os 2 | os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2' 3 | 4 | import tensorflow as tf 5 | from tensorflow import keras 6 | from onnx import numpy_helper 7 | from .dataloader import RandomLoader, ImageLoader 8 | 9 | from onnx2tflite.utils import OPERATOR 10 | from onnx2tflite.layers import conv_layers 11 | from onnx2tflite.utils.definitions import * 12 | from onnx2tflite.utils.graph_tools import build_tf_inputs, decode_node_attribute 13 | 14 | import logging 15 | 16 | # 设置日志 17 | LOG = logging.getLogger("keras_builder") 18 | LOG.setLevel(logging.INFO) # 设定日志级别 19 | 20 | # 添加日志处理器（如果未添加） 21 | if not LOG.hasHandlers(): 22 | handler = logging.StreamHandler() # 输出到控制台 23 | formatter = logging.Formatter("%(levelname)s: %(message)s") # 设置日志格式 24 | handler.setFormatter(formatter) 25 | LOG.addHandler(handler) 26 | 27 | 28 | def keras_builder(onnx_model, native_groupconv:bool=False): 29 | """ 30 | 将 ONNX 模型转换为 Keras 模型。 31 | 32 | 参数: 33 | - onnx_model (onnx.ModelProto): 需要转换的 ONNX 模型。 34 | - native_groupconv (bool, 可选): 是否保持 ONNX 原生的分组卷积操作。默认 False。 35 | 36 | 返回: 37 | - keras_model (keras.Model): 生成的 Keras 模型。 38 | - input_layout (dict): ONNX 输入张量的布局信息。 39 | - output_layout (dict): ONNX 输出张量的布局信息。 40 | """ 41 | # 设置全局变量，控制是否使用原生 ONNX 分组卷积 42 | conv_layers.USE_NATIVE_GROUP_CONV = native_groupconv 43 | 44 | # 解析 ONNX 计算图 45 | model_graph = onnx_model.graph 46 | layout_dict, tf_tensor = {}, {} # 存储 ONNX 层的布局信息 & TensorFlow 层的映射 47 | 48 | ''' 49 | 初始化 ONNX 的权重张量 50 | ''' 51 | 52 | onnx_weights = dict() 53 | 54 | for initializer in model_graph.initializer: 55 | # 将 ONNX 权重转换为 NumPy 数组 56 | onnx_weights[initializer.name] = numpy_helper.to_array(initializer) 57 | 58 | ''' 59 | 解析 ONNX 输入节点并转换为 TensorFlow 输入层 60 | ''' 61 | input_nodes = build_tf_inputs(model_graph, layout_dict) # 解析 ONNX 输入 62 | tf_tensor.update(input_nodes) # 更新 TensorFlow 层的映射字典 63 | 64 | ''' 65 | 遍历 ONNX 计算图中的所有节点，并转换为 TensorFlow 层 66 | ''' 67 | 68 | ########################################################################################################### 69 | for node in model_graph.node: 70 | 71 | op_name, node_inputs, node_outputs = node.op_type, node.input, node.output 72 | op_attr = decode_node_attribute(node) # 解析 ONNX 节点的属性 73 | 74 | # 查找 TensorFlow 对应的操作 75 | tf_operator = OPERATOR.get(op_name) 76 | if tf_operator is None: 77 | raise KeyError(f"{op_name} not implemented yet") 78 | 79 | _inputs = None 80 | if len(node_inputs) > 0: # 如果输入张量已在 `tf_tensor` 中，使用它，否则从 `onnx_weights` 取出 81 | _inputs = tf_tensor[node_inputs[0]] if node_inputs[0] in tf_tensor else onnx_weights[node_inputs[0]] 82 | 83 | # 初始化 layout（数据格式，例如 NHWC） 84 | for index in range(len(node_outputs)): 85 | layout_dict[node_outputs[index]] = layout_dict.get(node_inputs[0], Layout.Default) 86 | 87 | # 执行转换：ONNX 层 -> TensorFlow 层 88 | res = tf_operator(tf_tensor, onnx_weights, node_inputs, op_attr, node_outputs, layout_dict)(_inputs) 89 | 90 | ########################################################################################################### 91 | if isinstance(res, list): # 处理多个输出 92 | for index in range(len(node_outputs)): 93 | tf_tensor[node_outputs[index]] = res[index] 94 | else: 95 | tf_tensor[node_outputs[0]] = res 96 | 97 | ''' 98 | 构建 Keras 模型: 99 | INFO: Keras 模型输入形状: (1, 256, 256, 4) 100 | INFO:keras_builder:Keras 模型输入形状: (1, 256, 256, 4) 101 | INFO: Keras 模型输出形状: (1, 512, 512, 3) 102 | INFO:keras_builder:Keras 模型输出形状: (1, 512, 512, 3) 103 | ''' 104 | input_nodes = [tf_tensor[x.name] for x in model_graph.input] # 获取 ONNX 输入 105 | outputs_nodes = [tf_tensor[x.name] for x in model_graph.output] # 获取 ONNX 输出 106 | keras_model = keras.Model(inputs=input_nodes, outputs=outputs_nodes) # 构建 Keras 模型 107 | keras_model.trainable = False # 设定为不可训练 108 | # keras_model.summary() # 可选，打印模型结构 109 | # print(layout_dict) 110 | 111 | #################################################################### 112 | ''' 113 | 在返回模型之前，检查是否仍然包含动态输入/输出 114 | ''' 115 | # 1. 获取 Keras 模型的输入输出形状 116 | input_shape = keras_model.input_shape 117 | output_shape = keras_model.output_shape 118 | 119 | # 2. 检查是否存在动态输入 (None 表示动态形状) 120 | if any(dim is None for dim in input_shape): 121 | LOG.warning(f"Keras 模型仍然包含动态输入: {input_shape}") 122 | 123 | # 3. 检查是否存在动态输出 124 | if any(dim is None for dim in output_shape): 125 | LOG.warning(f"Keras 模型仍然包含动态输出: {output_shape}") 126 | 127 | # 4. 记录信息 128 | LOG.info(f"Keras 模型输入形状: {input_shape}") 129 | LOG.info(f"Keras 模型输出形状: {output_shape}") 130 | 131 | #################################################################### 132 | # 记录 ONNX 的输入和输出布局 133 | input_layout, output_layout = {}, {} 134 | for inp in model_graph.input: 135 | input_layout[inp.name] = layout_dict[inp.name] 136 | for oup in model_graph.output: 137 | output_layout[oup.name] = layout_dict[oup.name] 138 | 139 | return keras_model, input_layout, output_layout # 返回 Keras 模型和布局信息 140 | 141 | 142 | def tflite_builder(keras_model, weight_quant:bool=False, fp16_model=False, int8_model:bool=False, image_root:str=None, 143 | int8_mean:list or float = [123.675, 116.28, 103.53], int8_std:list or float = [58.395, 57.12, 57.375]): 144 | 145 | """ 146 | 将 Keras 模型转换为 TFLite 模型，并支持不同的量化模式。 147 | 148 | 参数: 149 | - keras_model (keras.Model): 需要转换的 Keras 模型。 150 | - weight_quant (bool, 可选): 是否进行权重量化。默认 False。 151 | - fp16_model (bool, 可选): 是否转换为 FP16 精度（适用于部分硬件优化）。默认 False。 152 | - int8_model (bool, 可选): 是否转换为 INT8 量化模型（适用于边缘设备）。默认 False。 153 | - image_root (str, 可选): 如果使用 INT8 量化，提供用于校准的图像数据目录。 154 | - int8_mean (list or float, 可选): INT8 量化校准的均值，默认 `[123.675, 116.28, 103.53]`。 155 | - int8_std (list or float, 可选): INT8 量化校准的标准差，默认 `[58.395, 57.12, 57.375]`。 156 | 157 | 返回: 158 | - tflite_model (bytes): 转换后的 TFLite 模型。 159 | """ 160 | # 1. 创建 TensorFlow Lite 转换器 161 | converter = tf.lite.TFLiteConverter.from_keras_model(keras_model) 162 | 163 | # 2. 设定转换支持的运算 # TensorFlow Lite 内置算子 # 允许使用部分 TensorFlow 原生算子 164 | converter.target_spec.supported_ops = [tf.lite.OpsSet.TFLITE_BUILTINS, tf.lite.OpsSet.SELECT_TF_OPS] 165 | 166 | # 3. 启用量化选项 167 | if weight_quant or int8_model or fp16_model: 168 | converter.experimental_new_converter = True # 使用新的 TFLite 转换器 169 | converter.optimizations = [tf.lite.Optimize.DEFAULT] # 启用优化 170 | 171 | # 4. 处理 FP16 量化（半精度浮点数） 172 | if fp16_model: 173 | converter.target_spec.supported_types = [tf.float16] 174 | converter.inference_input_type = tf.float32 175 | converter.inference_output_type = tf.float32 176 | 177 | # 5. 处理 INT8 量化 178 | elif int8_model: 179 | assert len(keras_model.inputs) == 1, f"help want, only support single input model." 180 | # 获取输入形状 181 | shape = list(keras_model.inputs[0].shape) 182 | # 选择数据集：使用 `image_root` 进行 INT8 量化校准 183 | dataset = RandomLoader(shape) if image_root is None else ImageLoader(image_root, shape, int8_mean, int8_std) 184 | # 设定代表性数据集（TFLite 量化需要一个校准数据集） 185 | converter.representative_dataset = lambda: dataset 186 | # 使用 INT8 计算 187 | converter.target_spec.supported_ops = [tf.lite.OpsSet.TFLITE_BUILTINS_INT8, tf.lite.OpsSet.SELECT_TF_OPS] 188 | converter.target_spec.supported_types = [] 189 | converter.inference_input_type = tf.uint8 190 | converter.inference_output_type = tf.uint8 191 | converter.experimental_new_converter = True # 启用新的转换器 192 | # 6. 进行 TFLite 转换 193 | tflite_model = converter.convert() 194 | return tflite_model # 返回 TFLite 模型 -------------------------------------------------------------------------------- /onnx2tflite/onnx2tflite.egg-info/PKG-INFO: -------------------------------------------------------------------------------- 1 | Metadata-Version: 2.1 2 | Name: onnx2tflite 3 | Version: 2.0 4 | Summary: onnx to keras/tensorflow lite 5 | Author: MPolaris 6 | License: Apache-2.0 7 | Platform: Windows 8 | Platform: linux 9 | Description-Content-Type: text/markdown 10 | Requires-Dist: onnx 11 | Requires-Dist: onnxruntime 12 | Requires-Dist: onnx-simplifier 13 | Requires-Dist: numpy<=1.24 14 | Requires-Dist: tensorflow<2.13,>=2.5 15 | Requires-Dist: opencv-python 16 | 17 | # ONNX->Keras and ONNX->TFLite tools 18 | ## Welcome 19 | If you have some good ideas, welcome to discuss or give project PRs. 20 | 21 | ## Install 22 | ```cmd 23 | git clone https://github.com/MPolaris/onnx2tflite.git 24 | cd onnx2tflite 25 | python setup.py install 26 | ``` 27 | ```python 28 | from onnx2tflite import onnx_converter 29 | res = onnx_converter( 30 | onnx_model_path = "./model.onnx", 31 | need_simplify = True, 32 | output_path = "./models/", 33 | target_formats = ['tflite'], 34 | ) 35 | ``` 36 | --- 37 | ```cmd 38 | # base 39 | python -m onnx2tflite --weights "./your_model.onnx" 40 | 41 | # give save path 42 | python -m onnx2tflite --weights "./your_model.onnx" --outpath "./save_path" 43 | 44 | # save tflite model 45 | python -m onnx2tflite --weights "./your_model.onnx" --outpath "./save_path" --formats "tflite" 46 | 47 | # save keras and tflite model 48 | python -m onnx2tflite --weights "./your_model.onnx" --outpath "./save_path" --formats "tflite" "keras" 49 | 50 | # cutoff model, redefine inputs and outputs, support middle layers 51 | python -m onnx2tflite --weights "./your_model.onnx" --outpath "./save_path" --formats "tflite" --input-node-names "layer_inputname" --output-node-names "layer_outname1" "layer_outname2" 52 | 53 | # quantify model weight, only weight 54 | python -m onnx2tflite --weights "./your_model.onnx" --formats "tflite" --weigthquant 55 | 56 | # quantify model weight, include input and output 57 | ## fp16 58 | python -m onnx2tflite --weights "./your_model.onnx" --formats "tflite" --fp16 59 | ## recommend 60 | python -m onnx2tflite --weights "./your_model.onnx" --formats "tflite" --int8 --imgroot "./dataset_path" --int8mean 0 0 0 --int8std 255 255 255 61 | ## generate random data, instead of read from image file 62 | python -m onnx2tflite --weights "./your_model.onnx" --formats "tflite" --int8 63 | ``` 64 | --- 65 | ## Features 66 | - High Consistency. Compare to ONNX outputs, average error less than 1e-5 per elements. 67 | - More Faster. Output tensorflow-lite model 30% faster than [onnx_tf](https://github.com/onnx/onnx-tensorflow). 68 | - Auto Channel Align. Auto convert pytorch format(NCWH) to tensorflow format(NWHC). 69 | - Deployment Support. Support output quantitative model, include fp16 quantization and uint8 quantization. 70 | - Code Friendly. I've been trying to keep the code structure simple and clear. 71 | --- 72 | 73 | ## Pytorch -> ONNX -> Tensorflow-Keras -> Tensorflow-Lite 74 | 75 | - ### From torchvision to tensorflow-lite 76 | ```python 77 | import torch 78 | import torchvision 79 | _input = torch.randn(1, 3, 224, 224) 80 | model = torchvision.models.mobilenet_v2(True) 81 | # use default settings is ok 82 | torch.onnx.export(model, _input, './mobilenetV2.onnx', opset_version=11)# or opset_version=13 83 | 84 | from converter import onnx_converter 85 | onnx_converter( 86 | onnx_model_path = "./mobilenetV2.onnx", 87 | need_simplify = True, 88 | output_path = "./", 89 | target_formats = ['tflite'], # or ['keras'], ['keras', 'tflite'] 90 | weight_quant = False, 91 | fp16_model=False, 92 | int8_model = False, 93 | int8_mean = None, 94 | int8_std = None, 95 | image_root = None 96 | ) 97 | ``` 98 | - ### From custom pytorch model to tensorflow-lite-int8 99 | ```python 100 | import torch 101 | import torch.nn as nn 102 | import torch.nn.functional as F 103 | 104 | class MyModel(nn.Module): 105 | def __init__(self): 106 | self.conv = nn.Sequential( 107 | nn.Conv2d(3, 64, kernel_size=3, padding=1), 108 | nn.BatchNorm2d(64), 109 | nn.ReLU(inplace=True), 110 | ) 111 | 112 | def forward(self, x): 113 | return self.conv(x) 114 | 115 | model = MyModel() 116 | model.load_state_dict(torch.load("model_checkpoint.pth", map_location="cpu")) 117 | 118 | _input = torch.randn(1, 3, 224, 224) 119 | torch.onnx.export(model, _input, './mymodel.onnx', opset_version=11)# or opset_version=13 120 | 121 | from converter import onnx_converter 122 | onnx_converter( 123 | onnx_model_path = "./mymodel.onnx", 124 | need_simplify = True, 125 | output_path = "./", 126 | target_formats = ['tflite'], #or ['keras'], ['keras', 'tflite'] 127 | weight_quant = False, 128 | int8_model = True, # do quantification 129 | int8_mean = [123.675, 116.28, 103.53], # give mean of image preprocessing 130 | int8_std = [58.395, 57.12, 57.375], # give std of image preprocessing 131 | image_root = "./dataset/train" # give image folder of train 132 | ) 133 | ``` 134 | --- 135 | ## Validated models 136 | - [SSD](https://github.com/qfgaohao/pytorch-ssd) 137 | - [HRNet](HRNet-Facial-Landmark-Detection) 138 | - [YOLOX](https://github.com/Megvii-BaseDetection/YOLOX) 139 | - [YOLOV3](https://github.com/ultralytics/yolov3) 140 | - [YOLOV4](https://github.com/Tianxiaomo/pytorch-YOLOv4) 141 | - [YOLOV5](https://github.com/ultralytics/yolov5) 142 | - [YOLOV6](https://github.com/meituan/YOLOv6) 143 | - [YOLOV7](https://github.com/WongKinYiu/yolov7) 144 | - [YOLOV10](https://github.com/THU-MIG/yolov10) 145 | - [MoveNet](https://github.com/fire717/movenet.pytorch) 146 | - [UNet\FPN](https://github.com/bigmb/Unet-Segmentation-Pytorch-Nest-of-Unets) 147 | - ViT(torchvision) 148 | - [SwinTransformerV1](https://github.com/microsoft/Swin-Transformer) 149 | - MLP(custom) 150 | - DCGAN(custom) 151 | - [AutoEncoder/VAE](https://github.com/AntixK/PyTorch-VAE) 152 | - all torchvision classification models 153 | - some segmation models in torchvision 154 | - 1D or 2D CNN without special operators(custom) 155 | --- 156 | ## Add operator by yourself 157 | When you counter unspported operator, you can choose to add it by yourself or make an issue.
