diff --git a/model_zoo/official/cv/faster_rcnn/README.md b/model_zoo/official/cv/faster_rcnn/README.md index 9d58387d90b..185c67a51c2 100644 --- a/model_zoo/official/cv/faster_rcnn/README.md +++ b/model_zoo/official/cv/faster_rcnn/README.md @@ -70,7 +70,7 @@ Dataset used: [COCO2017]() pip install mmcv==0.2.14 ``` - And change the COCO_ROOT and other settings you need in `config.py`. The directory structure is as follows: + And change the COCO_ROOT and other settings you need in `config_50.yaml、config_101.yaml or config_152.yaml`. The directory structure is as follows: ```path . @@ -90,7 +90,7 @@ Dataset used: [COCO2017]() train2017/0000001.jpg 0,259,401,459,7 35,28,324,201,2 0,30,59,80,2 ``` - Each row is an image annotation which split by space, the first column is a relative path of image, the others are box and class information of the format [xmin,ymin,xmax,ymax,class]. We read image from an image path joined by the `IMAGE_DIR`(dataset directory) and the relative path in `ANNO_PATH`(the TXT file path), `IMAGE_DIR` and `ANNO_PATH` are setting in `config.py`. + Each row is an image annotation which split by space, the first column is a relative path of image, the others are box and class information of the format [xmin,ymin,xmax,ymax,class]. We read image from an image path joined by the `IMAGE_DIR`(dataset directory) and the relative path in `ANNO_PATH`(the TXT file path), `IMAGE_DIR` and `ANNO_PATH` are setting in `config_50.yaml、config_101.yaml or config_152.yaml`. # Quick Start @@ -110,13 +110,13 @@ Note: python convert_checkpoint.py --ckpt_file=[BACKBONE_MODEL] # standalone training -sh run_standalone_train_ascend.sh [PRETRAINED_MODEL] +sh run_standalone_train_ascend.sh [PRETRAINED_MODEL] [BACKBONE] # distributed training -sh run_distribute_train_ascend.sh [RANK_TABLE_FILE] [PRETRAINED_MODEL] +sh run_distribute_train_ascend.sh [RANK_TABLE_FILE] [PRETRAINED_MODEL] [BACKBONE] # eval -sh run_eval_ascend.sh [VALIDATION_JSON_FILE] [CHECKPOINT_PATH] +sh run_eval_ascend.sh [VALIDATION_JSON_FILE] [CHECKPOINT_PATH] [BACKBONE] # inference sh run_infer_310.sh [AIR_PATH] [DATA_PATH] [ANN_FILE_PATH] @@ -130,13 +130,13 @@ sh run_infer_310.sh [AIR_PATH] [DATA_PATH] [ANN_FILE_PATH] python convert_checkpoint.py --ckpt_file=[BACKBONE_MODEL] # standalone training -sh run_standalone_train_gpu.sh [PRETRAINED_MODEL] +sh run_standalone_train_gpu.sh [PRETRAINED_MODEL] [BACKBONE] # distributed training -sh run_distribute_train_gpu.sh [DEVICE_NUM] [PRETRAINED_MODEL] +sh run_distribute_train_gpu.sh [DEVICE_NUM] [PRETRAINED_MODEL] [BACKBONE] # eval -sh run_eval_gpu.sh [VALIDATION_JSON_FILE] [CHECKPOINT_PATH] +sh run_eval_gpu.sh [VALIDATION_JSON_FILE] [CHECKPOINT_PATH] [BACKBONE] ``` @@ -160,17 +160,17 @@ bash scripts/docker_start.sh fasterrcnn:20.1.0 [DATA_DIR] [MODEL_DIR] ```shell # standalone training -sh run_standalone_train_ascend.sh [PRETRAINED_MODEL] +sh run_standalone_train_ascend.sh [PRETRAINED_MODEL] [BACKBONE] # distributed training -sh run_distribute_train_ascend.sh [RANK_TABLE_FILE] [PRETRAINED_MODEL] +sh run_distribute_train_ascend.sh [RANK_TABLE_FILE] [PRETRAINED_MODEL] [BACKBONE] ``` 4. Eval ```shell # eval -sh run_eval_ascend.sh [VALIDATION_JSON_FILE] [CHECKPOINT_PATH] +sh run_eval_ascend.sh [VALIDATION_JSON_FILE] [CHECKPOINT_PATH] [BACKBONE] ``` 5. Inference @@ -203,14 +203,19 @@ sh run_infer_310.sh [MINDIR_PATH] [DATA_PATH] [ANN_FILE] [DEVICE_ID] ├─anchor_generator.py // anchor generator ├─bbox_assign_sample.py // first stage sampler ├─bbox_assign_sample_stage2.py // second stage sampler - ├─faster_rcnn_r50.py // fasterrcnn network + ├─faster_rcnn_resnet.py // fasterrcnn network + ├─faster_rcnn_resnet50v1.py //fasterrcnn network for ResNet50v1.0 ├─fpn_neck.py //feature pyramid network ├─proposal_generator.py // proposal generator ├─rcnn.py // rcnn network - ├─resnet50.py // backbone network + ├─resnet.py // backbone network + ├─resnet50v1.py // backbone network for ResNet50v1.0 ├─roi_align.py // roi align network └─rpn.py // region proposal network - ├─config.py // total config + ├─config.py // config for yaml parsing + ├─config_50.yaml // config for ResNet50 + ├─config_101.yaml // config for ResNet101 + ├─config_152.yaml // config for ResNet152 ├─dataset.py // create dataset and process dataset ├─lr_schedule.py // learning ratio generator ├─network_define.py // network define for fasterrcnn @@ -239,10 +244,10 @@ sh run_distribute_train_ascend.sh [RANK_TABLE_FILE] [PRETRAINED_MODEL] ```shell # standalone training on gpu -sh run_standalone_train_gpu.sh [PRETRAINED_MODEL] +sh run_standalone_train_gpu.sh [PRETRAINED_MODEL] [BACKBONE] # distributed training on gpu -sh run_distribute_train_gpu.sh [DEVICE_NUM] [PRETRAINED_MODEL] +sh run_distribute_train_gpu.sh [DEVICE_NUM] [PRETRAINED_MODEL] [BACKBONE] ``` Notes: @@ -274,7 +279,7 @@ Notes: load_param_into_net(net, param_dict) ``` -3. The original dataset path needs to be in the config.py,you can select "coco_root" or "image_dir". +3. The original dataset path needs to be in the config_50.yaml、config_101.yaml、config_152.yaml,you can select "coco_root" or "image_dir". ### Result @@ -299,14 +304,14 @@ epoch: 12 step: 7393, rpn_loss: 0.00691, rcnn_loss: 0.10168, rpn_cls_loss: 0.005 ```shell # eval on ascend -sh run_eval_ascend.sh [VALIDATION_JSON_FILE] [CHECKPOINT_PATH] +sh run_eval_ascend.sh [VALIDATION_JSON_FILE] [CHECKPOINT_PATH] [BACKBONE] ``` #### on GPU ```shell # eval on GPU -sh run_eval_gpu.sh [VALIDATION_JSON_FILE] [CHECKPOINT_PATH] +sh run_eval_gpu.sh [VALIDATION_JSON_FILE] [CHECKPOINT_PATH] [BACKBONE] ``` > checkpoint can be produced in training process. @@ -335,7 +340,7 @@ Eval result will be stored in the example path, whose folder name is "eval". Und ## Model Export ```shell -python export.py --ckpt_file [CKPT_PATH] --device_target [DEVICE_TARGET] --file_format[EXPORT_FORMAT] +python export.py --ckpt_file [CKPT_PATH] --device_target [DEVICE_TARGET] --file_format[EXPORT_FORMAT] --backbone [BACKBONE] ``` `EXPORT_FORMAT` should be in ["AIR", "MINDIR"] diff --git a/model_zoo/official/cv/faster_rcnn/README_CN.md b/model_zoo/official/cv/faster_rcnn/README_CN.md index e352d51fc87..d0ad738fa2a 100644 --- a/model_zoo/official/cv/faster_rcnn/README_CN.md +++ b/model_zoo/official/cv/faster_rcnn/README_CN.md @@ -71,7 +71,7 @@ Faster R-CNN是一个两阶段目标检测网络，该网络采用RPN，可以 pip install mmcv==0.2.14 ``` - 在`config.py`中更改COCO_ROOT和其他您需要的设置。目录结构如下： + 根据模型运行需要，对应地在`config_50.yaml、config_101.yaml或config_152.yaml`中更改COCO_ROOT和其他需要的设置。目录结构如下： ```path . @@ -91,7 +91,7 @@ Faster R-CNN是一个两阶段目标检测网络，该网络采用RPN，可以 train2017/0000001.jpg 0,259,401,459,7 35,28,324,201,2 0,30,59,80,2 ``` - 每行是按空间分割的图像标注，第一列是图像的相对路径，其余为[xmin,ymin,xmax,ymax,class]格式的框和类信息。从`IMAGE_DIR`（数据集目录）图像路径以及`ANNO_PATH`（TXT文件路径）的相对路径中读取图像。`IMAGE_DIR`和`ANNO_PATH`可在`config.py`中设置。 + 每行是按空间分割的图像标注，第一列是图像的相对路径，其余为[xmin,ymin,xmax,ymax,class]格式的框和类信息。从`IMAGE_DIR`（数据集目录）图像路径以及`ANNO_PATH`（TXT文件路径）的相对路径中读取图像。`IMAGE_DIR`和`ANNO_PATH`可在`config_50.yaml、config_101.yaml或config_152.yaml`中设置。 # 快速入门 @@ -111,13 +111,13 @@ Faster R-CNN是一个两阶段目标检测网络，该网络采用RPN，可以 python convert_checkpoint.py --ckpt_file=[BACKBONE_MODEL] # 单机训练 -sh run_standalone_train_ascend.sh [PRETRAINED_MODEL] +sh run_standalone_train_ascend.sh [PRETRAINED_MODEL] [BACKBONE] # 分布式训练 -sh run_distribute_train_ascend.sh [RANK_TABLE_FILE] [PRETRAINED_MODEL] +sh run_distribute_train_ascend.sh [RANK_TABLE_FILE] [PRETRAINED_MODEL] [BACKBONE] # 评估 -sh run_eval_ascend.sh [VALIDATION_JSON_FILE] [CHECKPOINT_PATH] +sh run_eval_ascend.sh [VALIDATION_JSON_FILE] [CHECKPOINT_PATH] [BACKBONE] #推理 sh run_infer_310.sh [MINDIR_PATH] [DATA_PATH] [ANN_FILE] [DEVICE_ID] @@ -131,13 +131,13 @@ sh run_infer_310.sh [MINDIR_PATH] [DATA_PATH] [ANN_FILE] [DEVICE_ID] python convert_checkpoint.py --ckpt_file=[BACKBONE_MODEL] # 单机训练 -sh run_standalone_train_gpu.sh [PRETRAINED_MODEL] +sh run_standalone_train_gpu.sh [PRETRAINED_MODEL] [BACKBONE] # 分布式训练 -sh run_distribute_train_gpu.sh [DEVICE_NUM] [PRETRAINED_MODEL] +sh run_distribute_train_gpu.sh [DEVICE_NUM] [PRETRAINED_MODEL] [BACKBONE] # 评估 -sh run_eval_gpu.sh [VALIDATION_JSON_FILE] [CHECKPOINT_PATH] +sh run_eval_gpu.sh [VALIDATION_JSON_FILE] [CHECKPOINT_PATH] [BACKBONE] ``` @@ -161,17 +161,17 @@ bash scripts/docker_start.sh fasterrcnn:20.1.0 [DATA_DIR] [MODEL_DIR] ```shell # 单机训练 -sh run_standalone_train_ascend.sh [PRETRAINED_MODEL] +sh run_standalone_train_ascend.sh [PRETRAINED_MODEL] [BACKBONE] # 分布式训练 -sh run_distribute_train_ascend.sh [RANK_TABLE_FILE] [PRETRAINED_MODEL] +sh run_distribute_train_ascend.sh [RANK_TABLE_FILE] [PRETRAINED_MODEL] [BACKBONE] ``` 4. 评估 ```shell # 评估 -sh run_eval_ascend.sh [VALIDATION_JSON_FILE] [CHECKPOINT_PATH] +sh run_eval_ascend.sh [VALIDATION_JSON_FILE] [CHECKPOINT_PATH] [BACKBONE] ``` 5. 推理 @@ -204,14 +204,19 @@ sh run_infer_310.sh [AIR_PATH] [DATA_PATH] [ANN_FILE_PATH] [DEVICE_ID] ├─anchor_generator.py // 锚点生成器 ├─bbox_assign_sample.py // 第一阶段采样器 ├─bbox_assign_sample_stage2.py // 第二阶段采样器 - ├─faster_rcnn_r50.py // Faster R-CNN网络 + ├─faster_rcnn_resnet.py // Faster R-CNN网络 + ├─faster_rcnn_resnet50v1.py //以Resnet50v1.0作为backbone的Faster R-CNN网络 ├─fpn_neck.py // 特征金字塔网络 ├─proposal_generator.py // 候选生成器 ├─rcnn.py // R-CNN网络 - ├─resnet50.py // 骨干网络 + ├─resnet.py // 骨干网络 + ├─resnet50v1.py // Resnet50v1.0骨干网络 ├─roi_align.py // ROI对齐网络 └─rpn.py // 区域候选网络 - ├─config.py // 总配置 + ├─config.py // 读取yaml配置的config类 + ├─config_50.yaml // Resnet50相关配置 + ├─config_101.yaml // Resnet101相关配置 + ├─config_152.yaml // Resnet152相关配置 ├─dataset.py // 创建并处理数据集 ├─lr_schedule.py // 学习率生成器 ├─network_define.py // Faster R-CNN网络定义 @@ -230,20 +235,20 @@ sh run_infer_310.sh [AIR_PATH] [DATA_PATH] [ANN_FILE_PATH] [DEVICE_ID] ```shell # Ascend单机训练 -sh run_standalone_train_ascend.sh [PRETRAINED_MODEL] +sh run_standalone_train_ascend.sh [PRETRAINED_MODEL] [BACKBONE] # Ascend分布式训练 -sh run_distribute_train_ascend.sh [RANK_TABLE_FILE] [PRETRAINED_MODEL] +sh run_distribute_train_ascend.sh [RANK_TABLE_FILE] [PRETRAINED_MODEL] [BACKBONE] ``` #### 在GPU上运行 ```shell # GPU单机训练 -sh run_standalone_train_gpu.sh [PRETRAINED_MODEL] +sh run_standalone_train_gpu.sh [PRETRAINED_MODEL] [BACKBONE] # GPU分布式训练 -sh run_distribute_train_gpu.sh [DEVICE_NUM] [PRETRAINED_MODEL] +sh run_distribute_train_gpu.sh [DEVICE_NUM] [PRETRAINED_MODEL] [BACKBONE] ``` Notes: @@ -275,7 +280,7 @@ Notes: load_param_into_net(net, param_dict) ``` -3. config.py中包含原数据集路径，可以选择“coco_root”或“image_dir”。 +3. config_50.yaml、config_101.yaml、config_152.yaml中包含原数据集路径，可以选择“coco_root”或“image_dir”。 ### 结果 @@ -300,14 +305,14 @@ epoch: 12 step: 7393, rpn_loss: 0.00691, rcnn_loss: 0.10168, rpn_cls_loss: 0.005 ```shell # Ascend评估 -sh run_eval_ascend.sh [VALIDATION_JSON_FILE] [CHECKPOINT_PATH] +sh run_eval_ascend.sh [VALIDATION_JSON_FILE] [CHECKPOINT_PATH] [BACKBONE] ``` #### 在GPU上运行 ```shell # GPU评估 -sh run_eval_gpu.sh [VALIDATION_JSON_FILE] [CHECKPOINT_PATH] +sh run_eval_gpu.sh [VALIDATION_JSON_FILE] [CHECKPOINT_PATH] [BACKBONE] ``` > 在训练过程中生成检查点。 @@ -336,7 +341,7 @@ sh run_eval_gpu.sh [VALIDATION_JSON_FILE] [CHECKPOINT_PATH] ## 模型导出 ```shell -python export.py --ckpt_file [CKPT_PATH] --device_target [DEVICE_TARGET] --file_format[EXPORT_FORMAT] +python export.py --ckpt_file [CKPT_PATH] --device_target [DEVICE_TARGET] --file_format[EXPORT_FORMAT] --backbone [BACKBONE] ``` `EXPORT_FORMAT` 可选 ["AIR", "MINDIR"] diff --git a/model_zoo/official/cv/faster_rcnn/ascend310_infer/inc/utils.h b/model_zoo/official/cv/faster_rcnn/ascend310_infer/inc/utils.h index efebe03a8c1..88b18882a5d 100644 --- a/model_zoo/official/cv/faster_rcnn/ascend310_infer/inc/utils.h +++ b/model_zoo/official/cv/faster_rcnn/ascend310_infer/inc/utils.h @@ -1,5 +1,5 @@ /** - * Copyright 2021 Huawei Technologies Co., Ltd + * Copyright 2020 Huawei Technologies Co., Ltd * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. diff --git a/model_zoo/official/cv/faster_rcnn/eval.py b/model_zoo/official/cv/faster_rcnn/eval.py index 7a778a34c48..410c8187e32 100644 --- a/model_zoo/official/cv/faster_rcnn/eval.py +++ b/model_zoo/official/cv/faster_rcnn/eval.py @@ -24,10 +24,9 @@ from mindspore import context from mindspore.train.serialization import load_checkpoint, load_param_into_net from mindspore.common import set_seed, Parameter -from src.FasterRcnn.faster_rcnn_r50 import Faster_Rcnn_Resnet50 -from src.config import config from src.dataset import data_to_mindrecord_byte_image, create_fasterrcnn_dataset from src.util import coco_eval, bbox2result_1image, results2json +import src.config as cfg set_seed(1) @@ -38,14 +37,29 @@ parser.add_argument("--checkpoint_path", type=str, required=True, help="Checkpoi parser.add_argument("--device_target", type=str, default="Ascend", help="device where the code will be implemented, default is Ascend") parser.add_argument("--device_id", type=int, default=0, help="Device id, default is 0.") +parser.add_argument("--backbone", type=str, required=True, \ + help="backbone network name, options:resnet_v1_50, resnet_v1.5_50, resnet_v1_101, resnet_v1_152") args_opt = parser.parse_args() context.set_context(mode=context.GRAPH_MODE, device_target=args_opt.device_target, device_id=args_opt.device_id) +if args_opt.backbone in ("resnet_v1.5_50", "resnet_v1_101", "resnet_v1_152"): + from src.FasterRcnn.faster_rcnn_resnet import Faster_Rcnn_Resnet + if args_opt.backbone == "resnet_v1.5_50": + config = cfg.get_config("./src/config_50.yaml") + elif args_opt.backbone == "resnet_v1_101": + config = cfg.get_config("./src/config_101.yaml") + elif args_opt.backbone == "resnet_v1_152": + config = cfg.get_config("./src/config_152.yaml") + +elif args_opt.backbone == "resnet_v1_50": + config = cfg.get_config("./src/config_50.yaml") + from src.FasterRcnn.faster_rcnn_resnet50v1 import Faster_Rcnn_Resnet + def fasterrcnn_eval(dataset_path, ckpt_path, ann_file): """FasterRcnn evaluation.""" - ds = create_fasterrcnn_dataset(dataset_path, batch_size=config.test_batch_size, is_training=False) - net = Faster_Rcnn_Resnet50(config) + ds = create_fasterrcnn_dataset(config, dataset_path, batch_size=config.test_batch_size, is_training=False) + net = Faster_Rcnn_Resnet(config) param_dict = load_checkpoint(ckpt_path) if args_opt.device_target == "GPU": for key, value in param_dict.items(): @@ -123,14 +137,14 @@ if __name__ == '__main__': if args_opt.dataset == "coco": if os.path.isdir(config.coco_root): print("Create Mindrecord. It may take some time.") - data_to_mindrecord_byte_image("coco", False, prefix, file_num=1) + data_to_mindrecord_byte_image(config, "coco", False, prefix, file_num=1) print("Create Mindrecord Done, at {}".format(mindrecord_dir)) else: print("coco_root not exits.") else: if os.path.isdir(config.IMAGE_DIR) and os.path.exists(config.ANNO_PATH): print("Create Mindrecord. It may take some time.") - data_to_mindrecord_byte_image("other", False, prefix, file_num=1) + data_to_mindrecord_byte_image(config, "other", False, prefix, file_num=1) print("Create Mindrecord Done, at {}".format(mindrecord_dir)) else: print("IMAGE_DIR or ANNO_PATH not exits.") diff --git a/model_zoo/official/cv/faster_rcnn/export.py b/model_zoo/official/cv/faster_rcnn/export.py index 4da6cf81695..4ab062007f2 100644 --- a/model_zoo/official/cv/faster_rcnn/export.py +++ b/model_zoo/official/cv/faster_rcnn/export.py @@ -1,4 +1,4 @@ -# Copyright 2020 Huawei Technologies Co., Ltd +# Copyright 2020-2021 Huawei