158 | It's very simple to implement a new operator parser by following these steps below.
159 | Step 0: Select a corresponding layer code file in [layers folder](./onnx2tflite/layers/), such as activations_layers.py for 'HardSigmoid'.
160 | Step 1: Open it, and edit it: 161 | ```python 162 | # all operators regist through OPERATOR register. 163 | # regist operator's name is onnx operator name. 164 | @OPERATOR.register_operator("HardSigmoid") 165 | class TFHardSigmoid(): 166 | def __init__(self, tensor_grap, node_weights, node_inputs, node_attribute, node_outputs, layout_dict, *args, **kwargs) -> None: 167 | ''' 168 | :param tensor_grap: dict, key is node name, value is tensorflow-keras node output tensor. 169 | :param node_weights: dict, key is node name, value is static data, such as weight/bias/constant, weight should be transfom by dimension_utils.tensor_NCD_to_NDC_format at most time. 170 | :param node_inputs: List[str], stored node input names, indicates which nodes the input comes from, tensor_grap and node_weights are possible. 171 | :param node_attribute: dict, key is attribute name, such as 'axis' or 'perm'. value type is indeterminate, such as List[int] or int or float. notice that type of 'axis' value should be adjusted form NCHW to NHWC by dimension_utils.channel_to_last_dimension or dimension_utils.shape_NCD_to_NDC_format. 172 | :param node_inputs: List[str], stored node output names. 173 | :param layout_dict: List[Layout], stored all before node's layout. 174 | ''' 175 | super().__init__() 176 | self.alpha = node_attribute.get("alpha", 0.2) 177 | self.beta = node_attribute.get("beta", 0.5) 178 | 179 | def __call__(self, inputs): 180 | return tf.clip_by_value(self.alpha*inputs+self.beta, 0, 1) 181 | ``` 182 | Step 2: Make it work without error.
183 | Step 3: Convert model to tflite without any quantification.
184 | 185 | --- 186 | 187 | # License 188 | This software is covered by Apache-2.0 license. 189 | -------------------------------------------------------------------------------- /onnx2tflite/onnx2tflite/layers/mathematics_layers.py: -------------------------------------------------------------------------------- 1 | import logging 2 | import numpy as np 3 | import tensorflow as tf 4 | 5 | from onnx2tflite.utils.definitions import Layout 6 | from onnx2tflite.utils import OPERATOR, dimension_utils, np2tf_type 7 | 8 | LOG = logging.getLogger("calculations_layers :") 9 | 10 | def np2tf(x): 11 | if isinstance(x, np.ndarray): 12 | x = tf.convert_to_tensor(x, dtype=np2tf_type[x.dtype.name]) 13 | return x, False 14 | return x, True 15 | 16 | def match_tensor(x1:tf.Tensor or np.ndarray, x2:tf.Tensor or np.ndarray, x1_layout:Layout, x2_layout:Layout): 17 | 18 | x1, f1 = np2tf(x1) 19 | x2, f2 = np2tf(x2) 20 | 21 | # no need to transpose if all var are tensor, we assume tensor are computed by gragh. 22 | if f1 and f2: 23 | if x1_layout != x2_layout: 24 | if x1_layout == Layout.Channel_Last: 25 | x1 = dimension_utils.tensor_NDC_to_NCD_format(x1) 26 | elif x2_layout == Layout.Channel_Last: 27 | x2 = dimension_utils.tensor_NDC_to_NCD_format(x2) 28 | return x1, x2, Layout.Channel_First 29 | 30 | # ensure tensor is set to x1, const weights set to x2 31 | out_layout = x1_layout 32 | if f2: 33 | x1, x2 = x2, x1 34 | out_layout = x2_layout 35 | 36 | 37 | if out_layout == Layout.Channel_Last: 38 | if x1.shape.ndims != x2.shape.ndims: 39 | while x2.shape.ndims < x1.shape.ndims: 40 | x2 = tf.expand_dims(x2, axis=0) 41 | x2 = dimension_utils.tensor_NCD_to_NDC_format(x2) 42 | 43 | x2 = tf.cast(x2, x1.dtype) 44 | return (x2, x1, out_layout) if f2 else (x1, x2, out_layout) 45 | 46 | ''' 47 | tensor(NDC) + const 48 | tensor(NCD) + const 49 | tensor(NDC) + tensor(NDC) 50 | tensor(NCD) + tensor(NCD) 51 | ''' 52 | 53 | class BaseArithmetic: 54 | def __init__(self, tensor_grap, node_weights, node_inputs, node_attribute, node_outputs, layout_dict, *args, **kwargs): 55 | self.left_val, self.right_val = None, None 56 | left_layout, right_layout = Layout.Default, Layout.Default 57 | 58 | if node_inputs[0] in tensor_grap: 59 | self.left_val = tensor_grap[node_inputs[0]] 60 | left_layout = layout_dict[node_inputs[0]] 61 | else: 62 | self.left_val = node_weights[node_inputs[0]] 63 | 64 | if node_inputs[1] in tensor_grap: 65 | self.right_val = tensor_grap[node_inputs[1]] 66 | right_layout = layout_dict[node_inputs[1]] 67 | else: 68 | self.right_val = node_weights[node_inputs[1]] 69 | 70 | if left_layout == right_layout: 71 | return 72 | 73 | self.left_val, self.right_val, out_layout = match_tensor(self.left_val, self.right_val, left_layout, right_layout) 74 | layout_dict[node_outputs[0]] = out_layout 75 | 76 | @OPERATOR.register_operator("Add") 77 | class TFAdd(BaseArithmetic): 78 | def __init__(self, *args, **kwargs): 79 | super().__init__(*args, **kwargs) 80 | 81 | def __call__(self, *args, **kwargs): 82 | return self.left_val + self.right_val 83 | 84 | @OPERATOR.register_operator("Sub") 85 | class TFSub(BaseArithmetic): 86 | def __init__(self, *args, **kwargs): 87 | super().__init__(*args, **kwargs) 88 | 89 | def __call__(self, *args, **kwargs): 90 | return self.left_val - self.right_val 91 | 92 | @OPERATOR.register_operator("Mul") 93 | class TFMul(BaseArithmetic): 94 | def __init__(self,*args, **kwargs): 95 | super().__init__(*args, **kwargs) 96 | 97 | def __call__(self, *args, **kwargs): 98 | return self.left_val * self.right_val 99 | 100 | @OPERATOR.register_operator("Div") 101 | class TFDiv(BaseArithmetic): 102 | def __init__(self,*args, **kwargs): 103 | super().__init__(*args, **kwargs) 104 | 105 | def __call__(self, *args, **kwargs): 106 | return self.left_val / self.right_val 107 | 108 | @OPERATOR.register_operator("MatMul") 109 | class TFMatMul(): 110 | def __init__(self, tensor_grap, node_weights, node_inputs, node_attribute, node_outputs, layout_dict, *args, **kwargs): 111 | super().__init__() 112 | if node_inputs[0] in tensor_grap: 113 | self.A = tensor_grap[node_inputs[0]] 114 | if layout_dict[node_inputs[0]] == Layout.Channel_Last: 115 | self.A = dimension_utils.tensor_NDC_to_NCD_format(self.A) 116 | else: 117 | self.A = node_weights[node_inputs[0]] 118 | 119 | if node_inputs[1] in tensor_grap: 120 | self.B = tensor_grap[node_inputs[1]] 121 | if layout_dict[node_inputs[1]] == Layout.Channel_Last: 122 | self.B = dimension_utils.tensor_NDC_to_NCD_format(self.B) 123 | else: 124 | self.B = node_weights[node_inputs[1]] 125 | 126 | self.dense = tf.keras.layers.Dense(self.B.shape[-1], 127 | weights=[self.B], 128 | use_bias=False) 129 | 130 | layout_dict[node_outputs[0]] = Layout.Channel_First 131 | 132 | def __call__(self, *args, **kwargs): 133 | # out = tf.matmul(self.A, self.B) 134 | try: 135 | out = self.dense(self.A) 136 | except Exception: 137 | out = tf.matmul(self.A, self.B) 138 | return out 139 | 140 | @OPERATOR.register_operator("Mod") 141 | class TFMod(): 142 | def __init__(self, tensor_grap, node_weights, node_inputs, node_attribute, *args, **kwargs): 143 | super().__init__() 144 | self.fmod = bool(node_attribute.get("fmod", 0)) 145 | self.mod_value = None 146 | if node_inputs[1] in node_weights: 147 | self.mod_value = node_weights[node_inputs[1]] 148 | else: 149 | self.mod_value = tensor_grap[node_inputs[1]] 150 | 151 | def __call__(self, inputs): 152 | if self.fmod: 153 | return tf.math.floormod(inputs, tf.cast(self.mod_value, inputs.dtype)) 154 | else: 155 | return tf.math.mod(inputs, tf.cast(self.mod_value, inputs.dtype)) 156 | 157 | @OPERATOR.register_operator("Pow") 158 | class TFPow(): 159 | def __init__(self, tensor_grap, node_weights, node_inputs, *args, **kwargs): 160 | super().__init__() 161 | self.power_index = node_weights[node_inputs[1]] 162 | 163 | def __call__(self, inputs, *args, **kwargs): 164 | return tf.pow(inputs, self.power_index) 165 | 166 | @OPERATOR.register_operator("Reciprocal") 167 | class TFReciprocal(): 168 | def __init__(self, *args, **kwargs): 169 | super().__init__() 170 | 171 | def __call__(self, inputs, *args, **kwargs): 172 | return 1/inputs 173 | 174 | @OPERATOR.register_operator("Sqrt") 175 | class TFSqrt(): 176 | def __init__(self, *args, **kwargs): 177 | super().__init__() 178 | 179 | def __call__(self, inputs, *args, **kwargs): 180 | return tf.sqrt(inputs) 181 | 182 | @OPERATOR.register_operator("Exp") 183 | class TFSqrt(): 184 | def __init__(self, *args, **kwargs): 185 | super().__init__() 186 | 187 | def __call__(self, inputs, *args, **kwargs): 188 | return tf.exp(inputs) 189 | 190 | @OPERATOR.register_operator("Log") 191 | class TFLog(): 192 | def __init__(self, *args, **kwargs): 193 | super().__init__() 194 | 195 | def __call__(self, inputs, *args, **kwargs): 196 | return tf.log(inputs) 197 | 198 | class ReduceBase: 199 | def __init__(self, tensor_grap, node_weights, node_inputs, node_attribute, node_outputs, layout_dict, *args, **kwargs): 200 | self.keep_dims = node_attribute.get("keepdims", 1) == 1 201 | input_shape_len = len(tensor_grap[node_inputs[0]].shape) 202 | if layout_dict[node_inputs[0]] == Layout.Channel_Last: 203 | self.axes = [dimension_utils.channel_to_last_dimension(i) if i >=0 else dimension_utils.channel_to_last_dimension(input_shape_len + i) for i in node_attribute.get("axes", [-1])] 204 | else: 205 | self.axes = [i if i >=0 else input_shape_len + i for i in node_attribute.get("axes", [-1])] 206 | 207 | @OPERATOR.register_operator("ReduceSum") 208 | class TFReduceSum(ReduceBase): 209 | def __init__(self, *args, **kwargs): 210 | super().__init__(*args, **kwargs) 211 | 212 | def __call__(self, inputs, *args, **kwargs): 213 | return tf.math.reduce_sum(inputs, axis=self.axes, keepdims=self.keep_dims) 214 | 215 | @OPERATOR.register_operator("ReduceMean") 216 | class TFReduceMean(ReduceBase): 217 | def __init__(self, *args, **kwargs): 218 | super().__init__(*args, **kwargs) 219 | 220 | def __call__(self, inputs, *args, **kwargs): 221 | return tf.math.reduce_mean(inputs, axis=self.axes, keepdims=self.keep_dims) 222 | 223 | @OPERATOR.register_operator("ReduceMax") 224 | class TFReduceMax(ReduceBase): 225 | def __init__(self, *args, **kwargs): 226 | super().__init__(*args, **kwargs) 227 | 228 | def __call__(self, inputs, *args, **kwargs): 229 | return tf.math.reduce_max(inputs, axis=self.axes, keepdims=self.keep_dims) 230 | 231 | @OPERATOR.register_operator("ReduceMin") 232 | class TFReduceMin(ReduceBase): 233 | def __init__(self, *args, **kwargs): 234 | super().__init__(*args, **kwargs) 235 | 236 | def __call__(self, inputs, *args, **kwargs): 237 | return tf.math.reduce_min(inputs, axis=self.axes, keepdims=self.keep_dims) 238 | 239 | @OPERATOR.register_operator("ArgMax") 240 | class TFArgMax(): 241 | def __init__(self, tensor_grap, node_weights, node_inputs, node_attribute, node_outputs, layout_dict, *args, **kwargs): 242 | super().__init__() 243 | self.axis = node_attribute.get('axis', 0) 244 | if layout_dict[node_inputs[0]] == Layout.Channel_Last: 245 | self.axis = dimension_utils.channel_to_last_dimension(self.axis) 246 | self.keepdims = node_attribute.get("keepdims", 1) == 1 247 | 248 | def __call__(self, inputs, *args, **kwargs): 249 | _inputs = tf.argmax(inputs, axis=self.axis) 250 | if self.keepdims: 251 | _inputs = tf.expand_dims(_inputs, axis=self.axis) 252 | return _inputs 253 | 254 | @OPERATOR.register_operator("ArgMin") 255 | class TFArgMin(): 256 | def __init__(self, tensor_grap, node_weights, node_inputs, node_attribute, node_outputs, layout_dict, *args, **kwargs): 257 | super().__init__() 258 | self.axis = node_attribute.get('axis', 0) 259 | if layout_dict[node_inputs[0]] == Layout.Channel_Last: 260 | self.axis = dimension_utils.channel_to_last_dimension(self.axis) 261 | self.keepdims = node_attribute.get("keepdims", 1) == 1 262 | 263 | def __call__(self, inputs, *args, **kwargs): 264 | _inputs = tf.argmax(inputs, axis=self.axis) 265 | if self.keepdims: 266 | _inputs = tf.expand_dims(_inputs, axis=self.axis) 267 | return _inputs 268 | 269 | @OPERATOR.register_operator("Erf") 270 | class TFErf(): 271 | def __init__(self, *args, **kwargs) -> None: 272 | pass 273 | 274 | def __call__(self, inputs): 275 | inputs = tf.math.erf(inputs) 276 | return inputs -------------------------------------------------------------------------------- /LICENSE: -------------------------------------------------------------------------------- 1 | Apache License 2 | Version 2.0, January 2004 3 | http://www.apache.org/licenses/ 4 | 5 | TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION 6 | 7 | 1. Definitions. 8 | 9 | "License" shall mean the terms and conditions for use, reproduction, 10 | and distribution as defined by Sections 1 through 9 of this document. 11 | 12 | "Licensor" shall mean the copyright owner or entity authorized by 13 | the copyright owner that is granting the License. 14 | 15 | "Legal Entity" shall mean the union of the acting entity and all 16 | other entities that control, are controlled by, or are under common 17 | control with that entity. For the purposes of this definition, 18 | "control" means (i) the power, direct or indirect, to cause the 19 | direction or management of such entity, whether by contract or 20 | otherwise, or (ii) ownership of fifty percent (50%) or more of the 21 | outstanding shares, or (iii) beneficial ownership of such entity. 22 | 23 | "You" (or "Your") shall mean an individual or Legal Entity 24 | exercising permissions granted by this License. 25 | 26 | "Source" form shall mean the preferred form for making modifications, 27 | including but not limited to software source code, documentation 28 | source, and configuration files. 29 | 30 | "Object" form shall mean any form resulting from mechanical 31 | transformation or translation of a Source form, including but 32 | not limited to compiled object code, generated documentation, 33 | and conversions to other media types. 34 | 35 | "Work" shall mean the work of authorship, whether in Source or 36 | Object form, made available under the License, as indicated by a 37 | copyright notice that is included in or attached to the work 38 | (an example is provided in the Appendix below). 39 | 40 | "Derivative Works" shall mean any work, whether in Source or Object 41 | form, that is based on (or derived from) the Work and for which the 42 | editorial revisions, annotations, elaborations, or other modifications 43 | represent, as a whole, an original work of authorship. For the purposes 44 | of this License, Derivative Works shall not include works that remain 45 | separable from, or merely link (or bind by name) to the interfaces of, 46 | the Work and Derivative Works thereof. 47 | 48 | "Contribution" shall mean any work of authorship, including 49 | the original version of the Work and any modifications or additions 50 | to that Work or Derivative Works thereof, that is intentionally 51 | submitted to Licensor for inclusion in the Work by the copyright owner 52 | or by an individual or Legal Entity authorized to submit on behalf of 53 | the copyright owner. For the purposes of this definition, "submitted" 54 | means any form of electronic, verbal, or written communication sent 55 | to the Licensor or its representatives, including but not limited to 56 | communication on electronic mailing lists, source code control systems, 57 | and issue tracking systems that are managed by, or on behalf of, the 58 | Licensor for the purpose of discussing and improving the Work, but 59 | excluding communication that is conspicuously marked or otherwise 60 | designated in writing by the copyright owner as "Not a Contribution." 61 | 62 | "Contributor" shall mean Licensor and any individual or Legal Entity 63 | on behalf of whom a Contribution has been received by Licensor and 64 | subsequently incorporated within the Work. 65 | 66 | 2. Grant of Copyright License. Subject to the terms and conditions of 67 | this License, each Contributor hereby grants to You a perpetual, 68 | worldwide, non-exclusive, no-charge, royalty-free, irrevocable 69 | copyright license to reproduce, prepare Derivative Works of, 70 | publicly display, publicly perform, sublicense, and distribute the 71 | Work and such Derivative Works in Source or Object form. 72 | 73 | 3. Grant of Patent License. Subject to the terms and conditions of 74 | this License, each Contributor hereby grants to You a perpetual, 75 | worldwide, non-exclusive, no-charge, royalty-free, irrevocable 76 | (except as stated in this section) patent license to make, have made, 77 | use, offer to sell, sell, import, and otherwise transfer the Work, 78 | where such license applies only to those patent claims licensable 79 | by such Contributor that are necessarily infringed by their 80 | Contribution(s) alone or by combination of their Contribution(s) 81 | with the Work to which such Contribution(s) was submitted. If You 82 | institute patent litigation against any entity (including a 83 | cross-claim or counterclaim in a lawsuit) alleging that the Work 84 | or a Contribution incorporated within the Work constitutes direct 85 | or contributory patent infringement, then any patent licenses 86 | granted to You under this License for that Work shall terminate 87 | as of the date such litigation is filed. 88 | 89 | 4. Redistribution. You may reproduce and distribute copies of the 90 | Work or Derivative Works thereof in any medium, with or without 91 | modifications, and in Source or Object form, provided that You 92 | meet the following conditions: 93 | 94 | (a) You must give any other recipients of the Work or 95 | Derivative Works a copy of this License; and 96 | 97 | (b) You must cause any modified files to carry prominent notices 98 | stating that You changed the files; and 99 | 100 | (c) You must retain, in the Source form of any Derivative Works 101 | that You distribute, all copyright, patent, trademark, and 102 | attribution notices from the Source form of the Work, 103 | excluding those notices that do not pertain to any part of 104 | the Derivative Works; and 105 | 106 | (d) If the Work includes a "NOTICE" text file as part of its 107 | distribution, then any Derivative Works that You distribute must 108 | include a readable copy of the attribution notices contained 109 | within such NOTICE file, excluding those notices that do not 110 | pertain to any part of the Derivative Works, in at least one 111 | of the following places: within a NOTICE text file distributed 112 | as part of the Derivative Works; within the Source form or 113 | documentation, if provided along with the Derivative Works; or, 114 | within a display generated by the Derivative Works, if and 115 | wherever such third-party notices normally appear. The contents 116 | of the NOTICE file are for informational purposes only and 117 | do not modify the License. You may add Your own attribution 118 | notices within Derivative Works that You distribute, alongside 119 | or as an addendum to the NOTICE text from the Work, provided 120 | that such additional attribution notices cannot be construed 121 | as modifying the License. 122 | 123 | You may add Your own copyright statement to Your modifications and 124 | may provide additional or different license terms and conditions 125 | for use, reproduction, or distribution of Your modifications, or 126 | for any such Derivative Works as a whole, provided Your use, 127 | reproduction, and distribution of the Work otherwise complies with 128 | the conditions stated in this License. 129 | 130 | 5. Submission of Contributions. Unless You explicitly state otherwise, 131 | any Contribution intentionally submitted for inclusion in the Work 132 | by You to the Licensor shall be under the terms and conditions of 133 | this License, without any additional terms or conditions. 134 | Notwithstanding the above, nothing herein shall supersede or modify 135 | the terms of any separate license agreement you may have executed 136 | with Licensor regarding such Contributions. 137 | 138 | 6. Trademarks. This License does not grant permission to use the trade 139 | names, trademarks, service marks, or product names of the Licensor, 140 | except as required for reasonable and customary use in describing the 141 | origin of the Work and reproducing the content of the NOTICE file. 142 | 143 | 7. Disclaimer of Warranty. Unless required by applicable law or 144 | agreed to in writing, Licensor provides the Work (and each 145 | Contributor provides its Contributions) on an "AS IS" BASIS, 146 | WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or 147 | implied, including, without limitation, any warranties or conditions 148 | of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A 149 | PARTICULAR PURPOSE. You are solely responsible for determining the 150 | appropriateness of using or redistributing the Work and assume any 151 | risks associated with Your exercise of permissions under this License. 152 | 153 | 8. Limitation of Liability. In no event and under no legal theory, 154 | whether in tort (including negligence), contract, or otherwise, 155 | unless required by applicable law (such as deliberate and grossly 156 | negligent acts) or agreed to in writing, shall any Contributor be 157 | liable to You for damages, including any direct, indirect, special, 158 | incidental, or consequential damages of any character arising as a 159 | result of this License or out of the use or inability to use the 160 | Work (including but not limited to damages for loss of goodwill, 161 | work stoppage, computer failure or malfunction, or any and all 162 | other commercial damages or losses), even if such Contributor 163 | has been advised of the possibility of such damages. 164 | 165 | 9. Accepting Warranty or Additional Liability. While redistributing 166 | the Work or Derivative Works thereof, You may choose to offer, 167 | and charge a fee for, acceptance of support, warranty, indemnity, 168 | or other liability obligations and/or rights consistent with this 169 | License. However, in accepting such obligations, You may act only 170 | on Your own behalf and on Your sole responsibility, not on behalf 171 | of any other Contributor, and only if You agree to indemnify, 172 | defend, and hold each Contributor harmless for any liability 173 | incurred by, or claims asserted against, such Contributor by reason 174 | of your accepting any such warranty or additional liability. 175 | 176 | END OF TERMS AND CONDITIONS 177 | 178 | APPENDIX: How to apply the Apache License to your work. 179 | 180 | To apply the Apache License to your work, attach the following 181 | boilerplate notice, with the fields enclosed by brackets "[]" 182 | replaced with your own identifying information. (Don't include 183 | the brackets!) The text should be enclosed in the appropriate 184 | comment syntax for the file format. We also recommend that a 185 | file or class name and description of purpose be included on the 186 | same "printed page" as the copyright notice for easier 187 | identification within third-party archives. 