Technologies Co., Ltd # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. @@ -16,12 +16,9 @@ import argparse import numpy as np -import mindspore as ms import mindspore.common.dtype as mstype from mindspore import Tensor, load_checkpoint, load_param_into_net, export, context - -from src.FasterRcnn.faster_rcnn_r50 import FasterRcnn_Infer -from src.config import config +import src.config as cfg parser = argparse.ArgumentParser(description='fasterrcnn_export') parser.add_argument("--device_id", type=int, default=0, help="Device id") @@ -30,12 +27,27 @@ parser.add_argument("--file_format", type=str, choices=["AIR", "ONNX", "MINDIR"] parser.add_argument("--device_target", type=str, choices=["Ascend", "GPU", "CPU"], default="Ascend", help="device target") parser.add_argument('--ckpt_file', type=str, default='', help='fasterrcnn ckpt file.') +parser.add_argument("--backbone", type=str, required=True, \ + help="backbone network name, options:resnet_v1_50, resnet_v1.5_50, resnet_v1_101, resnet_v1_152") args = parser.parse_args() context.set_context(mode=context.GRAPH_MODE, device_target=args.device_target) if args.device_target == "Ascend": context.set_context(device_id=args.device_id) +if args.backbone in ("resnet_v1.5_50", "resnet_v1_101", "resnet_v1_152"): + from src.FasterRcnn.faster_rcnn_resnet import FasterRcnn_Infer + if args.backbone == "resnet_v1.5_50": + config = cfg.get_config("./src/config_50.yaml") + elif args.backbone == "resnet_v1_101": + config = cfg.get_config("./src/config_101.yaml") + elif args.backbone == "resnet_v1_152": + config = cfg.get_config("./src/config_152.yaml") + +elif args.backbone == "resnet_v1_50": + config = cfg.get_config("./src/config_50.yaml") + from src.FasterRcnn.faster_rcnn_resnet50v1 import FasterRcnn_Infer + if __name__ == '__main__': net = FasterRcnn_Infer(config=config) @@ -51,7 +63,7 @@ if __name__ == '__main__': if device_type == "Ascend": net.to_float(mstype.float16) - img = Tensor(np.zeros([config.test_batch_size, 3, config.img_height, config.img_width]), ms.float32) - img_metas = Tensor(np.random.uniform(0.0, 1.0, size=[config.test_batch_size, 4]), ms.float32) + img = Tensor(np.zeros([config.test_batch_size, 3, config.img_height, config.img_width]), mstype.float32) + img_metas = Tensor(np.random.uniform(0.0, 1.0, size=[config.test_batch_size, 4]), mstype.float32) export(net, img, img_metas, file_name=args.file_name, file_format=args.file_format) diff --git a/model_zoo/official/cv/faster_rcnn/mindspore_hub_conf.py b/model_zoo/official/cv/faster_rcnn/mindspore_hub_conf.py index 50f1cc517ab..63abac8fdb1 100644 --- a/model_zoo/official/cv/faster_rcnn/mindspore_hub_conf.py +++ b/model_zoo/official/cv/faster_rcnn/mindspore_hub_conf.py @@ -1,4 +1,4 @@ -# Copyright 2020 Huawei Technologies Co., Ltd +# Copyright 2020-2021 Huawei Technologies Co., Ltd # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. @@ -13,10 +13,28 @@ # limitations under the License. # ============================================================================ """hub config.""" -from src.FasterRcnn.faster_rcnn_r50 import Faster_Rcnn_Resnet50 -from src.config import config +import argparse +import src.config as cfg + +parser = argparse.ArgumentParser(description="FasterRcnn") +parser.add_argument("--backbone", type=str, required=True, \ + help="backbone network name, options:resnet_v1_50, resnet_v1.5_50, resnet_v1_101, resnet_v1_152") +args_opt = parser.parse_args() + +if args_opt.backbone in ("resnet_v1.5_50", "resnet_v1_101", "resnet_v1_152"): + from src.FasterRcnn.faster_rcnn_resnet import Faster_Rcnn_Resnet + if args_opt.backbone == "resnet_v1.5_50": + config = cfg.get_config("./src/config_50.yaml") + elif args_opt.backbone == "resnet_v1_101": + config = cfg.get_config("./src/config_101.yaml") + elif args_opt.backbone == "resnet_v1_152": + config = cfg.get_config("./src/config_152.yaml") + +elif args_opt.backbone == "resnet_v1_50": + config = cfg.get_config("./src/config_50.yaml") + from src.FasterRcnn.faster_rcnn_resnet50v1 import Faster_Rcnn_Resnet def create_network(name, *args, **kwargs): if name == "faster_rcnn": - return Faster_Rcnn_Resnet50(config=config) + return Faster_Rcnn_Resnet(config=config) raise NotImplementedError(f"{name} is not implemented in the repo") diff --git a/model_zoo/official/cv/faster_rcnn/postprocess.py b/model_zoo/official/cv/faster_rcnn/postprocess.py index b866c3c172b..ffe051ebcec 100644 --- a/model_zoo/official/cv/faster_rcnn/postprocess.py +++ b/model_zoo/official/cv/faster_rcnn/postprocess.py @@ -1,4 +1,4 @@ -# Copyright 2020 Huawei Technologies Co., Ltd +# Copyright 2020-2021 Huawei Technologies Co., Ltd # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. @@ -18,8 +18,8 @@ import argparse import numpy as np from pycocotools.coco import COCO -from src.config import config from src.util import coco_eval, bbox2result_1image, results2json +import src.config as cfg dst_width = 1280 dst_height = 768 @@ -28,6 +28,7 @@ parser = argparse.ArgumentParser(description="FasterRcnn inference") parser.add_argument("--ann_file", type=str, required=True, help="ann file.") parser.add_argument("--result_path", type=str, required=True, help="result file path.") args = parser.parse_args() +config = cfg.get_config("./src/config_50.yaml") def get_eval_result(ann_file, result_path): """ get evaluation result of faster rcnn""" diff --git a/model_zoo/official/cv/faster_rcnn/scripts/run_distribute_train_ascend.sh b/model_zoo/official/cv/faster_rcnn/scripts/run_distribute_train_ascend.sh index 1cb0898bc20..92652e43056 100755 --- a/model_zoo/official/cv/faster_rcnn/scripts/run_distribute_train_ascend.sh +++ b/model_zoo/official/cv/faster_rcnn/scripts/run_distribute_train_ascend.sh @@ -1,5 +1,5 @@ #!/bin/bash -# Copyright 2020 Huawei Technologies Co., Ltd +# Copyright 2020-2021 Huawei Technologies Co., Ltd # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. @@ -14,9 +14,15 @@ # limitations under the License. # ============================================================================ -if [ $# -ne 2 ] +if [ $# -ne 3 ] then - echo "Usage: sh run_distribute_train_ascend.sh [RANK_TABLE_FILE] [PRETRAINED_PATH]" + echo "Usage: sh run_distribute_train_ascend.sh [RANK_TABLE_FILE] [PRETRAINED_PATH] [BACKBONE]" +exit 1 +fi + +if [ $3 != "resnet_v1_50" ] && [ $3 != "resnet_v1.5_50" ] && [ $3 != "resnet_v1_101" ] && [ $3 != "resnet_v1_152" ] +then + echo "error: the selected backbone must be resnet_v1_50, resnet_v1.5_50, resnet_v1_101, resnet_v1_152" exit 1 fi @@ -63,6 +69,6 @@ do cd ./train_parallel$i || exit echo "start training for rank $RANK_ID, device $DEVICE_ID" env > env.log - python train.py --device_id=$i --rank_id=$i --run_distribute=True --device_num=$DEVICE_NUM --pre_trained=$PATH2 &> log & + python train.py --device_id=$i --rank_id=$i --run_distribute=True --device_num=$DEVICE_NUM --pre_trained=$PATH2 --backbone=$3 &> log & cd .. -done +done \ No newline at end of file diff --git a/model_zoo/official/cv/faster_rcnn/scripts/run_distribute_train_gpu.sh b/model_zoo/official/cv/faster_rcnn/scripts/run_distribute_train_gpu.sh index 515073e761c..4df8beb49a4 100755 --- a/model_zoo/official/cv/faster_rcnn/scripts/run_distribute_train_gpu.sh +++ b/model_zoo/official/cv/faster_rcnn/scripts/run_distribute_train_gpu.sh @@ -1,5 +1,5 @@ #!