188 | 189 | Copyright [yyyy] [name of copyright owner] 190 | 191 | Licensed under the Apache License, Version 2.0 (the "License"); 192 | you may not use this file except in compliance with the License. 193 | You may obtain a copy of the License at 194 | 195 | http://www.apache.org/licenses/LICENSE-2.0 196 | 197 | Unless required by applicable law or agreed to in writing, software 198 | distributed under the License is distributed on an "AS IS" BASIS, 199 | WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 200 | See the License for the specific language governing permissions and 201 | limitations under the License. 202 | -------------------------------------------------------------------------------- /model/utils.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import torch.nn as nn 3 | import torch.nn.functional as F 4 | 5 | class MBRConv5(nn.Module): 6 | def __init__(self, in_channels, out_channels, rep_scale=4): 7 | super(MBRConv5, self).__init__() 8 | self.in_channels = in_channels 9 | self.out_channels = out_channels 10 | self.conv = nn.Conv2d(in_channels, out_channels * rep_scale, 5, 1, 2) 11 | self.conv_bn = nn.Sequential( 12 | nn.BatchNorm2d(out_channels * rep_scale) 13 | ) 14 | self.conv1 = nn.Conv2d(in_channels, out_channels * rep_scale, 1) 15 | self.conv1_bn = nn.Sequential( 16 | nn.BatchNorm2d(out_channels * rep_scale) 17 | ) 18 | self.conv2 = nn.Conv2d(in_channels, out_channels * rep_scale, 3, 1, 1) 19 | self.conv2_bn = nn.Sequential( 20 | nn.BatchNorm2d(out_channels * rep_scale) 21 | ) 22 | self.conv_crossh = nn.Conv2d(in_channels, out_channels * rep_scale, (3, 1), 1, (1, 0)) 23 | self.conv_crossh_bn = nn.Sequential( 24 | nn.BatchNorm2d(out_channels * rep_scale) 25 | ) 26 | self.conv_crossv = nn.Conv2d(in_channels, out_channels * rep_scale, (1, 3), 1, (0, 1)) 27 | self.conv_crossv_bn = nn.Sequential( 28 | nn.BatchNorm2d(out_channels * rep_scale) 29 | ) 30 | self.conv_out = nn.Conv2d(out_channels * rep_scale * 10, out_channels, 1) 31 | 32 | def forward(self, inp): 33 | x1 = self.conv(inp) 34 | x2 = self.conv1(inp) 35 | x3 = self.conv2(inp) 36 | x4 = self.conv_crossh(inp) 37 | x5 = self.conv_crossv(inp) 38 | x = torch.cat( 39 | [x1, x2, x3, x4, x5, 40 | self.conv_bn(x1), 41 | self.conv1_bn(x2), 42 | self.conv2_bn(x3), 43 | self.conv_crossh_bn(x4), 44 | self.conv_crossv_bn(x5)], 45 | 1 46 | ) 47 | out = self.conv_out(x) 48 | return out 49 | 50 | def slim(self): 51 | conv_weight = self.conv.weight 52 | conv_bias = self.conv.bias 53 | 54 | conv1_weight = self.conv1.weight 55 | conv1_bias = self.conv1.bias 56 | conv1_weight = nn.functional.pad(conv1_weight, (2, 2, 2, 2)) 57 | 58 | conv2_weight = self.conv2.weight 59 | conv2_weight = nn.functional.pad(conv2_weight, (1, 1, 1, 1)) 60 | conv2_bias = self.conv2.bias 61 | 62 | conv_crossv_weight = self.conv_crossv.weight 63 | conv_crossv_weight = nn.functional.pad(conv_crossv_weight, (1, 1, 2, 2)) 64 | conv_crossv_bias = self.conv_crossv.bias 65 | 66 | conv_crossh_weight = self.conv_crossh.weight 67 | conv_crossh_weight = nn.functional.pad(conv_crossh_weight, (2, 2, 1, 1)) 68 | conv_crossh_bias = self.conv_crossh.bias 69 | 70 | conv1_bn_weight = self.conv1.weight 71 | conv1_bn_weight = nn.functional.pad(conv1_bn_weight, (2, 2, 2, 2)) 72 | 73 | conv2_bn_weight = self.conv2.weight 74 | conv2_bn_weight = nn.functional.pad(conv2_bn_weight, (1, 1, 1, 1)) 75 | 76 | conv_crossv_bn_weight = self.conv_crossv.weight 77 | conv_crossv_bn_weight = nn.functional.pad(conv_crossv_bn_weight, (1, 1, 2, 2)) 78 | 79 | conv_crossh_bn_weight = self.conv_crossh.weight 80 | conv_crossh_bn_weight = nn.functional.pad(conv_crossh_bn_weight, (2, 2, 1, 1)) 81 | 82 | bn = self.conv_bn[0] 83 | k = 1 / (bn.running_var + bn.eps) ** .5 84 | b = - bn.running_mean / (bn.running_var + bn.eps) ** .5 85 | 86 | conv_bn_weight = self.conv.weight * k.unsqueeze(-1).unsqueeze(-1).unsqueeze(-1) 87 | conv_bn_weight = conv_bn_weight * bn.weight.unsqueeze(-1).unsqueeze(-1).unsqueeze(-1) 88 | conv_bn_bias = self.conv.bias * k + b 89 | conv_bn_bias = conv_bn_bias * bn.weight + bn.bias 90 | 91 | bn = self.conv1_bn[0] 92 | k = 1 / (bn.running_var + bn.eps) ** .5 93 | b = - bn.running_mean / (bn.running_var + bn.eps) ** .5 94 | conv1_bn_weight = conv1_bn_weight * k.unsqueeze(-1).unsqueeze(-1).unsqueeze(-1) 95 | conv1_bn_weight = conv1_bn_weight * bn.weight.unsqueeze(-1).unsqueeze(-1).unsqueeze(-1) 96 | conv1_bn_bias = self.conv1.bias * k + b 97 | conv1_bn_bias = conv1_bn_bias * bn.weight + bn.bias 98 | 99 | bn = self.conv2_bn[0] 100 | k = 1 / (bn.running_var + bn.eps) ** .5 101 | b = - bn.running_mean / (bn.running_var + bn.eps) ** .5 102 | conv2_bn_weight = conv2_bn_weight * k.unsqueeze(-1).unsqueeze(-1).unsqueeze(-1) 103 | conv2_bn_weight = conv2_bn_weight * bn.weight.unsqueeze(-1).unsqueeze(-1).unsqueeze(-1) 104 | conv2_bn_bias = self.conv2.bias * k + b 105 | conv2_bn_bias = conv2_bn_bias * bn.weight + bn.bias 106 | 107 | bn = self.conv_crossv_bn[0] 108 | k = 1 / (bn.running_var + bn.eps) ** .5 109 | b = - bn.running_mean / (bn.running_var + bn.eps) ** .5 110 | conv_crossv_bn_weight = conv_crossv_bn_weight * k.unsqueeze(-1).unsqueeze(-1).unsqueeze(-1) 111 | conv_crossv_bn_weight = conv_crossv_bn_weight * bn.weight.unsqueeze(-1).unsqueeze(-1).unsqueeze(-1) 112 | conv_crossv_bn_bias = self.conv_crossv.bias * k + b 113 | conv_crossv_bn_bias = conv_crossv_bn_bias * bn.weight + bn.bias 114 | 115 | bn = self.conv_crossh_bn[0] 116 | k = 1 / (bn.running_var + bn.eps) ** .5 117 | b = - bn.running_mean / (bn.running_var + bn.eps) ** .5 118 | conv_crossh_bn_weight = conv_crossh_bn_weight * k.unsqueeze(-1).unsqueeze(-1).unsqueeze(-1) 119 | conv_crossh_bn_weight = conv_crossh_bn_weight * bn.weight.unsqueeze(-1).unsqueeze(-1).unsqueeze(-1) 120 | conv_crossh_bn_bias = self.conv_crossh.bias * k + b 121 | conv_crossh_bn_bias = conv_crossh_bn_bias * bn.weight + bn.bias 122 | 123 | weight = torch.cat( 124 | [conv_weight, conv1_weight, conv2_weight, 125 | conv_crossh_weight, conv_crossv_weight, 126 | conv_bn_weight, conv1_bn_weight, conv2_bn_weight, 127 | conv_crossh_bn_weight, conv_crossv_bn_weight], 128 | 0 129 | ) 130 | weight_compress = self.conv_out.weight.squeeze() 131 | weight = torch.matmul(weight_compress, weight.permute([2, 3, 0, 1])).permute([2, 3, 0, 1]) 132 | bias_ = torch.cat( 133 | [conv_bias, conv1_bias, conv2_bias, 134 | conv_crossh_bias, conv_crossv_bias, 135 | conv_bn_bias, conv1_bn_bias, conv2_bn_bias, 136 | conv_crossh_bn_bias, conv_crossv_bn_bias], 137 | 0 138 | ) 139 | bias = torch.matmul(weight_compress, bias_) 140 | if isinstance(self.conv_out.bias, torch.Tensor): 141 | bias = bias + self.conv_out.bias 142 | return weight, bias 143 | 144 | 145 | ############################################################################################################## 146 | class MBRConv3(nn.Module): 147 | def __init__(self, in_channels, out_channels, rep_scale=4): 148 | super(MBRConv3, self).__init__() 149 | 150 | self.in_channels = in_channels 151 | self.out_channels = out_channels 152 | self.rep_scale = rep_scale 153 | 154 | self.conv = nn.Conv2d(in_channels, out_channels * rep_scale, 3, 1, 1) 155 | self.conv_bn = nn.Sequential( 156 | nn.BatchNorm2d(out_channels * rep_scale) 157 | ) 158 | self.conv1 = nn.Conv2d(in_channels, out_channels * rep_scale, 1) 159 | self.conv1_bn = nn.Sequential( 160 | nn.BatchNorm2d(out_channels * rep_scale) 161 | ) 162 | self.conv_crossh = nn.Conv2d(in_channels, out_channels * rep_scale, (3, 1), 1, (1, 0)) 163 | self.conv_crossh_bn = nn.Sequential( 164 | nn.BatchNorm2d(out_channels * rep_scale) 165 | ) 166 | self.conv_crossv = nn.Conv2d(in_channels, out_channels * rep_scale, (1, 3), 1, (0, 1)) 167 | self.conv_crossv_bn = nn.Sequential( 168 | nn.BatchNorm2d(out_channels * rep_scale) 169 | ) 170 | self.conv_out = nn.Conv2d(out_channels * rep_scale * 8, out_channels, 1) 171 | 172 | def forward(self, inp): 173 | x0 = self.conv(inp) 174 | x1 = self.conv1(inp) 175 | x2 = self.conv_crossh(inp) 176 | x3 = self.conv_crossv(inp) 177 | x = torch.cat( 178 | [ x0,x1,x2,x3, 179 | self.conv_bn(x0), 180 | self.conv1_bn(x1), 181 | self.conv_crossh_bn(x2), 182 | self.conv_crossv_bn(x3)], 183 | 1 184 | ) 185 | out = self.conv_out(x) 186 | return out 187 | 188 | def slim(self): 189 | conv_weight = self.conv.weight 190 | conv_bias = self.conv.bias 191 | 192 | conv1_weight = self.conv1.weight 193 | conv1_bias = self.conv1.bias 194 | conv1_weight = F.pad(conv1_weight, (1, 1, 1, 1)) 195 | 196 | conv_crossh_weight = self.conv_crossh.weight 197 | conv_crossh_bias = self.conv_crossh.bias 198 | conv_crossh_weight = F.pad(conv_crossh_weight, (1, 1, 0, 0)) 199 | 200 | conv_crossv_weight = self.conv_crossv.weight 201 | conv_crossv_bias = self.conv_crossv.bias 202 | conv_crossv_weight = F.pad(conv_crossv_weight, (0, 0, 1, 1)) 203 | 204 | # conv_bn 205 | bn = self.conv_bn[0] 206 | k = 1 / torch.sqrt(bn.running_var + bn.eps) 207 | conv_bn_weight = self.conv.weight * k.unsqueeze(-1).unsqueeze(-1).unsqueeze(-1) 208 | conv_bn_weight = conv_bn_weight * bn.weight.unsqueeze(-1).unsqueeze(-1).unsqueeze(-1) 209 | conv_bn_bias = self.conv.bias * k + (-bn.running_mean * k) 210 | conv_bn_bias = conv_bn_bias * bn.weight + bn.bias 211 | 212 | # conv1_bn 213 | bn = self.conv1_bn[0] 214 | k = 1 / torch.sqrt(bn.running_var + bn.eps) 215 | conv1_bn_weight = self.conv1.weight * k.unsqueeze(-1).unsqueeze(-1).unsqueeze(-1) 216 | conv1_bn_weight = conv1_bn_weight * bn.weight.unsqueeze(-1).unsqueeze(-1).unsqueeze(-1) 217 | conv1_bn_weight = F.pad(conv1_bn_weight, (1, 1, 1, 1)) 218 | conv1_bn_bias = self.conv1.bias * k + (-bn.running_mean * k) 219 | conv1_bn_bias = conv1_bn_bias * bn.weight + bn.bias 220 | 221 | # conv_crossh_bn 222 | bn = self.conv_crossh_bn[0] 223 | k = 1 / torch.sqrt(bn.running_var + bn.eps) 224 | conv_crossh_bn_weight = self.conv_crossh.weight * k.unsqueeze(-1).unsqueeze(-1).unsqueeze(-1) 225 | conv_crossh_bn_weight = conv_crossh_bn_weight * bn.weight.unsqueeze(-1).unsqueeze(-1).unsqueeze(-1) 226 | conv_crossh_bn_weight = F.pad(conv_crossh_bn_weight, (1, 1, 0, 0)) 227 | conv_crossh_bn_bias = self.conv_crossh.bias * k + (-bn.running_mean * k) 228 | conv_crossh_bn_bias = conv_crossh_bn_bias * bn.weight + bn.bias 229 | 230 | # conv_crossv_bn 231 | bn = self.conv_crossv_bn[0] 232 | k = 1 / torch.sqrt(bn.running_var + bn.eps) 233 | conv_crossv_bn_weight = self.conv_crossv.weight * k.unsqueeze(-1).unsqueeze(-1).unsqueeze(-1) 234 | conv_crossv_bn_weight = conv_crossv_bn_weight * bn.weight.unsqueeze(-1).unsqueeze(-1).unsqueeze(-1) 235 | conv_crossv_bn_weight = F.pad(conv_crossv_bn_weight, (0, 0, 1, 1)) 236 | conv_crossv_bn_bias = self.conv_crossv.bias * k + (-bn.running_mean * k) 237 | conv_crossv_bn_bias = conv_crossv_bn_bias * bn.weight + bn.bias 238 | 239 | weight = torch.cat([ 240 | conv_weight, 241 | conv1_weight, 242 | conv_crossh_weight, 243 | conv_crossv_weight, 244 | conv_bn_weight, 245 | conv1_bn_weight, 246 | conv_crossh_bn_weight, 247 | conv_crossv_bn_weight 248 | ], dim=0) 249 | 250 | bias = torch.cat([ 251 | conv_bias, 252 | conv1_bias, 253 | conv_crossh_bias, 254 | conv_crossv_bias, 255 | conv_bn_bias, 256 | conv1_bn_bias, 257 | conv_crossh_bn_bias, 258 | conv_crossv_bn_bias 259 | ], dim=0) 260 | 261 | weight_compress = self.conv_out.weight.squeeze() 262 | weight = torch.matmul(weight_compress, weight.view(weight.size(0), -1)) 263 | weight = weight.view(self.conv_out.out_channels, self.in_channels, 3, 3) 264 | 265 | bias = torch.matmul(weight_compress, bias.unsqueeze(-1)).squeeze(-1) 266 | if self.conv_out.bias is not None: 267 | bias += self.conv_out.bias 268 | 269 | return weight, bias 270 | 271 | ###################################################################################################### 272 | class MBRConv1(nn.Module): 273 | def __init__(self, in_channels, out_channels, rep_scale=4): 274 | super(MBRConv1, self).__init__() 275 | 276 | self.in_channels = in_channels 277 | self.out_channels = out_channels 278 | self.rep_scale = rep_scale 279 | 280 | self.conv = nn.Conv2d(in_channels, out_channels * rep_scale, 1) 281 | self.conv_bn = nn.Sequential( 282 | nn.BatchNorm2d(out_channels * rep_scale) 283 | ) 284 | self.conv_out = nn.Conv2d(out_channels * rep_scale * 2, out_channels, 1) 285 | 286 | def forward(self, inp): 287 | x0 = self.conv(inp) 288 | x = torch.cat([x0, self.conv_bn(x0)], 1) 289 | out = self.conv_out(x) 290 | return out 291 | 292 | def slim(self): 293 | conv_weight = self.conv.weight 294 | conv_bias = self.conv.bias 295 | 296 | bn = self.conv_bn[0] 297 | k = 1 / (bn.running_var + bn.eps) ** .5 298 | b = - bn.running_mean / (bn.running_var + bn.eps) ** .5 299 | conv_bn_weight = self.conv.weight * k.unsqueeze(-1).unsqueeze(-1).unsqueeze(-1) 300 | conv_bn_weight = conv_bn_weight * bn.weight.unsqueeze(-1).unsqueeze(-1).unsqueeze(-1) 301 | conv_bn_bias = self.conv.bias * k + b 302 | conv_bn_bias = conv_bn_bias * bn.weight + bn.bias 303 | 304 | weight = torch.cat([conv_weight, conv_bn_weight], 0) 305 | weight_compress = self.conv_out.weight.squeeze() 306 | weight = torch.matmul(weight_compress, weight.permute([2, 3, 0, 1])).permute([2, 3, 0, 1]) 307 | 308 | bias = torch.cat([conv_bias, conv_bn_bias], 0) 309 | bias = torch.matmul(weight_compress, bias) 310 | 311 | if isinstance(self.conv_out.bias, torch.Tensor): 312 | bias = bias + self.conv_out.bias 313 | return weight, bias 314 | 315 | class FST(nn.Module): 316 | def __init__(self, block1, channels): 317 | super(FST, self).__init__() 318 | self.block1 = block1 319 | self.weight1 = nn.Parameter(torch.randn(1)) 320 | self.weight2 = nn.Parameter(torch.randn(1)) 321 | self.bias = nn.Parameter(torch.randn((1, channels, 1, 1))) 322 | 323 | def forward(self, x): 324 | x1 = self.block1(x) 325 | weighted_block1 = self.weight1 * x1 326 | weighted_block2 = self.weight2 * x1 327 | return weighted_block1 * weighted_block2 + self.bias 328 | 329 | class FSTS(nn.Module): 330 | def __init__(self, block1, channels): 331 | super(FSTS, self).__init__() 332 | self.block1 = block1 333 | self.weight1 = nn.Parameter(torch.randn(1)) 334 | self.weight2 = nn.Parameter(torch.randn(1)) 335 | self.bias = nn.Parameter(torch.randn((1, channels, 1, 1))) 336 | 337 | def forward(self, x): 338 | x1 = self.block1(x) 339 | weighted_block1 = self.weight1 * x1 340 | weighted_block2 = self.weight2 * x1 341 | return weighted_block1 * weighted_block2 + self.bias 342 | ################################################################################## 343 | class DropBlock(nn.Module): 344 | def __init__(self, block_size, p=0.5): 345 | super(DropBlock, self).__init__() 346 | self.block_size = block_size 347 | self.p = p / block_size / block_size 348 | 349 | def forward(self, x): 350 | mask = 1 - (torch.rand_like(x[:, :1]) >= self.p).float() 351 | mask = nn.functional.max_pool2d(mask, self.block_size, 1, self.block_size // 2) 352 | return x * (1 - mask) 353 | -------------------------------------------------------------------------------- /onnx2tflite/onnx2tflite/layers/deformation_layers.py: -------------------------------------------------------------------------------- 1 | import logging 2 | import tensorflow as tf 3 | 4 | from onnx2tflite.utils.definitions import Layout 5 | from onnx2tflite.utils import OPERATOR, dimension_utils 6 | 7 | LOG = logging.getLogger("deformation_layers :") 8 | 9 | @OPERATOR.register_operator("Transpose") 10 | class TFTranspose(): 11 | def __init__(self, tensor_grap, node_weights, node_inputs, node_attribute, node_outputs, layout_dict, *args, **kwargs)->None: 12 | super().__init__() 13 | for nop in node_outputs: 14 | layout_dict[nop] = Layout.Channel_First 15 | if kwargs.get("perm_list"): 16 | self.perm_list = kwargs.get("perm_list") 17 | return 18 | self.trans_in = None 19 | self.perm_list = [i for i in node_attribute['perm']] 20 | if layout_dict[node_inputs[0]] == Layout.Channel_Last: 21 | # LOG.info("Transpose will process tensor after change back to NCHW format.") 22 | shape_len = len(tensor_grap[node_inputs[0]].shape) 23 | self.trans_in = [0, shape_len-1] + [n for n in range(1, shape_len-1)] 24 | 25 | def __call__(self, inputs): 26 | if self.trans_in: 27 | inputs = tf.transpose(inputs, perm=self.trans_in) 28 | return tf.transpose(inputs, perm=self.perm_list) 29 | 30 | @OPERATOR.register_operator("Slice") 31 | class TFSlice(): 32 | def __init__(self, tensor_grap, node_weights, node_inputs, node_attribute, node_outputs, layout_dict, *args, **kwargs) -> None: 33 | super().__init__() 34 | if len(node_inputs) == 1: 35 | self.starts = node_attribute['starts'][0] 36 | self.ends = node_attribute['ends'][0] 37 | self.axis = node_attribute['axes'][0] 38 | self.steps = 1 39 | else: 40 | self.starts = node_weights[node_inputs[1]][0] if node_inputs[1] in node_weights else tensor_grap[node_inputs[1]][0] 41 | self.axis = node_weights[node_inputs[3]][0] if node_inputs[3] in node_weights else tensor_grap[node_inputs[3]][0] 42 | self.ends = node_weights[node_inputs[2]][0] if node_inputs[2] in node_weights else tensor_grap[node_inputs[2]][0] 43 | self.ends = min(self.ends, tensor_grap[node_inputs[0]].shape[self.axis]) 44 | if len(node_inputs) < 5: 45 | self.steps = 1 46 | else: 47 | self.steps = node_weights[node_inputs[4]][0] if node_inputs[4] in node_weights else tensor_grap[node_inputs[4]][0] 48 | 49 | shape = tensor_grap[node_inputs[0]].shape.as_list() 50 | if self.starts < 0: 51 | self.starts = shape[self.axis] + self.starts 52 | if self.ends < 0: 53 | self.ends = shape[self.axis] + self.ends 54 | 55 | if layout_dict[node_inputs[0]] == Layout.Channel_Last: 56 | self.axis = dimension_utils.channel_to_last_dimension(self.axis) 57 | 58 | def __call__(self, inputs): 59 | indices = tf.keras.backend.arange(self.starts, self.ends, step=self.steps) 60 | return tf.gather(inputs, indices, axis=self.axis) 61 | 62 | @OPERATOR.register_operator("Gather") 63 | class TFGather(): 64 | def __init__(self, tensor_grap, node_weights, node_inputs, node_attribute, node_outputs, layout_dict, *args, **kwargs) -> None: 65 | super().__init__() 66 | self.axis = node_attribute.get('axis', 0) 67 | self.indices = tensor_grap[node_inputs[1]] if node_inputs[1] in tensor_grap else node_weights[node_inputs[1]] 68 | if layout_dict[node_inputs[0]] == Layout.Channel_Last: 69 | self.axis = dimension_utils.channel_to_last_dimension(self.axis) 70 | 71 | def __call__(self, inputs): 72 | return tf.gather(inputs, self.indices, axis=self.axis) 73 | 74 | @OPERATOR.register_operator("Concat") 75 | class TFConcat(): 76 | def __init__(self, tensor_grap, node_weights, node_inputs, node_attribute, node_outputs, layout_dict, *args, **kwargs): 77 | super().__init__() 78 | #TODO can be optimzer by watch after node, if conv to be channel last. 79 | self._axis = node_attribute['axis'] 80 | # use `count` to count how much more for channel-last to channel-first 81 | count = 0 82 | for inp in node_inputs: 83 | if inp in node_weights: 84 | count -= 1 85 | elif layout_dict[inp] == Layout.Channel_Last: 86 | count += 1 87 | else: 88 | count -= 1 89 | 90 | self._gather = [] 91 | if count < 0: 92 | # align to Channel_First 93 | layout_dict[node_outputs[0]] = Layout.Channel_First 94 | for inp in node_inputs: 95 | if inp in tensor_grap: 96 | if layout_dict[inp] == Layout.Channel_Last: 97 | tensor_grap[inp] = dimension_utils.tensor_NDC_to_NCD_format(tensor_grap[inp]) 98 | self._gather.append(tensor_grap[inp]) 99 | else: 100 | self._gather.append(node_weights[inp]) 101 | else: 102 | # align to Channel_Last 103 | layout_dict[node_outputs[0]] = Layout.Channel_Last 104 | self._axis = dimension_utils.channel_to_last_dimension(self._axis) 105 | for inp in node_inputs: 106 | if inp in tensor_grap: 107 | if layout_dict[inp] != Layout.Channel_Last: 108 | tensor_grap[inp] = dimension_utils.tensor_NCD_to_NDC_format(tensor_grap[inp]) 109 | self._gather.append(tensor_grap[inp]) 110 | else: 111 | self._gather.append(dimension_utils.tensor_NCD_to_NDC_format(node_weights[inp])) 112 | 113 | def __call__(self, *args, **kwargs): 114 | return tf.concat(self._gather, axis=self._axis) 115 | 116 | @OPERATOR.register_operator("Reshape") 117 | class TFReshape(): 118 | def __init__(self, tensor_grap, node_weights, node_inputs, node_attribute, node_outputs, layout_dict, *args, **kwargs): 119 | super().__init__() 120 | self.out_shape = node_weights[node_inputs[1]] 121 | self.trans_in = None 122 | # LOG.info("Reshape will process tensor after change back to NCHW format.") 123 | if layout_dict[node_inputs[0]] == Layout.Channel_Last: 124 | shape_len = len(tensor_grap[node_inputs[0]].shape) 125 | self.trans_in = [0, shape_len-1] + [n for n in range(1, shape_len-1)] 126 | for nop in node_outputs: 127 | layout_dict[nop] = Layout.Channel_First 128 | 129 | def __call__(self, inputs): 130 | if self.trans_in: 131 | inputs = tf.transpose(inputs, perm=self.trans_in) 132 | inputs = tf.reshape(inputs, shape=self.out_shape) 133 | return inputs 134 | 135 | @OPERATOR.register_operator("Flatten") 136 | class TFFlatten(): 137 | def __init__(self, tensor_grap, node_weights, node_inputs, node_attribute, node_outputs, layout_dict, *args, **kwargs)->None: 138 | super().__init__() 139 | num_elements = int(tensor_grap[node_inputs[0]].shape.num_elements()/tensor_grap[node_inputs[0]].shape[0]) 140 | input_shape = tensor_grap[node_inputs[0]].shape 141 | self.flat = tf.keras.layers.Flatten() 142 | ''' 143 | ensure memory order match, for example: 144 | onnx = (B, 2, 3, 4).reshape(B, -1) 145 | tflite = (B, 3, 4, 2).reshape(B, -1) 146 | we can observe that: 147 | onnx.shape == tflite.shape, but np.sum(onnx-tflite) != 0 148 | it's cause memory order of two vars is different, we must make tflite back to onnx by transpose. 149 | generally, this situation is general one, below is just special situation and most appear in cnn. 150 | onnx = (B, 512, 1, 1) 151 | tflite = (B, 1, 1, 512) 152 | or = (B, 1, 512, 1) 153 | these memory order are all same. 154 | ''' 155 | self.perm = None 156 | if layout_dict[node_inputs[0]] == Layout.Channel_Last and num_elements != max(input_shape[1:]): 157 | self.perm = [0, len(input_shape)-1] 158 | for i in range(len(input_shape)-2): 159 | self.perm.append(i+1) 160 | 161 | def __call__(self, inputs): 162 | if self.perm: 163 | inputs = tf.transpose(inputs, perm=self.perm) 164 | return self.flat(inputs) 165 | 166 | @OPERATOR.register_operator("Split") 167 | class TFSplit(): 168 | def __init__(self, tensor_grap, node_weights, node_inputs, node_attribute, node_outputs, layout_dict, *args, **kwargs)->None: 169 | super().__init__() 170 | self.outputs_nums = len(node_outputs) 171 | self.axis = node_attribute.get("axis", 0) 172 | if layout_dict[node_inputs[0]] == Layout.Channel_Last: 173 | self.axis = dimension_utils.channel_to_last_dimension(self.axis) 174 | split_args = None 175 | if 'split' in node_attribute: 176 | split_args = node_attribute['split'] 177 | else: 178 | assert len(node_inputs) == 2 and node_inputs[1] in node_weights 179 | split_args = node_weights[node_inputs[1]] 180 | 181 | self.indices = [] 182 | start, end = 0, 0 183 | for i in range(self.outputs_nums): 184 | end = start + int(split_args[i]) 185 | self.indices.append(tf.keras.backend.arange(start, end, 1)) 186 | start = end 187 | 188 | def __call__(self, inputs): 189 | return [tf.gather(inputs, indices=indice, axis=self.axis) for indice in self.indices] 190 | 191 | @OPERATOR.register_operator("Expand") 192 | class TFExpand(): 193 | def __init__(self, tensor_grap, node_weights, node_inputs, node_attribute, node_outputs, layout_dict, *args, **kwargs)->None: 194 | super().__init__() 195 | self.shape = node_weights[node_inputs[1]] 196 | if layout_dict[node_inputs[0]] == Layout.Channel_Last: 197 | self.shape = dimension_utils.shape_NCD_to_NDC_format(self.shape) 198 | def __call__(self, inputs): 199 | for i in range(len(self.shape)): 200 | if int(self.shape[i]//inputs.shape[i]) > 1: 201 | inputs = tf.repeat(inputs, repeats=int(self.shape[i]//inputs.shape[i]), axis=i) 202 | elif self.shape[i] < inputs.shape[i] and self.shape[i] != 1: 203 | inputs = tf.repeat(inputs, repeats=int(self.shape[i]), axis=i) 204 | return inputs 205 | 206 | @OPERATOR.register_operator("GatherElements") 207 | class TFGatherElements(): 208 | def __init__(self, tensor_grap, node_weights, node_inputs, node_attribute, node_outputs, layout_dict, *args, **kwargs) -> None: 209 | super().