/bin/bash -# Copyright 2021 Huawei Technologies Co., Ltd +# Copyright 2020-2021 Huawei Technologies Co., Ltd # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. @@ -16,14 +16,20 @@ echo "==============================================================================================================" echo "Please run the script as: " -echo "sh run_distribute_train_gpu.sh DEVICE_NUM PRETRAINED_PATH" -echo "for example: sh run_distribute_train_gpu.sh 8 /path/pretrain.ckpt" +echo "sh run_distribute_train_gpu.sh DEVICE_NUM PRETRAINED_PATH BACKBONE" +echo "for example: sh run_distribute_train_gpu.sh 8 /path/pretrain.ckpt resnet_v1_50" echo "It is better to use absolute path." echo "==============================================================================================================" -if [ $# != 2 ] +if [ $# != 3 ] then - echo "Usage: sh run_distribute_train_gpu.sh [DEVICE_NUM] [PRETRAINED_PATH]" + echo "Usage: sh run_distribute_train_gpu.sh [DEVICE_NUM] [PRETRAINED_PATH] [BACKBONE]" +exit 1 +fi + +if [ $3 != "resnet_v1_50" ] && [ $3 != "resnet_v1.5_50" ] && [ $3 != "resnet_v1_101" ] && [ $3 != "resnet_v1_152" ] +then + echo "error: the selected backbone must be resnet_v1_50, resnet_v1.5_50, resnet_v1_101, resnet_v1_152" exit 1 fi @@ -41,4 +47,5 @@ mpirun -n $RANK_SIZE \ --run_distribute=True \ --device_target="GPU" \ --device_num=$RANK_SIZE \ - --pre_trained=$PRETRAINED_PATH > log 2>&1 & + --pre_trained=$PRETRAINED_PATH \ + --backbone=$3 > log 2>&1 & \ No newline at end of file diff --git a/model_zoo/official/cv/faster_rcnn/scripts/run_eval_ascend.sh b/model_zoo/official/cv/faster_rcnn/scripts/run_eval_ascend.sh index 606331ea2cd..27c86ca1a1b 100755 --- a/model_zoo/official/cv/faster_rcnn/scripts/run_eval_ascend.sh +++ b/model_zoo/official/cv/faster_rcnn/scripts/run_eval_ascend.sh @@ -1,5 +1,5 @@ #!/bin/bash -# Copyright 2020 Huawei Technologies Co., Ltd +# Copyright 2020-2021 Huawei Technologies Co., Ltd # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. @@ -14,9 +14,15 @@ # limitations under the License. # ============================================================================ -if [ $# != 2 ] +if [ $# != 3 ] then - echo "Usage: sh run_eval_ascend.sh [VALIDATION_JSON_FILE] [CHECKPOINT_PATH]" + echo "Usage: sh run_eval_ascend.sh [VALIDATION_JSON_FILE] [CHECKPOINT_PATH] [BACKBONE]" +exit 1 +fi + +if [ $3 != "resnet_v1_50" ] && [ $3 != "resnet_v1.5_50" ] && [ $3 != "resnet_v1_101" ] && [ $3 != "resnet_v1_152" ] +then + echo "error: the selected backbone must be resnet_v1_50, resnet_v1.5_50, resnet_v1_101, resnet_v1_152" exit 1 fi @@ -61,5 +67,5 @@ cp -r ../src ./eval cd ./eval || exit env > env.log echo "start eval for device $DEVICE_ID" -python eval.py --device_id=$DEVICE_ID --ann_file=$PATH1 --checkpoint_path=$PATH2 &> log & -cd .. +python eval.py --device_id=$DEVICE_ID --ann_file=$PATH1 --checkpoint_path=$PATH2 --backbone=$3 &> log & +cd .. \ No newline at end of file diff --git a/model_zoo/official/cv/faster_rcnn/scripts/run_eval_gpu.sh b/model_zoo/official/cv/faster_rcnn/scripts/run_eval_gpu.sh index dd6a70e5e5b..5f091980b01 100755 --- a/model_zoo/official/cv/faster_rcnn/scripts/run_eval_gpu.sh +++ b/model_zoo/official/cv/faster_rcnn/scripts/run_eval_gpu.sh @@ -1,5 +1,5 @@ #!/bin/bash -# Copyright 2021 Huawei Technologies Co., Ltd +# Copyright 2020-2021 Huawei Technologies Co., Ltd # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. @@ -14,9 +14,15 @@ # limitations under the License. # ============================================================================ -if [ $# != 2 ] +if [ $# != 3 ] then - echo "Usage: sh run_eval_gpu.sh [VALIDATION_JSON_FILE] [CHECKPOINT_PATH]" + echo "Usage: sh run_eval_gpu.sh [VALIDATION_JSON_FILE] [CHECKPOINT_PATH] [BACKBONE]" +exit 1 +fi + +if [ $3 != "resnet_v1_50" ] && [ $3 != "resnet_v1.5_50" ] && [ $3 != "resnet_v1_101" ] && [ $3 != "resnet_v1_152" ] +then + echo "error: the selected backbone must be resnet_v1_50, resnet_v1.5_50, resnet_v1_101, resnet_v1_152" exit 1 fi @@ -42,7 +48,7 @@ if [ ! -f $PATH2 ] then echo "error: CHECKPOINT_PATH=$PATH2 is not a file" exit 1 -fi +fi export DEVICE_NUM=1 export RANK_SIZE=$DEVICE_NUM @@ -60,5 +66,5 @@ cp -r ../src ./eval cd ./eval || exit env > env.log echo "start eval for device $DEVICE_ID" -python eval.py --device_target="GPU" --device_id=$DEVICE_ID --ann_file=$PATH1 --checkpoint_path=$PATH2 &> log & +python eval.py --device_target="GPU" --device_id=$DEVICE_ID --ann_file=$PATH1 --checkpoint_path=$PATH2 --backbone=$3 &> log & cd .. diff --git a/model_zoo/official/cv/faster_rcnn/scripts/run_standalone_train_ascend.sh b/model_zoo/official/cv/faster_rcnn/scripts/run_standalone_train_ascend.sh index d970f353006..e10976576b7 100755 --- a/model_zoo/official/cv/faster_rcnn/scripts/run_standalone_train_ascend.sh +++ b/model_zoo/official/cv/faster_rcnn/scripts/run_standalone_train_ascend.sh @@ -1,5 +1,5 @@ #!/bin/bash -# Copyright 2020 Huawei Technologies Co., Ltd +# Copyright 2020-2021 Huawei Technologies Co., Ltd # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. @@ -14,9 +14,15 @@ # limitations under the License. # ============================================================================ -if [ $# -ne 1 ] +if [ $# -ne 2 ] then - echo "Usage: sh run_standalone_train_ascend.sh [PRETRAINED_PATH]" + echo "Usage: sh run_standalone_train_ascend.sh [PRETRAINED_PATH] [BACKBONE]" +exit 1 +fi + +if [ $2 != "resnet_v1_50" ] && [ $2 != "resnet_v1.5_50" ] && [ $2 != "resnet_v1_101" ] && [ $2 != "resnet_v1_152" ] +then + echo "error: the selected backbone must be resnet_v1_50, resnet_v1.5_50, resnet_v1_101, resnet_v1_152" exit 1 fi @@ -54,5 +60,5 @@ cp -r ../src ./train cd ./train || exit echo "start training for device $DEVICE_ID" env > env.log -python train.py --device_id=$DEVICE_ID --pre_trained=$PATH1 &> log & -cd .. +python train.py --device_id=$DEVICE_ID --pre_trained=$PATH1 --backbone=$2 &> log & +cd .. \ No newline at end of file diff --git a/model_zoo/official/cv/faster_rcnn/scripts/run_standalone_train_gpu.sh b/model_zoo/official/cv/faster_rcnn/scripts/run_standalone_train_gpu.sh index 61bd7153fdb..8984fcb3be6 100755 --- a/model_zoo/official/cv/faster_rcnn/scripts/run_standalone_train_gpu.sh +++ b/model_zoo/official/cv/faster_rcnn/scripts/run_standalone_train_gpu.sh @@ -1,5 +1,5 @@ #!/bin/bash -# Copyright 2021 Huawei Technologies Co., Ltd +# Copyright 2020-2021 Huawei Technologies Co., Ltd # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. @@ -14,9 +14,15 @@ # limitations under the License. # ============================================================================ -if [ $# -ne 1 ] +if [ $# -ne 2 ] then - echo "Usage: sh run_standalone_train_gpu.sh [PRETRAINED_PATH]" + echo "Usage: sh run_standalone_train_gpu.sh [PRETRAINED_PATH] [BACKBONE]" +exit 1 +fi + +if [ $2 != "resnet_v1_50" ] && [ $2 != "resnet_v1.5_50" ] && [ $2 != "resnet_v1_101" ] && [ $2 != "resnet_v1_152" ] +then + echo "error: the selected backbone must be resnet_v1_50, resnet_v1.5_50, resnet_v1_101, resnet_v1_152" exit 1 fi @@ -54,5 +60,5 @@ cp -r ../src ./train cd ./train || exit echo "start training for device $DEVICE_ID" env > env.log -python train.py --device_id=$DEVICE_ID --pre_trained=$PATH1 --device_target="GPU" &> log & +python train.py --device_id=$DEVICE_ID --pre_trained=$PATH1 --device_target="GPU" --backbone=$2 &> log & cd .. diff --git a/model_zoo/official/cv/faster_rcnn/src/FasterRcnn/__init__.py b/model_zoo/official/cv/faster_rcnn/src/FasterRcnn/__init__.py index cbc0a27c661..78f136dcedb 100644 --- a/model_zoo/official/cv/faster_rcnn/src/FasterRcnn/__init__.py +++ b/model_zoo/official/cv/faster_rcnn/src/FasterRcnn/__init__.py @@ -1,4 +1,4 @@ -# Copyright 2020 Huawei Technologies Co., Ltd +# Copyright 2020-2021 Huawei Technologies Co., Ltd # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. @@ -14,7 +14,8 @@ # ============================================================================ """FasterRcnn Init.""" -from .resnet50 import ResNetFea, ResidualBlockUsing +from .resnet import