__init__() 210 | self.axis = node_attribute.get("axis", 1) 211 | self.indices = None 212 | if 'indices' in node_attribute: 213 | self.indices = node_attribute['indices'] 214 | self.indices = dimension_utils.tensor_NCD_to_NDC_format(self.indices) 215 | elif node_inputs[1] in node_weights: 216 | self.indices = node_weights[node_inputs[1]] 217 | self.indices = dimension_utils.tensor_NCD_to_NDC_format(self.indices) 218 | else: 219 | self.indices = tensor_grap[node_inputs[1]] 220 | if layout_dict[node_inputs[0]] == Layout.Channel_Last: 221 | self.axis = dimension_utils.channel_to_last_dimension(self.axis) 222 | if len(node_inputs) == 1 or layout_dict[node_inputs[1]] != Layout.Channel_Last: 223 | self.indices = dimension_utils.tensor_NCD_to_NDC_format(self.indices) 224 | 225 | def gather_elements(self, input_tensor, indices, axis): 226 | # Get the shape of the input tensor and the indices tensor 227 | input_shape = tf.shape(input_tensor) 228 | indices_shape = tf.shape(indices) 229 | 230 | # Create indices for all dimensions 231 | idx = tf.meshgrid(*[tf.range(s) for s in indices_shape], indexing='ij') 232 | idx = [tf.cast(i, tf.int64) for i in idx] 233 | 234 | # Replace the axis index with the provided indices 235 | idx[axis] = tf.cast(indices, tf.int64) 236 | 237 | # Stack indices to form the final gather indices 238 | gather_indices = tf.stack(idx, axis=-1) 239 | 240 | # Use tf.gather_nd to gather elements 241 | output_tensor = tf.gather_nd(input_tensor, gather_indices) 242 | 243 | return output_tensor 244 | 245 | def __call__(self, inputs): 246 | return self.gather_elements(inputs, self.indices, self.axis) 247 | 248 | @OPERATOR.register_operator("Tile") 249 | class TFTile(): 250 | def __init__(self, tensor_grap, node_weights, node_inputs, node_attribute, node_outputs, layout_dict, *args, **kwargs)->None: 251 | super().__init__() 252 | self.repeats = node_attribute['repeats'] if 'repeats' in node_attribute else node_weights[node_inputs[1]] 253 | if layout_dict[node_inputs[0]] == Layout.Channel_Last: 254 | self.repeats = dimension_utils.shape_NCD_to_NDC_format(self.repeats) 255 | 256 | def __call__(self, inputs): 257 | for i in range(len(self.repeats)): 258 | if self.repeats[i] > 1: 259 | inputs = tf.repeat(inputs, self.repeats[i], axis=i) 260 | return inputs 261 | 262 | @OPERATOR.register_operator("Unsqueeze") 263 | class TFUnsqueeze(): 264 | def __init__(self, tensor_grap, node_weights, node_inputs, node_attribute, node_outputs, layout_dict, *args, **kwargs)->None: 265 | super().__init__() 266 | self.axis = node_attribute['axes'] if 'axes' in node_attribute else node_weights[node_inputs[1]] 267 | if not isinstance(self.axis, int): 268 | self.axis = int(self.axis[0]) 269 | input_shape = tensor_grap[node_inputs[0]].shape 270 | if len(input_shape) == 1: 271 | layout_dict[node_outputs[0]] = Layout.Channel_None 272 | elif len(input_shape) == 2: 273 | layout_dict[node_outputs[0]] = Layout.Channel_First 274 | else: 275 | layout_dict[node_outputs[0]] = layout_dict[node_inputs[0]] 276 | if layout_dict[node_inputs[0]] == Layout.Channel_Last: 277 | self.axis = dimension_utils.channel_to_last_dimension(self.axis) 278 | 279 | def __call__(self, inputs): 280 | return tf.expand_dims(inputs, self.axis) 281 | 282 | @OPERATOR.register_operator("Squeeze") 283 | class TFSqueeze(): 284 | def __init__(self, tensor_grap, node_weights, node_inputs, node_attribute, node_outputs, layout_dict, *args, **kwargs)->None: 285 | super().__init__() 286 | self.axis = node_attribute['axes'] if 'axes' in node_attribute else node_weights[node_inputs[1]] 287 | if not isinstance(self.axis, int): 288 | self.axis = int(self.axis[0]) 289 | input_shape = tensor_grap[node_inputs[0]].shape 290 | if len(input_shape) <= 3: 291 | layout_dict[node_outputs[0]] = Layout.Channel_None 292 | if len(input_shape) > 2 and layout_dict[node_inputs[0]] == Layout.Channel_Last: 293 | self.axis = dimension_utils.channel_to_last_dimension(self.axis) 294 | 295 | def __call__(self, inputs): 296 | return tf.squeeze(inputs, self.axis) 297 | 298 | @OPERATOR.register_operator("DepthToSpace") 299 | class TFDepthToSpace(): 300 | def __init__(self, tensor_grap, node_weights, node_inputs, node_attribute, node_outputs, layout_dict, *args, **kwargs)->None: 301 | super().__init__() 302 | self.block_size = node_attribute.get("blocksize", 2) 303 | self.mode = node_attribute.get("mode", "DCR") 304 | self.channel_last = layout_dict[node_inputs[0]] == Layout.Channel_Last 305 | 306 | def __call__(self, inputs): 307 | if not self.channel_last: 308 | inputs = dimension_utils.tensor_NDC_to_NCD_format(inputs) 309 | if self.mode == "DCR": 310 | return tf.nn.depth_to_space(inputs, self.block_size) 311 | elif self.mode == "CRD": 312 | # help want, native tensorflow is not support CRD mode, this way will generate 5 dims op. 313 | b, h, w, c = inputs.shape 314 | inputs = tf.reshape(inputs, [b, h, w, c//(self.block_size * self.block_size), self.block_size, self.block_size]) 315 | inputs = tf.transpose(inputs, perm=[0, 1, 4, 2, 5, 3]) 316 | inputs = tf.reshape(inputs, [b, h*self.block_size, w*self.block_size, c//(self.block_size * self.block_size)]) 317 | return inputs 318 | else: 319 | raise KeyError(f"For DepthToSpace, mode must be [DCR, CRD], not {self.mode}") 320 | -------------------------------------------------------------------------------- /onnx2tflite/onnx2tflite/layers/common_layers.py: -------------------------------------------------------------------------------- 1 | import math 2 | import logging 3 | import numpy as np 4 | import tensorflow as tf 5 | from tensorflow import keras 6 | 7 | from onnx2tflite.utils.definitions import Layout 8 | from onnx2tflite.utils import OPERATOR, intfloat_to_list, dimension_utils 9 | 10 | LOG = logging.getLogger("common_layers :") 11 | 12 | @OPERATOR.register_operator("BatchNormalization") 13 | class TFBatchNormalization(): 14 | def __init__(self, tensor_grap, node_weights, node_inputs, node_attribute, *args, **kwargs): 15 | super().__init__() 16 | epsilon = node_attribute.get("epsilon", 1e-5) 17 | momentum = node_attribute.get("momentum", 0.9) 18 | 19 | self.bn = keras.layers.BatchNormalization( 20 | gamma_initializer=keras.initializers.Constant(node_weights[node_inputs[1]]), 21 | beta_initializer=keras.initializers.Constant(node_weights[node_inputs[2]]), 22 | moving_mean_initializer=keras.initializers.Constant(node_weights[node_inputs[3]]), 23 | moving_variance_initializer=keras.initializers.Constant(node_weights[node_inputs[4]]), 24 | epsilon=epsilon, 25 | momentum=momentum) 26 | 27 | def __call__(self, inputs): 28 | return self.bn(inputs) 29 | 30 | @OPERATOR.register_operator("InstanceNormalization") 31 | class TFInstanceNormalization(): 32 | def __init__(self, tensor_grap, node_weights, node_inputs, node_attribute, node_outputs, layout_dict, *args, **kwargs): 33 | super().__init__() 34 | self.epsilon = node_attribute.get("epsilon", 1e-5) 35 | self.scale = node_weights[node_inputs[1]] 36 | self.bias = node_weights[node_inputs[2]] 37 | self.channel_last = layout_dict[node_inputs[0]] == Layout.Channel_Last 38 | 39 | def __call__(self, inputs): 40 | axes = tuple(range(1, len(inputs.shape)-1)) if self.channel_last else tuple(range(2, len(inputs.shape))) 41 | mean = tf.reduce_mean(inputs, axis=axes, keepdims=True) 42 | var = tf.math.reduce_variance(inputs, axis= axes, keepdims=True) 43 | return self.scale*(inputs - mean)/tf.sqrt(var + self.epsilon) + self.bias 44 | 45 | @OPERATOR.register_operator("Pad") 46 | class TFPad(): 47 | def __init__(self, tensor_grap, node_weights, node_inputs, node_attribute, node_outputs, layout_dict, *args, **kwargs): 48 | super().__init__() 49 | if node_attribute.get("pads") is not None: 50 | pads = node_attribute['pads'] 51 | elif node_inputs[1] in node_weights: 52 | pads = node_weights[node_inputs[1]] 53 | else: 54 | pads = tensor_grap[node_inputs[1]] 55 | self.pad = [[pads[0], pads[4]], [pads[2], pads[6]], [pads[3], pads[7]], [pads[1], pads[5]]] 56 | self.model = node_attribute.get("mode", "constant").upper() 57 | self.channel_last = layout_dict[node_inputs[0]] == Layout.Channel_Last 58 | layout_dict[node_outputs[0]] = Layout.Channel_Last 59 | 60 | def __call__(self, inputs): 61 | if not self.channel_last: 62 | inputs = dimension_utils.tensor_NCD_to_NDC_format(inputs) 63 | return tf.pad(inputs, self.pad, mode=self.model) 64 | 65 | @OPERATOR.register_operator("Clip") 66 | class TFClip(): 67 | def __init__(self, tensor_grap, node_weights, node_inputs, node_attribute, *args, **kwargs): 68 | super().__init__() 69 | if "min" in node_attribute: 70 | self.min = node_attribute.get("min") 71 | else: 72 | self.min = tensor_grap[node_inputs[1]] if node_inputs[1] in tensor_grap else node_weights[node_inputs[1]] 73 | if "max" in node_attribute: 74 | self.max = node_attribute.get("max") 75 | else: 76 | self.max = tensor_grap[node_inputs[2]] if node_inputs[2] in tensor_grap else node_weights[node_inputs[2]] 77 | 78 | def __call__(self, inputs): 79 | if float(self.min) == 0 and float(self.max) == 6: 80 | return tf.nn.relu6(inputs) 81 | return tf.clip_by_value(inputs, self.min, self.max) 82 | 83 | @OPERATOR.register_operator("TFGlobalMaxPool") 84 | class TFGlobalMaxPool(): 85 | def __init__(self, tensor_grap, node_weights, node_inputs, node_attribute, node_outputs, layout_dict, *args, **kwargs) -> None: 86 | super().__init__() 87 | self.channel_last = layout_dict[node_inputs[0]] == Layout.Channel_Last 88 | 89 | def __call__(self, inputs): 90 | if self.channel_last: 91 | return tf.reduce_max(inputs, axis=[i for i in range(1, len(inputs.shape)-1)], keepdims=True) 92 | else: 93 | return tf.reduce_max(inputs, axis=[i for i in range(2, len(inputs.shape))], keepdims=True) 94 | 95 | @OPERATOR.register_operator("GlobalAveragePool") 96 | class TFGlobalAveragePool(): 97 | def __init__(self, tensor_grap, node_weights, node_inputs, node_attribute, node_outputs, layout_dict, *args, **kwargs) -> None: 98 | super().__init__() 99 | self.channel_last = layout_dict[node_inputs[0]] == Layout.Channel_Last 100 | 101 | def __call__(self, inputs): 102 | if self.channel_last: 103 | return tf.reduce_mean(inputs, axis=[i for i in range(1, len(inputs.shape)-1)], keepdims=True) 104 | else: 105 | return tf.reduce_mean(inputs, axis=[i for i in range(2, len(inputs.shape))], keepdims=True) 106 | 107 | @OPERATOR.register_operator("AveragePool") 108 | class TFAveragePool(): 109 | def __init__(self, tensor_grap, node_weights, node_inputs, node_attribute, node_outputs, layout_dict, *args, **kwargs) -> None: 110 | super().__init__() 111 | kernel_shape = intfloat_to_list(node_attribute.get("kernel_shape", [2, 2]), 2) 112 | strides = intfloat_to_list(node_attribute.get("strides", [1, 1]), 2) 113 | dilations = intfloat_to_list(node_attribute.get("dilations", [1, 1]), 2) 114 | ceil_mode = node_attribute.get("ceil_mode", 0) 115 | pads = intfloat_to_list(node_attribute.get("pads", [0, 0, 0, 0]), 4) 116 | 117 | func = math.floor if ceil_mode == 0 else math.ceil 118 | 119 | pad_mode = "SAME" 120 | input_shape = tensor_grap[node_inputs[0]].shape 121 | for i in range(len(input_shape)-2): 122 | pad_shape = pads[i] + pads[i+2] 123 | onnx_output_shape = func((input_shape[1+i]+pad_shape-((kernel_shape[i]-1)*dilations[i]+1))/strides[i]+1) 124 | tf_output_shape = math.floor((input_shape[1+i] - kernel_shape[i]) / strides[i]) + 1 125 | pads[2+i] = max(onnx_output_shape-tf_output_shape, pads[2+i]) # right_down pad 126 | if pad_mode == "SAME" and onnx_output_shape != input_shape[1+i]: 127 | pad_mode = "VALID" 128 | self.avg_pool = keras.layers.AveragePooling2D(pool_size=kernel_shape, strides=strides, padding=pad_mode) 129 | 130 | self.pad = None 131 | if pad_mode == "VALID" and pads is not None and np.sum(pads) > 0: 132 | if np.sum(pads) > 0: 133 | self.pad = keras.layers.ZeroPadding2D(padding=((pads[0], pads[2]), (pads[1], pads[3]))) 134 | 135 | self.channel_last = layout_dict[node_inputs[0]] == Layout.Channel_Last 136 | layout_dict[node_outputs[0]] = Layout.Channel_Last 137 | 138 | def __call__(self, inputs): 139 | if not self.channel_last: 140 | inputs = dimension_utils.tensor_NCD_to_NDC_format(inputs) 141 | if self.pad: 142 | inputs = self.pad(inputs) 143 | return self.avg_pool(inputs) 144 | 145 | @OPERATOR.register_operator("MaxPool") 146 | class TFMaxPool(): 147 | def __init__(self, tensor_grap, node_weights, node_inputs, node_attribute, node_outputs, layout_dict, *args, **kwargs) -> None: 148 | super().