ResNetFea, ResidualBlockUsing +from .resnet50v1 import ResidualBlockUsing_V1 from .bbox_assign_sample import BboxAssignSample from .bbox_assign_sample_stage2 import BboxAssignSampleForRcnn from .fpn_neck import FeatPyramidNeck @@ -27,5 +28,5 @@ from .anchor_generator import AnchorGenerator __all__ = [ "ResNetFea", "BboxAssignSample", "BboxAssignSampleForRcnn", "FeatPyramidNeck", "Proposal", "Rcnn", - "RPN", "SingleRoIExtractor", "AnchorGenerator", "ResidualBlockUsing" + "RPN", "SingleRoIExtractor", "AnchorGenerator", "ResidualBlockUsing", "ResidualBlockUsing_V1" ] diff --git a/model_zoo/official/cv/faster_rcnn/src/FasterRcnn/anchor_generator.py b/model_zoo/official/cv/faster_rcnn/src/FasterRcnn/anchor_generator.py index 666508c051a..1bc8935926d 100644 --- a/model_zoo/official/cv/faster_rcnn/src/FasterRcnn/anchor_generator.py +++ b/model_zoo/official/cv/faster_rcnn/src/FasterRcnn/anchor_generator.py @@ -1,4 +1,4 @@ -# Copyright 2020 Huawei Technologies Co., Ltd +# Copyright 2020-2021 Huawei Technologies Co., Ltd # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. diff --git a/model_zoo/official/cv/faster_rcnn/src/FasterRcnn/faster_rcnn_r50.py b/model_zoo/official/cv/faster_rcnn/src/FasterRcnn/faster_rcnn_resnet.py similarity index 92% rename from model_zoo/official/cv/faster_rcnn/src/FasterRcnn/faster_rcnn_r50.py rename to model_zoo/official/cv/faster_rcnn/src/FasterRcnn/faster_rcnn_resnet.py index 765f13d6bc3..4fa7bf0608f 100644 --- a/model_zoo/official/cv/faster_rcnn/src/FasterRcnn/faster_rcnn_r50.py +++ b/model_zoo/official/cv/faster_rcnn/src/FasterRcnn/faster_rcnn_resnet.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. # ============================================================================ -"""FasterRcnn based on ResNet50.""" +"""FasterRcnn based on ResNet.""" import numpy as np import mindspore.nn as nn @@ -21,7 +21,7 @@ from mindspore.ops import operations as P from mindspore.common.tensor import Tensor import mindspore.common.dtype as mstype from mindspore.ops import functional as F -from .resnet50 import ResNetFea, ResidualBlockUsing +from .resnet import ResNetFea, ResidualBlockUsing from .bbox_assign_sample_stage2 import BboxAssignSampleForRcnn from .fpn_neck import FeatPyramidNeck from .proposal_generator import Proposal @@ -31,12 +31,12 @@ from .roi_align import SingleRoIExtractor from .anchor_generator import AnchorGenerator -class Faster_Rcnn_Resnet50(nn.Cell): +class Faster_Rcnn_Resnet(nn.Cell): """ FasterRcnn Network. Note: - backbone = resnet50 + backbone = resnet Returns: Tuple, tuple of output tensor. @@ -48,10 +48,10 @@ class Faster_Rcnn_Resnet50(nn.Cell): rcnn_reg_loss: Scalar, Regression loss of RCNN subnet. Examples: - net = Faster_Rcnn_Resnet50() + net = Faster_Rcnn_Resnet() """ def __init__(self, config): - super(Faster_Rcnn_Resnet50, self).__init__() + super(Faster_Rcnn_Resnet, self).__init__() self.dtype = np.float32 self.ms_type = mstype.float32 self.train_batch_size = config.batch_size @@ -79,7 +79,7 @@ class Faster_Rcnn_Resnet50(nn.Cell): self.anchor_list = self.get_anchors(featmap_sizes) - # Backbone resnet50 + # Backbone resnet self.backbone = ResNetFea(ResidualBlockUsing, config.resnet_block, config.resnet_in_channels, @@ -121,7 +121,7 @@ class Faster_Rcnn_Resnet50(nn.Cell): self.roi_init(config) # Rcnn - self.rcnn = Rcnn(config, config.rcnn_in_channels * config.roi_layer['out_size'] * config.roi_layer['out_size'], + self.rcnn = Rcnn(config, config.rcnn_in_channels * config.roi_layer.out_size * config.roi_layer.out_size, self.train_batch_size, self.num_classes) # Op declare @@ -148,6 +148,19 @@ class Faster_Rcnn_Resnet50(nn.Cell): self.device_type = "Ascend" if context.get_context("device_target") == "Ascend" else "Others" def roi_init(self, config): + """ + Initialize roi from the config file + + Args: + config (file): config file. + roi_layer (dict): Numbers of block in different layers. + roi_align_out_channels (int): Out channel in each layer. + config.roi_align_featmap_strides (list): featmap_strides in each layer. + roi_align_finest_scale (int): finest_scale in roi. + + Examples: + self.roi_init(config) + """ self.roi_align = SingleRoIExtractor(config, config.roi_layer, config.roi_align_out_channels, @@ -164,6 +177,19 @@ class Faster_Rcnn_Resnet50(nn.Cell): self.roi_align_test.set_train_local(config, False) def test_mode_init(self, config): + """ + Initialize test_mode from the config file. + + Args: + config (file): config file. + test_batch_size (int): Size of test batch. + rpn_max_num (int): max num of rpn. + test_score_thresh (float): threshold of test score. + test_iou_thr (float): threshold of test iou. + + Examples: + self.test_mode_init(config) + """ self.test_batch_size = config.test_batch_size self.split = P.Split(axis=0, output_num=self.test_batch_size) self.split_shape = P.Split(axis=0, output_num=4) @@ -195,6 +221,7 @@ class Faster_Rcnn_Resnet50(nn.Cell): self.test_num_proposal = self.test_batch_size * self.rpn_max_num def init_tensor(self, config): + roi_align_index = [np.array(np.ones((config.num_expected_pos_stage2 + config.num_expected_neg_stage2, 1)) * i, dtype=self.dtype) for i in range(self.train_batch_size)] @@ -205,6 +232,19 @@ class Faster_Rcnn_Resnet50(nn.Cell): self.roi_align_index_test_tensor = Tensor(np.concatenate(roi_align_index_test)) def construct(self, img_data, img_metas, gt_bboxes, gt_labels, gt_valids): + """ + construct the FasterRcnn Network. + + Args: + img_data: input image data. + img_metas: meta label of img. + gt_bboxes (Tensor): get the value of bboxes. + gt_labels (Tensor): get the value of labels. + gt_valids (Tensor): get the valid part of bboxes. + + Returns: + Tuple,tuple of output tensor + """ x = self.backbone(img_data) x = self.fpn_ncek(x) @@ -440,7 +480,7 @@ class Faster_Rcnn_Resnet50(nn.Cell): class FasterRcnn_Infer(nn.Cell): def __init__(self, config): super(FasterRcnn_Infer, self).__init__() - self.network = Faster_Rcnn_Resnet50(config) + self.network = Faster_Rcnn_Resnet(config) self.network.set_train(False) def construct(self, img_data, img_metas): diff --git a/model_zoo/official/cv/faster_rcnn/src/FasterRcnn/faster_rcnn_resnet50v1.py b/model_zoo/official/cv/faster_rcnn/src/FasterRcnn/faster_rcnn_resnet50v1.py new file mode 100644 index 00000000000..c56867a35da --- /dev/null +++ b/model_zoo/official/cv/faster_rcnn/src/FasterRcnn/faster_rcnn_resnet50v1.py @@ -0,0 +1,488 @@ +# Copyright 2020-2021 Huawei Technologies Co., Ltd +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# ============================================================================ +"""FasterRcnn based on ResNet50v1.0.""" + +import numpy as np +import mindspore.nn as nn +from mindspore import context +from mindspore.ops import operations as P +from mindspore.common.tensor import Tensor +import mindspore.common.dtype as mstype +from mindspore.ops import functional as F +from .resnet import ResNetFea +from .resnet50v1 import ResidualBlockUsing_V1 +from .bbox_assign_sample_stage2 import BboxAssignSampleForRcnn +from .fpn_neck import FeatPyramidNeck +from .proposal_generator import Proposal +from .rcnn import Rcnn +from .rpn import RPN +from .roi_align import SingleRoIExtractor +from .anchor_generator import AnchorGenerator + + +class Faster_Rcnn_Resnet(nn.Cell): + """ + FasterRcnn Network. + + Note: + backbone = resnet + + Returns: + Tuple, tuple of output tensor. + rpn_loss: Scalar, Total loss of RPN subnet. + rcnn_loss: Scalar, Total loss of RCNN subnet. + rpn_cls_loss: Scalar, Classification loss of RPN subnet. + rpn_reg_loss: Scalar, Regression loss of RPN subnet. + rcnn_cls_loss: Scalar, Classification loss of RCNN subnet. + rcnn_reg_loss: Scalar, Regression loss of RCNN subnet. + + Examples: + net = Faster_Rcnn_Resnet() + """ + def __init__(self, config): + super(Faster_Rcnn_Resnet, self).__init__() + self.dtype = np.float32 + self.ms_type = mstype.float32 + self.train_batch_size = config.batch_size + self.num_classes = config.num_classes + self.anchor_scales = config.anchor_scales + self.anchor_ratios = config.anchor_ratios + self.anchor_strides = config.anchor_strides + self.target_means = tuple(config.rcnn_target_means) + self.target_stds = tuple(config.rcnn_target_stds) + + # Anchor generator + anchor_base_sizes = None + self.anchor_base_sizes = list( + self.anchor_strides) if anchor_base_sizes is None else anchor_base_sizes + + self.anchor_generators = [] + for anchor_base in self.anchor_base_sizes: + self.anchor_generators.append( + AnchorGenerator(anchor_base, self.anchor_scales, self.anchor_ratios)) + + self.num_anchors = len(self.anchor_ratios) * len(self.anchor_scales) + + featmap_sizes = config.feature_shapes + assert len(featmap_sizes) == len(self.anchor_generators) + + self.anchor_list = self.get_anchors(featmap_sizes) + + # Backbone resnet + self.backbone = ResNetFea(ResidualBlockUsing_V1, + config.resnet_block, + config.resnet_in_channels, + config.resnet_out_channels, + False) + + # Fpn + self.fpn_ncek = FeatPyramidNeck(config.fpn_in_channels, + config.fpn_out_channels, + config.fpn_num_outs) + + # Rpn and rpn loss + self.gt_labels_stage1 = Tensor(np.ones((self.train_batch_size, config.num_gts)).astype(np.uint8)) + self.rpn_with_loss = RPN(config, + self.train_batch_size, + config.rpn_in_channels, + config.rpn_feat_channels, + config.num_anchors, + config.rpn_cls_out_channels) + + # Proposal + self.proposal_generator = Proposal(config, + self.train_batch_size, + config.activate_num_classes, + config.use_sigmoid_cls) + self.proposal_generator.set_train_local(config, True) + self.proposal_generator_test = Proposal(config, + config.test_batch_size, + config.activate_num_classes, + config.use_sigmoid_cls) + self.proposal_generator_test.set_train_local(config, False) + + # Assign and sampler stage two + self.bbox_assigner_sampler_for_rcnn = BboxAssignSampleForRcnn(config, self.train_batch_size, + config.num_bboxes_stage2, True) + self.decode = P.BoundingBoxDecode(max_shape=(config.img_height, config.img_width), means=self.target_means, \ + stds=self.target_stds) + # Roi + self.roi_init(config) + + # Rcnn + self.rcnn = Rcnn(config, config.rcnn_in_channels * config.roi_layer.out_size * config.roi_layer.out_size, + self.train_batch_size, self.num_classes) + + # Op declare + self.squeeze = P.Squeeze() + self.cast = P.Cast() + + self.concat = P.Concat(axis=0) + self.concat_1 = P.Concat(axis=1) + self.concat_2 = P.Concat(axis=2) + self.reshape = P.Reshape() + self.select = P.Select() + self.greater = P.Greater() + self.transpose = P.Transpose() + + # Improve speed + self.concat_start = min(self.num_classes - 2, 55) + self.concat_end = (self.num_classes - 1) + + # Test mode + self.test_mode_init(config) + + # Init tensor + self.init_tensor(config) + self.device_type = "Ascend" if context.get_context("device_target") == "Ascend" else "Others" + + def roi_init(self, config): + """ + Initialize roi from the config file + + Args: + config (file): config file. + roi_layer (dict): Numbers of block in different layers. + roi_align_out_channels (int): Out channel in each layer. + config.roi_align_featmap_strides (list): featmap_strides in each layer. + roi_align_finest_scale (int): finest_scale in roi. + + Examples: + self.roi_init(config) + """ + self.roi_align = SingleRoIExtractor(config, + config.roi_layer, + config.roi_align_out_channels, + config.roi_align_featmap_strides, + self.train_batch_size, + config.roi_align_finest_scale) + self.roi_align.set_train_local(config, True) + self.roi_align_test = SingleRoIExtractor(config, + config.roi_layer, + config.roi_align_out_channels, + config.roi_align_featmap_strides, + 1, + config.roi_align_finest_scale) + self.roi_align_test.set_train_local(config, False) + + def test_mode_init(self, config): + """ + Initialize test_mode from the config file. + + Args: + config (file): config file. + test_batch_size (int): Size of test batch. + rpn_max_num (int): max num of rpn. + test_score_thresh (float): threshold of test score. + test_iou_thr (float): threshold of test iou. + + Examples: + self.test_mode_init(config) + """ + self.test_batch_size = config.test_batch_size + self.split = P.Split(axis=0, output_num=self.test_batch_size) + self.split_shape = P.Split(axis=0, output_num=4) + self.split_scores = P.Split(axis=1, output_num=self.num_classes) + self.split_cls = P.Split(axis=0, output_num=self.num_classes-1) + self.tile = P.Tile() + self.gather = P.GatherNd() + + self.rpn_max_num = config.rpn_max_num + + self.zeros_for_nms = Tensor(np.zeros((self.rpn_max_num, 3)).astype(self.dtype)) + self.ones_mask = np.ones((self.rpn_max_num, 1)).astype(np.bool) + self.zeros_mask = np.zeros((self.rpn_max_num, 1)).astype(np.bool) + self.bbox_mask = Tensor(np.concatenate((self.ones_mask, self.zeros_mask, + self.ones_mask, self.zeros_mask), axis=1)) + self.nms_pad_mask = Tensor(np.concatenate((self.ones_mask, self.ones_mask, + self.ones_mask, self.ones_mask, self.zeros_mask), axis=1)) + + self.test_score_thresh = Tensor(np.ones((self.rpn_max_num, 1)).astype(self.dtype) * config.test_score_thr) + self.test_score_zeros = Tensor(np.ones((self.rpn_max_num, 1)).astype(self.dtype) * 0) + self.test_box_zeros = Tensor(np.ones((self.rpn_max_num, 4)).astype(self.dtype) * -1) + self.test_iou_thr = Tensor(np.ones((self.rpn_max_num, 1)).astype(self.dtype) * config.test_iou_thr) + self.test_max_per_img = config.test_max_per_img + self.nms_test = P.NMSWithMask(config.test_iou_thr) + self.softmax = P.Softmax(axis=1) + self.logicand = P.LogicalAnd() + self.oneslike = P.OnesLike() + self.test_topk = P.TopK(sorted=True) + self.test_num_proposal = self.test_batch_size * self.rpn_max_num + + def init_tensor(self, config): + roi_align_index = [np.array(np.ones((config.num_expected_pos_stage2 + config.num_expected_neg_stage2, 1)) * i, + dtype=self.dtype) for i in range(self.train_batch_size)] + + roi_align_index_test = [np.array(np.ones((config.rpn_max_num, 1)) * i, dtype=self.dtype) \ + for i in range(self.test_batch_size)] + + self.roi_align_index_tensor = Tensor(np.concatenate(roi_align_index)) + self.roi_align_index_test_tensor = Tensor(np.concatenate(roi_align_index_test)) + + def construct(self, img_data, img_metas, gt_bboxes, gt_labels, gt_valids): + """ + construct the FasterRcnn Network. + + Args: + img_data: input image data. + img_metas: meta label of img. + gt_bboxes (Tensor): get the value of bboxes. + gt_labels (Tensor): get the value of labels. + gt_valids (Tensor): get the valid part of bboxes. + + Returns: + Tuple,tuple of output tensor + """ + x = self.backbone(img_data) + x = self.fpn_ncek(x) + + rpn_loss, cls_score, bbox_pred, rpn_cls_loss, rpn_reg_loss, _ = self.rpn_with_loss(x, + img_metas, + self.anchor_list, + gt_bboxes, + self.gt_labels_stage1, + gt_valids) + + if self.training: + proposal, proposal_mask = self.proposal_generator(cls_score, bbox_pred, self.anchor_list) + else: + proposal, proposal_mask = self.proposal_generator_test(cls_score, bbox_pred, self.anchor_list) + + gt_labels = self.cast(gt_labels, mstype.int32) + gt_valids = self.cast(gt_valids, mstype.int32) + bboxes_tuple = () + deltas_tuple = () + labels_tuple = () + mask_tuple = () + if self.training: + for i in range(self.train_batch_size): + gt_bboxes_i = self.squeeze(gt_bboxes[i:i + 1:1, ::]) + + gt_labels_i = self.squeeze(gt_labels[i:i + 1:1, ::]) + gt_labels_i = self.cast(gt_labels_i, mstype.uint8) + + gt_valids_i = self.squeeze(gt_valids[i:i + 1:1, ::]) + gt_valids_i = self.cast(gt_valids_i, mstype.bool_) + + bboxes, deltas, labels, mask = self.bbox_assigner_sampler_for_rcnn(gt_bboxes_i, + gt_labels_i, + proposal_mask[i], + proposal[i][::, 0:4:1], + gt_valids_i) + bboxes_tuple += (bboxes,) + deltas_tuple += (deltas,) + labels_tuple += (labels,) + mask_tuple += (mask,) + + bbox_targets = self.concat(deltas_tuple) + rcnn_labels = self.concat(labels_tuple) + bbox_targets = F.stop_gradient(bbox_targets) + rcnn_labels = F.stop_gradient(rcnn_labels) + rcnn_labels = self.cast(rcnn_labels, mstype.int32) + else: + mask_tuple += proposal_mask + bbox_targets = proposal_mask + rcnn_labels = proposal_mask + for p_i in proposal: + bboxes_tuple += (p_i[::, 0:4:1],) + + if self.training: + if self.train_batch_size > 1: + bboxes_all = self.concat(bboxes_tuple) + else: + bboxes_all = bboxes_tuple[0] + rois = self.concat_1((self.roi_align_index_tensor, bboxes_all)) + else: + if self.test_batch_size > 1: + bboxes_all = self.concat(bboxes_tuple) + else: + bboxes_all = bboxes_tuple[0] + if self.device_type == "Ascend": + bboxes_all = self.cast(bboxes_all, mstype.float16) + rois = self.concat_1((self.roi_align_index_test_tensor, bboxes_all)) + + rois = self.cast(rois, mstype.float32) + rois = F.stop_gradient(rois) + + if self.training: + roi_feats = self.roi_align(rois, + self.cast(x[0], mstype.float32), + self.cast(x[1], mstype.float32), + self.cast(x[2], mstype.float32), + self.cast(x[3], mstype.float32)) + else: + roi_feats = self.roi_align_test(rois, + self.cast(x[0], mstype.float32), + self.cast(x[1], mstype.float32), + self.cast(x[2], mstype.float32), + self.cast(x[3], mstype.float32)) + + roi_feats = self.cast(roi_feats, self.ms_type) + rcnn_masks = self.concat(mask_tuple) + rcnn_masks = F.stop_gradient(rcnn_masks) + rcnn_mask_squeeze = self.squeeze(self.cast(rcnn_masks, mstype.bool_)) + rcnn_loss, rcnn_cls_loss, rcnn_reg_loss, _ = self.rcnn(roi_feats, + bbox_targets, + rcnn_labels, + rcnn_mask_squeeze) + + output = () + if self.training: + output += (rpn_loss, rcnn_loss, rpn_cls_loss, rpn_reg_loss, rcnn_cls_loss, rcnn_reg_loss) + else: + output = self.get_det_bboxes(rcnn_cls_loss, rcnn_reg_loss, rcnn_masks, bboxes_all, img_metas) + + return output + + def get_det_bboxes(self, cls_logits, reg_logits, mask_logits, rois, img_metas): + """Get the actual detection box.""" + scores = self.softmax(cls_logits) + + boxes_all = () + for i in range(self.num_classes): + k = i * 4 + reg_logits_i = self.squeeze(reg_logits[::, k:k+4:1]) + out_boxes_i = self.decode(rois, reg_logits_i) + boxes_all += (out_boxes_i,) + + img_metas_all = self.split(img_metas) + scores_all = self.split(scores) + mask_all = self.split(self.cast(mask_logits, mstype.int32)) + + boxes_all_with_batchsize = () + for i in range(self.test_batch_size): + scale = self.split_shape(self.squeeze(img_metas_all[i])) + scale_h = scale[2] + scale_w = scale[3] + boxes_tuple = () + for j in range(self.num_classes): + boxes_tmp = self.split(boxes_all[j]) + out_boxes_h = boxes_tmp[i] / scale_h + out_boxes_w = boxes_tmp[i] / scale_w + boxes_tuple += (self.select(self.bbox_mask, out_boxes_w, out_boxes_h),) + boxes_all_with_batchsize += (boxes_tuple,) + + output = self.multiclass_nms(boxes_all_with_batchsize, scores_all, mask_all) + + return output + + def multiclass_nms(self, boxes_all, scores_all, mask_all): + """Multiscale postprocessing.""" + all_bboxes = () + all_labels = () + all_masks = () + + for i in range(self.test_batch_size): + bboxes = boxes_all[i] + scores = scores_all[i] + masks = self.cast(mask_all[i], mstype.bool_) + + res_boxes_tuple = () + res_labels_tuple = () + res_masks_tuple = () + + for j in range(self.num_classes - 1): + k = j + 1 + _cls_scores = scores[::, k:k + 1:1] + _bboxes = self.squeeze(bboxes[k]) + _mask_o = self.reshape(masks, (self.rpn_max_num, 1)) + + cls_mask = self.greater(_cls_scores, self.test_score_thresh) + _mask = self.logicand(_mask_o, cls_mask) + + _reg_mask = self.cast(self.tile(self.cast(_mask, mstype.int32), (1, 4)), mstype.bool_) + + _bboxes = self.select(_reg_mask, _bboxes, self.test_box_zeros) + _cls_scores = self.select(_mask, _cls_scores, self.test_score_zeros) + __cls_scores = self.squeeze(_cls_scores) + scores_sorted, topk_inds = self.test_topk(__cls_scores, self.rpn_max_num) + topk_inds = self.reshape(topk_inds, (self.rpn_max_num, 1)) + scores_sorted = self.reshape(scores_sorted, (self.rpn_max_num, 1)) + _bboxes_sorted = self.gather(_bboxes, topk_inds) + _mask_sorted = self.gather(_mask, topk_inds) + + scores_sorted = self.tile(scores_sorted, (1, 4)) + cls_dets = self.concat_1((_bboxes_sorted, scores_sorted)) + cls_dets = P.Slice()(cls_dets, (0, 0), (self.rpn_max_num, 5)) + + cls_dets, _index, _mask_nms = self.nms_test(cls_dets) + _index = self.reshape(_index, (self.rpn_max_num, 1)) + _mask_nms = self.reshape(_mask_nms, (self.rpn_max_num, 1)) + + _mask_n = self.gather(_mask_sorted, _index) + + _mask_n = self.logicand(_mask_n, _mask_nms) + cls_labels = self.oneslike(_index) * j + res_boxes_tuple += (cls_dets,) + res_labels_tuple += (cls_labels,) + res_masks_tuple += (_mask_n,) + + res_boxes_start = self.concat(res_boxes_tuple[:self.concat_start]) + res_labels_start = self.concat(res_labels_tuple[:self.concat_start]) + res_masks_start = self.concat(res_masks_tuple[:self.concat_start]) + + res_boxes_end = self.concat(res_boxes_tuple[self.concat_start:self.concat_end]) + res_labels_end = self.concat(res_labels_tuple[self.concat_start:self.concat_end]) + res_masks_end = self.concat(res_masks_tuple[self.concat_start:self.concat_end]) + + res_boxes = self.concat((res_boxes_start, res_boxes_end)) + res_labels = self.concat((res_labels_start, res_labels_end)) + res_masks = self.concat((res_masks_start, res_masks_end)) + + reshape_size = (self.num_classes - 1) * self.rpn_max_num + res_boxes = self.reshape(res_boxes, (1, reshape_size, 5)) + res_labels = self.reshape(res_labels, (1, reshape_size, 1)) + res_masks = self.reshape(res_masks, (1, reshape_size, 1)) + + all_bboxes += (res_boxes,) + all_labels += (res_labels,) + all_masks += (res_masks,) + + all_bboxes = self.concat(all_bboxes) + all_labels = self.concat(all_labels) + all_masks = self.concat(all_masks) + return all_bboxes, all_labels, all_masks + + def get_anchors(self, featmap_sizes): + """Get anchors according to feature map sizes. + + Args: + featmap_sizes (list[tuple]): Multi-level feature map sizes. + img_metas (list[dict]): Image meta info. + + Returns: + tuple: anchors of each image, valid flags of each image + """ + num_levels = len(featmap_sizes) + + # since feature map sizes of all images are the same, we only compute + # anchors for one time + multi_level_anchors = () + for i in range(num_levels): + anchors = self.anchor_generators[i].grid_anchors( + featmap_sizes[i], self.anchor_strides[i]) + multi_level_anchors += (Tensor(anchors.astype(self.dtype)),) + + return multi_level_anchors + +class FasterRcnn_Infer(nn.Cell): + def __init__(self, config): + super(FasterRcnn_Infer, self).