__init__() 149 | kernel_shape = intfloat_to_list(node_attribute.get("kernel_shape", [2, 2]), 2) 150 | strides = intfloat_to_list(node_attribute.get("strides", [1, 1]), 2) 151 | dilations = intfloat_to_list(node_attribute.get("dilations", [1, 1]), 2) 152 | ceil_mode = node_attribute.get("ceil_mode", 0) 153 | pads = intfloat_to_list(node_attribute.get("pads", [0, 0, 0, 0]), 4) 154 | 155 | func = math.floor if ceil_mode == 0 else math.ceil 156 | 157 | pad_mode = "SAME" 158 | input_shape = tensor_grap[node_inputs[0]].shape 159 | for i in range(len(input_shape)-2): 160 | pad_shape = pads[i] + pads[i+2] 161 | onnx_output_shape = func((input_shape[1+i]+pad_shape-((kernel_shape[i]-1)*dilations[i]+1))/strides[i]+1) 162 | tf_output_shape = math.floor((input_shape[1+i] - kernel_shape[i]) / strides[i]) + 1 163 | pads[2+i] = max(onnx_output_shape-tf_output_shape, pads[2+i]) # right_down pad 164 | if pad_mode == "SAME" and onnx_output_shape != input_shape[1+i]: 165 | pad_mode = "VALID" 166 | self.max_pool = keras.layers.MaxPool2D(pool_size=kernel_shape, strides=strides, padding=pad_mode) 167 | 168 | self.pad = None 169 | if pad_mode == "VALID" and pads is not None and np.sum(pads) > 0: 170 | if np.sum(pads) > 0: 171 | self.pad = keras.layers.ZeroPadding2D(padding=((pads[0], pads[2]), (pads[1], pads[3]))) 172 | 173 | self.channel_last = layout_dict[node_inputs[0]] == Layout.Channel_Last 174 | layout_dict[node_outputs[0]] = Layout.Channel_Last 175 | 176 | def __call__(self, inputs): 177 | if not self.channel_last: 178 | inputs = dimension_utils.tensor_NCD_to_NDC_format(inputs) 179 | if self.pad: 180 | inputs = self.pad(inputs) 181 | return self.max_pool(inputs) 182 | 183 | @OPERATOR.register_operator("Upsample") 184 | class TFUpsample(): 185 | def __init__(self, tensor_grap, node_weights, node_inputs, node_attribute, node_outputs, layout_dict, *args, **kwargs): 186 | super().__init__() 187 | _, h, w, _ = tensor_grap[node_inputs[0]].shape 188 | scale = node_weights[node_inputs[1]] 189 | 190 | self.scale = (int(h*scale[2]), int(w*scale[3])) 191 | if node_attribute.get("mode", "nearest").lower() == 'nearest': 192 | self.method = tf.image.ResizeMethod.NEAREST_NEIGHBOR 193 | else: 194 | self.method = tf.image.ResizeMethod.BILINEAR 195 | 196 | self.channel_last = layout_dict[node_inputs[0]] == Layout.Channel_Last 197 | layout_dict[node_outputs[0]] = Layout.Channel_Last 198 | 199 | def __call__(self, inputs): 200 | if not self.channel_last: 201 | inputs = dimension_utils.tensor_NCD_to_NDC_format(inputs) 202 | return tf.image.resize(inputs, self.scale, method=self.method) 203 | 204 | @OPERATOR.register_operator("Constant") 205 | class TFConstant(): 206 | def __init__(self, tensor_grap, node_weights, node_inputs, node_attribute, *args, **kwargs): 207 | super().__init__() 208 | self.val = node_attribute['value'] 209 | 210 | def __call__(self, *args, **kwargs): 211 | return self.val 212 | 213 | @OPERATOR.register_operator("ScatterND") 214 | class TFScatterND(): 215 | def __init__(self, tensor_grap, node_weights, node_inputs, node_attribute, node_outputs, layout_dict, *args, **kwargs): 216 | super().__init__() 217 | self.indices = node_weights[node_inputs[1]] 218 | self.channle_last = layout_dict[node_inputs[0]] == Layout.Channel_Last 219 | if node_inputs[2] in tensor_grap: 220 | self.updates = tensor_grap[node_inputs[2]] 221 | if self.channle_last: 222 | self.updates = dimension_utils.tensor_NDC_to_NCD_format(self.updates) 223 | else: 224 | self.updates = node_weights[node_inputs[2]] 225 | 226 | layout_dict[node_outputs[0]] = Layout.Channel_First 227 | 228 | def __call__(self, inputs): 229 | if self.channle_last: 230 | inputs = dimension_utils.tensor_NDC_to_NCD_format(inputs) 231 | inputs = tf.tensor_scatter_nd_update(inputs, self.indices, self.updates) 232 | return inputs 233 | 234 | @OPERATOR.register_operator("Resize") 235 | class TFResize(): 236 | def __init__(self, tensor_grap, node_weights, node_inputs, node_attribute, node_outputs, layout_dict, *args, **kwargs): 237 | super().__init__() 238 | if node_inputs[-1] in node_weights: 239 | _, _, nh, nw = node_weights[node_inputs[-1]] 240 | if len(node_inputs) != 4: 241 | _, h, w, _ = tensor_grap[node_inputs[0]].shape 242 | nh, nw = int(h*nh), int(w*nw) 243 | self.scale = (nh, nw) 244 | else: 245 | scales = tensor_grap[node_inputs[0]].shape[1:3]*tensor_grap[node_inputs[2]][2:3] 246 | self.scale = scales 247 | 248 | if node_attribute.get("mode", "nearest").lower() == 'nearest': 249 | self.method = tf.image.ResizeMethod.NEAREST_NEIGHBOR 250 | else: 251 | self.method = tf.image.ResizeMethod.BILINEAR 252 | 253 | self.channel_last = layout_dict[node_inputs[0]] == Layout.Channel_Last 254 | layout_dict[node_outputs[0]] = Layout.Channel_Last 255 | 256 | def __call__(self, inputs): 257 | if not self.channel_last: 258 | inputs = dimension_utils.tensor_NCD_to_NDC_format(inputs) 259 | return tf.image.resize(inputs, self.scale, method=self.method) 260 | 261 | @OPERATOR.register_operator("Gemm") 262 | class TFGemm(): 263 | def __init__(self, tensor_grap, node_weights, node_inputs, node_attribute, node_outputs, layout_dict, *args, **kwargs) -> None: 264 | super().__init__() 265 | if len(node_inputs) > 2: 266 | weights = [node_weights[node_inputs[1]].T, node_weights[node_inputs[2]]] 267 | else: 268 | weights = [node_weights[node_inputs[1]].T] 269 | 270 | self.dense = keras.layers.Dense(weights[0].shape[1], 271 | weights=weights, 272 | use_bias=len(weights)==2) 273 | 274 | self.channel_last = layout_dict[node_inputs[0]] == Layout.Channel_Last 275 | layout_dict[node_outputs[0]] = Layout.Channel_Last 276 | 277 | def __call__(self, inputs): 278 | if not self.channel_last: 279 | inputs = dimension_utils.tensor_NCD_to_NDC_format(inputs) 280 | return self.dense(inputs) 281 | 282 | @OPERATOR.register_operator("Identity") 283 | class TFIdentity(): 284 | def __init__(self, *args, **kwargs): 285 | super().__init__() 286 | 287 | def __call__(self, inputs): 288 | return inputs 289 | 290 | @OPERATOR.register_operator("Dropout") 291 | class TFDropout(): 292 | ''' 293 | Dropout will be ignored in deployment. 294 | ''' 295 | def __init__(self, *args, **kwargs): 296 | super().__init__() 297 | 298 | def __call__(self, inputs): 299 | return inputs 300 | 301 | @OPERATOR.register_operator("TopK") 302 | class TFTopK(): 303 | def __init__(self, tensor_grap, node_weights, node_inputs, node_attribute, *args, **kwargs) -> None: 304 | 305 | self.axis = node_attribute.get("axis", -1) 306 | self.largest = node_attribute.get("largest", 1) 307 | self.sorted = bool(node_attribute.get("sorted", 1)) 308 | self.K = node_attribute.get('K') if len(node_inputs)==1 else node_weights[node_inputs[1]][0] 309 | 310 | def __call__(self, inputs): 311 | res = tf.math.top_k(inputs, k=self.K, sorted=self.sorted) 312 | return [res[0], res[1]] 313 | 314 | @OPERATOR.register_operator("Cast") 315 | class TFCast(): 316 | def __init__(self, tensor_grap, node_weights, node_inputs, node_attribute, *args, **kwargs): 317 | super().__init__() 318 | self.cast_to = int(node_attribute.get("to", 1)) 319 | assert self.cast_to > 0 and self.cast_to < 12, f"Unknown cast type [{self.cast_to}]" 320 | self.np_cast_map = { 321 | 1: np.float32, 322 | 2: np.uint8, 323 | 3: np.int8, 324 | 5: np.int16, 325 | 6: np.int32, 326 | 7: np.int64, 327 | 9: np.bool_, 328 | 10: np.float16, 329 | 11: np.double, 330 | } 331 | self.tf_cast_map = { 332 | 1: tf.float32, 333 | 2: tf.uint8, 334 | 3: tf.int8, 335 | 5: tf.int16, 336 | 6: tf.int32, 337 | 7: tf.int64, 338 | 9: tf.bool, 339 | 10: tf.float16, 340 | 11: tf.double, 341 | } 342 | 343 | def __call__(self, inputs): 344 | if isinstance(inputs, list): 345 | for i in range(len(inputs)): 346 | if isinstance(inputs[i], np.ndarray) or isinstance(inputs[i], np.generic): 347 | inputs[i] = self.np_cast_map[self.cast_to](inputs[i]) 348 | else: 349 | inputs[i] = tf.cast(input[i], dtype=self.tf_cast_map[self.cast_to]) 350 | else: 351 | if isinstance(inputs, np.ndarray) or isinstance(inputs, np.generic): 352 | inputs = self.np_cast_map[self.cast_to](inputs) 353 | else: 354 | inputs = tf.cast(inputs, dtype=self.tf_cast_map[self.cast_to]) 355 | 356 | return inputs 357 | -------------------------------------------------------------------------------- /model/utils_IWO.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import torch.nn as nn 3 | import torch.nn.functional as F 4 | 5 | class MBRConv5(nn.Module): 6 | def __init__(self, in_channels, out_channels, rep_scale=4): 7 | super(MBRConv5, self).__init__() 8 | self.in_channels = in_channels 9 | self.out_channels = out_channels 10 | self.conv = nn.Conv2d(in_channels, out_channels * rep_scale, 5, 1, 2) 11 | self.conv_bn = nn.Sequential( 12 | nn.BatchNorm2d(out_channels * rep_scale) 13 | ) 14 | self.conv1 = nn.Conv2d(in_channels, out_channels * rep_scale, 1) 15 | self.conv1_bn = nn.Sequential( 16 | nn.BatchNorm2d(out_channels * rep_scale) 17 | ) 18 | self.conv2 = nn.Conv2d(in_channels, out_channels * rep_scale, 3, 1, 1) 19 | self.conv2_bn = nn.Sequential( 20 | nn.BatchNorm2d(out_channels * rep_scale) 21 | ) 22 | self.conv_crossh = nn.Conv2d(in_channels, out_channels * rep_scale, (3, 1), 1, (1, 0)) 23 | self.conv_crossh_bn = nn.Sequential( 24 | nn.BatchNorm2d(out_channels * rep_scale) 25 | ) 26 | self.conv_crossv = nn.Conv2d(in_channels, out_channels * rep_scale, (1, 3), 1, (0, 1)) 27 | self.conv_crossv_bn = nn.Sequential( 28 | nn.BatchNorm2d(out_channels * rep_scale) 29 | ) 30 | self.conv_out = nn.Conv2d(out_channels * rep_scale * 10, out_channels, 1) 31 | self.conv_out.weight.requires_grad = False 32 | self.weight1 = nn.Parameter(torch.zeros_like(self.conv_out.weight)) 33 | nn.init.xavier_normal_(self.weight1) 34 | 35 | def forward(self, inp): 36 | x1 = self.conv(inp) 37 | x2 = self.conv1(inp) 38 | x3 = self.conv2(inp) 39 | x4 = self.conv_crossh(inp) 40 | x5 = self.conv_crossv(inp) 41 | x = torch.cat( 42 | [x1, x2, x3, x4, x5, 43 | self.conv_bn(x1), 44 | self.conv1_bn(x2), 45 | self.conv2_bn(x3), 46 | self.conv_crossh_bn(x4), 47 | self.conv_crossv_bn(x5)], 48 | 1 49 | ) 50 | final_weight = self.conv_out.weight + self.weight1 51 | out = F.conv2d(x, final_weight, self.conv_out.bias) 52 | return out 53 | 54 | def slim(self): 55 | conv_weight = self.conv.weight 56 | conv_bias = self.conv.bias 57 | 58 | conv1_weight = self.conv1.weight 59 | conv1_bias = self.conv1.bias 60 | conv1_weight = nn.functional.pad(conv1_weight, (2, 2, 2, 2)) 61 | 62 | conv2_weight = self.conv2.weight 63 | conv2_weight = nn.functional.pad(conv2_weight, (1, 1, 1, 1)) 64 | conv2_bias = self.conv2.bias 65 | 66 | conv_crossv_weight = self.conv_crossv.weight 67 | conv_crossv_weight = nn.functional.pad(conv_crossv_weight, (1, 1, 2, 2)) 68 | conv_crossv_bias = self.conv_crossv.bias 69 | 70 | conv_crossh_weight = self.conv_crossh.weight 71 | conv_crossh_weight = nn.functional.pad(conv_crossh_weight, (2, 2, 1, 1)) 72 | conv_crossh_bias = self.conv_crossh.bias 73 | 74 | conv1_bn_weight = self.conv1.weight 75 | conv1_bn_weight = nn.functional.pad(conv1_bn_weight, (2, 2, 2, 2)) 76 | 77 | conv2_bn_weight = self.conv2.weight 78 | conv2_bn_weight = nn.functional.pad(conv2_bn_weight, (1, 1, 1, 1)) 79 | 80 | conv_crossv_bn_weight = self.conv_crossv.weight 81 | conv_crossv_bn_weight = nn.functional.pad(conv_crossv_bn_weight, (1, 1, 2, 2)) 82 | 83 | conv_crossh_bn_weight = self.conv_crossh.weight 84 | conv_crossh_bn_weight = nn.functional.pad(conv_crossh_bn_weight, (2, 2, 1, 1)) 85 | 86 | bn = self.conv_bn[0] 87 | k = 1 / (bn.running_var + bn.eps) ** .5 88 | b = - bn.running_mean / (bn.running_var + bn.eps) ** .5 89 | 90 | conv_bn_weight = self.conv.weight * k.unsqueeze(-1).unsqueeze(-1).unsqueeze(-1) 91 | conv_bn_weight = conv_bn_weight * bn.weight.unsqueeze(-1).unsqueeze(-1).unsqueeze(-1) 92 | conv_bn_bias = self.conv.bias * k + b 93 | conv_bn_bias = conv_bn_bias * bn.weight + bn.bias 94 | 95 | bn = self.conv1_bn[0] 96 | k = 1 / (bn.running_var + bn.eps) ** .5 97 | b = - bn.running_mean / (bn.running_var + bn.eps) ** .5 98 | conv1_bn_weight = conv1_bn_weight * k.unsqueeze(-1).unsqueeze(-1).unsqueeze(-1) 99 | conv1_bn_weight = conv1_bn_weight * bn.weight.unsqueeze(-1).unsqueeze(-1).unsqueeze(-1) 100 | conv1_bn_bias = self.conv1.bias * k + b 101 | conv1_bn_bias = conv1_bn_bias * bn.weight + bn.bias 102 | 103 | bn = self.conv2_bn[0] 104 | k = 1 / (bn.running_var + bn.eps) ** .5 105 | b = - bn.running_mean / (bn.running_var + bn.eps) ** .5 106 | conv2_bn_weight = conv2_bn_weight * k.unsqueeze(-1).unsqueeze(-1).unsqueeze(-1) 107 | conv2_bn_weight = conv2_bn_weight * bn.weight.unsqueeze(-1).unsqueeze(-1).unsqueeze(-1) 108 | conv2_bn_bias = self.conv2.bias * k + b 109 | conv2_bn_bias = conv2_bn_bias * bn.weight + bn.bias 110 | 111 | bn = self.conv_crossv_bn[0] 112 | k = 1 / (bn.running_var + bn.eps) ** .5 113 | b = - bn.running_mean / (bn.running_var + bn.eps) ** .5 114 | conv_crossv_bn_weight = conv_crossv_bn_weight * k.unsqueeze(-1).unsqueeze(-1).unsqueeze(-1) 115 | conv_crossv_bn_weight = conv_crossv_bn_weight * bn.weight.unsqueeze(-1).unsqueeze(-1).unsqueeze(-1) 116 | conv_crossv_bn_bias = self.conv_crossv.bias * k + b 117 | conv_crossv_bn_bias = conv_crossv_bn_bias * bn.weight + bn.bias 118 | 119 | bn = self.conv_crossh_bn[0] 120 | k = 1 / (bn.running_var + bn.eps) ** .5 121 | b = - bn.running_mean / (bn.running_var + bn.eps) ** .5 122 | conv_crossh_bn_weight = conv_crossh_bn_weight * k.unsqueeze(-1).unsqueeze(-1).unsqueeze(-1) 123 | conv_crossh_bn_weight = conv_crossh_bn_weight * bn.weight.unsqueeze(-1).unsqueeze(-1).unsqueeze(-1) 124 | conv_crossh_bn_bias = self.conv_crossh.bias * k + b 125 | conv_crossh_bn_bias = conv_crossh_bn_bias * bn.weight + bn.bias 126 | 127 | weight = torch.cat( 128 | [conv_weight, conv1_weight, conv2_weight, 129 | conv_crossh_weight, conv_crossv_weight, 130 | conv_bn_weight, conv1_bn_weight, conv2_bn_weight, 131 | conv_crossh_bn_weight, conv_crossv_bn_weight], 132 | 0 133 | ) 134 | #weight_compress = self.conv_out.weight.squeeze() 135 | weight_compress = (self.conv_out.weight + self.weight1).squeeze() 136 | weight = torch.matmul(weight_compress, weight.permute([2, 3, 0, 1])).permute([2, 3, 0, 1]) 137 | bias_ = torch.cat( 138 | [conv_bias, conv1_bias, conv2_bias, 139 | conv_crossh_bias, conv_crossv_bias, 140 | conv_bn_bias, conv1_bn_bias, conv2_bn_bias, 141 | conv_crossh_bn_bias, conv_crossv_bn_bias], 142 | 0 143 | ) 144 | bias = torch.matmul(weight_compress, bias_) 145 | if isinstance(self.conv_out.bias, torch.Tensor): 146 | bias = bias + self.conv_out.bias 147 | return weight, bias 148 | 149 | 150 | ############################################################################################################## 151 | class MBRConv3(nn.Module): 152 | def __init__(self, in_channels, out_channels, rep_scale=4): 153 | super(MBRConv3, self).__init__() 154 | 155 | self.in_channels = in_channels 156 | self.out_channels = out_channels 157 | self.rep_scale = rep_scale 158 | 159 | self.conv = nn.Conv2d(in_channels, out_channels * rep_scale, 3, 1, 1) 160 | self.conv_bn = nn.Sequential( 161 | nn.BatchNorm2d(out_channels * rep_scale) 162 | ) 163 | self.conv1 = nn.Conv2d(in_channels, out_channels * rep_scale, 1) 164 | self.conv1_bn = nn.Sequential( 165 | nn.BatchNorm2d(out_channels * rep_scale) 166 | ) 167 | self.conv_crossh = nn.Conv2d(in_channels, out_channels * rep_scale, (3, 1), 1, (1, 0)) 168 | self.conv_crossh_bn = nn.Sequential( 169 | nn.BatchNorm2d(out_channels * rep_scale) 170 | ) 171 | self.conv_crossv = nn.Conv2d(in_channels, out_channels * rep_scale, (1, 3), 1, (0, 1)) 172 | self.conv_crossv_bn = nn.Sequential( 173 | nn.BatchNorm2d(out_channels * rep_scale) 174 | ) 175 | self.conv_out = nn.Conv2d(out_channels * rep_scale * 8, out_channels, 1) 176 | self.conv_out.weight.requires_grad = False 177 | self.weight1 = nn.Parameter(torch.zeros_like(self.conv_out.weight)) 178 | nn.init.xavier_normal_(self.weight1) 179 | 180 | def forward(self, inp): 181 | x0 = self.conv(inp) 182 | x1 = self.conv1(inp) 183 | x2 = self.conv_crossh(inp) 184 | x3 = self.conv_crossv(inp) 185 | x = torch.cat( 186 | [ x0,x1,x2,x3, 187 | self.conv_bn(x0), 188 | self.conv1_bn(x1), 189 | self.conv_crossh_bn(x2), 190 | self.conv_crossv_bn(x3)], 191 | 1 192 | ) 193 | final_weight = self.conv_out.weight + self.weight1 194 | out = F.conv2d(x, final_weight, self.conv_out.bias) 195 | return out 196 | 197 | def slim(self): 198 | conv_weight = self.conv.weight 199 | conv_bias = self.conv.bias 200 | 201 | conv1_weight = self.conv1.weight 202 | conv1_bias = self.conv1.bias 203 | conv1_weight = F.pad(conv1_weight, (1, 1, 1, 1)) 204 | 205 | conv_crossh_weight = self.conv_crossh.weight 206 | conv_crossh_bias = self.conv_crossh.bias 207 | conv_crossh_weight = F.pad(conv_crossh_weight, (1, 1, 0, 0)) 208 | 209 | conv_crossv_weight = self.conv_crossv.weight 210 | conv_crossv_bias = self.conv_crossv.bias 211 | conv_crossv_weight = F.pad(conv_crossv_weight, (0, 0, 1, 1)) 212 | 213 | # conv_bn 214 | bn = self.conv_bn[0] 215 | k = 1 / torch.sqrt(bn.running_var + bn.eps) 216 | conv_bn_weight = self.conv.weight * k.unsqueeze(-1).unsqueeze(-1).unsqueeze(-1) 217 | conv_bn_weight = conv_bn_weight * bn.weight.unsqueeze(-1).unsqueeze(-1).unsqueeze(-1) 218 | conv_bn_bias = self.conv.bias * k + (-bn.running_mean * k) 219 | conv_bn_bias = conv_bn_bias * bn.weight + bn.bias 220 | 221 | # conv1_bn 222 | bn = self.conv1_bn[0] 223 | k = 1 / torch.sqrt(bn.running_var + bn.eps) 224 | conv1_bn_weight = self.conv1.weight * k.unsqueeze(-1).unsqueeze(-1).unsqueeze(-1) 225 | conv1_bn_weight = conv1_bn_weight * bn.weight.unsqueeze(-1).unsqueeze(-1).unsqueeze(-1) 226 | conv1_bn_weight = F.pad(conv1_bn_weight, (1, 1, 1, 1)) 227 | conv1_bn_bias = self.conv1.bias * k + (-bn.running_mean * k) 228 | conv1_bn_bias = conv1_bn_bias * bn.weight + bn.bias 229 | 230 | # conv_crossh_bn 231 | bn = self.conv_crossh_bn[0] 232 | k = 1 / torch.sqrt(bn.running_var + bn.eps) 233 | conv_crossh_bn_weight = self.conv_crossh.weight * k.unsqueeze(-1).unsqueeze(-1).unsqueeze(-1) 234 | conv_crossh_bn_weight = conv_crossh_bn_weight * bn.weight.unsqueeze(-1).unsqueeze(-1).unsqueeze(-1) 235 | conv_crossh_bn_weight = F.pad(conv_crossh_bn_weight, (1, 1, 0, 0)) 236 | conv_crossh_bn_bias = self.conv_crossh.bias * k + (-bn.running_mean * k) 237 | conv_crossh_bn_bias = conv_crossh_bn_bias * bn.weight + bn.bias 238 | 239 | # conv_crossv_bn 240 | bn = self.conv_crossv_bn[0] 241 | k = 1 / torch.sqrt(bn.running_var + bn.eps) 242 | conv_crossv_bn_weight = self.conv_crossv.weight * k.unsqueeze(-1).unsqueeze(-1).unsqueeze(-1) 243 | conv_crossv_bn_weight = conv_crossv_bn_weight * bn.weight.unsqueeze(-1).unsqueeze(-1).unsqueeze(-1) 244 | conv_crossv_bn_weight = F.pad(conv_crossv_bn_weight, (0, 0, 1, 1)) 245 | conv_crossv_bn_bias = self.conv_crossv.bias * k + (-bn.running_mean * k) 246 | conv_crossv_bn_bias = conv_crossv_bn_bias * bn.weight + bn.bias 247 | 248 | weight = torch.cat([ 249 | conv_weight, 250 | conv1_weight, 251 | conv_crossh_weight, 252 | conv_crossv_weight, 253 | conv_bn_weight, 254 | conv1_bn_weight, 255 | conv_crossh_bn_weight, 256 | conv_crossv_bn_weight 257 | ], dim=0) 258 | 259 | bias = torch.cat([ 260 | conv_bias, 261 | conv1_bias, 262 | conv_crossh_bias, 263 | conv_crossv_bias, 264 | conv_bn_bias, 265 | conv1_bn_bias, 266 | conv_crossh_bn_bias, 267 | conv_crossv_bn_bias 268 | ], dim=0) 269 | 270 | #weight_compress = self.conv_out.weight.squeeze() 271 | weight_compress = (self.conv_out.weight + self.weight1).squeeze() 272 | weight = torch.matmul(weight_compress, weight.view(weight.size(0), -1)) 273 | weight = weight.view(self.conv_out.out_channels, self.in_channels, 3, 3) 274 | 275 | bias = torch.matmul(weight_compress, bias.unsqueeze(-1)).squeeze(-1) 276 | if self.conv_out.bias is not None: 277 | bias += self.conv_out.bias 278 | 279 | return weight, bias 280 | 281 | ###################################################################################################### 282 | class MBRConv1(nn.Module): 283 | def __init__(self, in_channels, out_channels, rep_scale=4): 284 | super(MBRConv1, self).__init__() 285 | 286 | self.in_channels = in_channels 287 | self.out_channels = out_channels 288 | self.rep_scale = rep_scale 289 | 290 | self.conv = nn.Conv2d(in_channels, out_channels * rep_scale, 1) 291 | self.conv_bn = nn.Sequential( 292 | nn.BatchNorm2d(out_channels * rep_scale) 293 | ) 294 | self.conv_out = nn.Conv2d(out_channels * rep_scale * 2, out_channels, 1) 295 | self.conv_out.weight.requires_grad = False 296 | 297 | self.weight1 = nn.Parameter(torch.zeros_like(self.conv_out.weight)) 298 | nn.init.xavier_normal_(self.weight1) 299 | 300 | def forward(self, inp): 301 | x1 = self.conv(inp) 302 | x = torch.cat([x1, self.conv_bn(x1)], 1) 303 | final_weight = self.conv_out.weight + self.weight1 304 | out = F.conv2d(x, final_weight, self.conv_out.bias) 305 | return out 306 | 307 | def slim(self): 308 | conv_weight = self.conv.weight 309 | conv_bias = self.conv.bias 310 | 311 | bn = self.conv_bn[0] 312 | k = 1 / (bn.running_var + bn.eps) ** .5 313 | b = - bn.running_mean / (bn.running_var + bn.eps) ** .5 314 | conv_bn_weight = self.conv.weight * k.unsqueeze(-1).unsqueeze(-1).unsqueeze(-1) 315 | conv_bn_weight = conv_bn_weight * bn.weight.unsqueeze(-1).unsqueeze(-1).unsqueeze(-1) 316 | conv_bn_bias = self.conv.bias * k + b 317 | conv_bn_bias = conv_bn_bias * bn.weight + bn.bias 318 | 319 | weight = torch.cat([conv_weight, conv_bn_weight], 0) 320 | #weight_compress = self.conv_out.weight.squeeze() 321 | weight_compress = (self.conv_out.weight + self.weight1).squeeze() 322 | weight = torch.matmul(weight_compress, weight.permute([2, 3, 0, 1])).permute([2, 3, 0, 1]) 323 | 324 | bias = torch.cat([conv_bias, conv_bn_bias], 0) 325 | bias = torch.matmul(weight_compress, bias) 326 | 327 | if isinstance(self.conv_out.bias, torch.Tensor): 328 | bias = bias + self.conv_out.bias 329 | return weight, bias 330 | 331 | class FST(nn.Module): 332 | def __init__(self, block1, channels): 333 | super(FST, self).__init__() 334 | self.block1 = block1 335 | self.weight1 = nn.Parameter(torch.randn(1)) 336 | self.weight2 = nn.Parameter(torch.randn(1)) 337 | self.bias = nn.Parameter(torch.randn((1, channels, 1, 1))) 338 | 339 | def forward(self, x): 340 | x1 = self.block1(x) 341 | weighted_block1 = self.weight1 * x1 342 | weighted_block2 = self.weight2 * x1 343 | return weighted_block1 * weighted_block2 + self.bias 344 | 345 | class FSTS(nn.Module): 346 | def __init__(self, block1, channels): 347 | super(FSTS, self).__init__() 348 | self.block1 = block1 349 | self.weight1 = nn.Parameter(torch.randn(1)) 350 | self.weight2 = nn.Parameter(torch.randn(1)) 351 | self.bias = nn.Parameter(torch.randn((1, channels, 1, 1))) 352 | 353 | def forward(self, x): 354 | x1 = self.block1(x) 355 | weighted_block1 = self.weight1 * x1 356 | weighted_block2 = self.weight2 * x1 357 | return weighted_block1 * weighted_block2 + self.bias 358 | ################################################################################## 359 | class DropBlock(nn.Module): 360 | def __init__(self, block_size, p=0.5): 361 | super(DropBlock, self).__init__() 362 | self.block_size = block_size 363 | self.p = p / block_size / block_size 364 | 365 | def forward(self, x): 366 | mask = 1 - (torch.rand_like(x[:, :1]) >= self.p).float() 367 | mask = nn.functional.max_pool2d(mask, self.block_size, 1, self.block_size // 2) 368 | return x * (1 - mask) 369 | --------------------------------------------------------------------------------