__init__() + self.network = Faster_Rcnn_Resnet(config) + self.network.set_train(False) + + def construct(self, img_data, img_metas): + output = self.network(img_data, img_metas, None, None, None) + return output diff --git a/model_zoo/official/cv/faster_rcnn/src/FasterRcnn/resnet50.py b/model_zoo/official/cv/faster_rcnn/src/FasterRcnn/resnet.py similarity index 96% rename from model_zoo/official/cv/faster_rcnn/src/FasterRcnn/resnet50.py rename to model_zoo/official/cv/faster_rcnn/src/FasterRcnn/resnet.py index 2480ecfc921..587f84b028c 100644 --- a/model_zoo/official/cv/faster_rcnn/src/FasterRcnn/resnet50.py +++ b/model_zoo/official/cv/faster_rcnn/src/FasterRcnn/resnet.py @@ -1,244 +1,262 @@ -# Copyright 2020-2021 Huawei Technologies Co., Ltd -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. -# ============================================================================ -"""Resnet50 backbone.""" - -import numpy as np -import mindspore.nn as nn -from mindspore.ops import operations as P -from mindspore.common.tensor import Tensor -from mindspore.ops import functional as F - - -def weight_init_ones(shape): - """Weight init.""" - return Tensor(np.full(shape, 0.01).astype(np.float32)) - - -def _conv(in_channels, out_channels, kernel_size=3, stride=1, padding=0, pad_mode='pad'): - """Conv2D wrapper.""" - shape = (out_channels, in_channels, kernel_size, kernel_size) - weights = weight_init_ones(shape) - return nn.Conv2d(in_channels, out_channels, - kernel_size=kernel_size, stride=stride, padding=padding, - pad_mode=pad_mode, weight_init=weights, has_bias=False) - - -def _BatchNorm2dInit(out_chls, momentum=0.1, affine=True, use_batch_statistics=True): - """Batchnorm2D wrapper.""" - dtype = np.float32 - gamma_init = Tensor(np.array(np.ones(out_chls)).astype(dtype)) - beta_init = Tensor(np.array(np.ones(out_chls) * 0).astype(dtype)) - moving_mean_init = Tensor(np.array(np.ones(out_chls) * 0).astype(dtype)) - moving_var_init = Tensor(np.array(np.ones(out_chls)).astype(dtype)) - return nn.BatchNorm2d(out_chls, momentum=momentum, affine=affine, gamma_init=gamma_init, - beta_init=beta_init, moving_mean_init=moving_mean_init, - moving_var_init=moving_var_init, use_batch_statistics=use_batch_statistics) - - -class ResNetFea(nn.Cell): - """ - ResNet architecture. - - Args: - block (Cell): Block for network. - layer_nums (list): Numbers of block in different layers. - in_channels (list): Input channel in each layer. - out_channels (list): Output channel in each layer. - weights_update (bool): Weight update flag. - Returns: - Tensor, output tensor. - - Examples: - >>> ResNet(ResidualBlock, - >>> [3, 4, 6, 3], - >>> [64, 256, 512, 1024], - >>> [256, 512, 1024, 2048], - >>> False) - """ - def __init__(self, - block, - layer_nums, - in_channels, - out_channels, - weights_update=False): - super(ResNetFea, self).__init__() - - if not len(layer_nums) == len(in_channels) == len(out_channels) == 4: - raise ValueError("the length of " - "layer_num, inchannel, outchannel list must be 4!") - - bn_training = False - self.conv1 = _conv(3, 64, kernel_size=7, stride=2, padding=3, pad_mode='pad') - self.bn1 = _BatchNorm2dInit(64, affine=bn_training, use_batch_statistics=bn_training) - self.relu = P.ReLU() - self.maxpool = P.MaxPool(kernel_size=3, strides=2, pad_mode="SAME") - self.weights_update = weights_update - - if not self.weights_update: - self.conv1.weight.requires_grad = False - - self.layer1 = self._make_layer(block, - layer_nums[0], - in_channel=in_channels[0], - out_channel=out_channels[0], - stride=1, - training=bn_training, - weights_update=self.weights_update) - self.layer2 = self._make_layer(block, - layer_nums[1], - in_channel=in_channels[1], - out_channel=out_channels[1], - stride=2, - training=bn_training, - weights_update=True) - self.layer3 = self._make_layer(block, - layer_nums[2], - in_channel=in_channels[2], - out_channel=out_channels[2], - stride=2, - training=bn_training, - weights_update=True) - self.layer4 = self._make_layer(block, - layer_nums[3], - in_channel=in_channels[3], - out_channel=out_channels[3], - stride=2, - training=bn_training, - weights_update=True) - - def _make_layer(self, block, layer_num, in_channel, out_channel, stride, training=False, weights_update=False): - """Make block layer.""" - layers = [] - down_sample = False - if stride != 1 or in_channel != out_channel: - down_sample = True - resblk = block(in_channel, - out_channel, - stride=stride, - down_sample=down_sample, - training=training, - weights_update=weights_update) - layers.append(resblk) - - for _ in range(1, layer_num): - resblk = block(out_channel, out_channel, stride=1, training=training, weights_update=weights_update) - layers.append(resblk) - - return nn.SequentialCell(layers) - - def construct(self, x): - x = self.conv1(x) - x = self.bn1(x) - x = self.relu(x) - c1 = self.maxpool(x) - - c2 = self.layer1(c1) - identity = c2 - if not self.weights_update: - identity = F.stop_gradient(c2) - c3 = self.layer2(identity) - c4 = self.layer3(c3) - c5 = self.layer4(c4) - - return identity, c3, c4, c5 - - -class ResidualBlockUsing(nn.Cell): - """ - ResNet V1 residual block definition. - - Args: - in_channels (int) - Input channel. - out_channels (int) - Output channel. - stride (int) - Stride size for the initial convolutional layer. Default: 1. - down_sample (bool) - If to do the downsample in block. Default: False. - momentum (float) - Momentum for batchnorm layer. Default: 0.1. - training (bool) - Training flag. Default: False. - weights_updata (bool) - Weights update flag. Default: False. - - Returns: - Tensor, output tensor. - - Examples: - ResidualBlock(3,256,stride=2,down_sample=True) - """ - expansion = 4 - - def __init__(self, - in_channels, - out_channels, - stride=1, - down_sample=False, - momentum=0.1, - training=False, - weights_update=False): - super(ResidualBlockUsing, self).__init__() - - self.affine = weights_update - - out_chls = out_channels // self.expansion - self.conv1 = _conv(in_channels, out_chls, kernel_size=1, stride=1, padding=0) - self.bn1 = _BatchNorm2dInit(out_chls, momentum=momentum, affine=self.affine, use_batch_statistics=training) - - self.conv2 = _conv(out_chls, out_chls, kernel_size=3, stride=stride, padding=1) - self.bn2 = _BatchNorm2dInit(out_chls, momentum=momentum, affine=self.affine, use_batch_statistics=training) - - self.conv3 = _conv(out_chls, out_channels, kernel_size=1, stride=1, padding=0) - self.bn3 = _BatchNorm2dInit(out_channels, momentum=momentum, affine=self.affine, use_batch_statistics=training) - - if training: - self.bn1 = self.bn1.set_train() - self.bn2 = self.bn2.set_train() - self.bn3 = self.bn3.set_train() - - if not weights_update: - self.conv1.weight.requires_grad = False - self.conv2.weight.requires_grad = False - self.conv3.weight.requires_grad = False - - self.relu = P.ReLU() - self.downsample = down_sample - if self.downsample: - self.conv_down_sample = _conv(in_channels, out_channels, kernel_size=1, stride=stride, padding=0) - self.bn_down_sample = _BatchNorm2dInit(out_channels, momentum=momentum, affine=self.affine, - use_batch_statistics=training) - if training: - self.bn_down_sample = self.bn_down_sample.set_train() - if not weights_update: - self.conv_down_sample.weight.requires_grad = False - self.add = P.Add() - - def construct(self, x): - identity = x - - out = self.conv1(x) - out = self.bn1(out) - out = self.relu(out) - - out = self.conv2(out) - out = self.bn2(out) - out = self.relu(out) - - out = self.conv3(out) - out = self.bn3(out) - - if self.downsample: - identity = self.conv_down_sample(identity) - identity = self.bn_down_sample(identity) - - out = self.add(out, identity) - out = self.relu(out) - - return out +# Copyright 2020-2021 Huawei Technologies Co., Ltd +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# ============================================================================ +"""Resnet backbone.""" + +import numpy as np +import mindspore.nn as nn +from mindspore.ops import operations as P +from mindspore.common.tensor import Tensor +from mindspore.ops import functional as F + + +def weight_init_ones(shape): + """Weight init.""" + return Tensor(np.full(shape, 0.01).astype(np.float32)) + + +def _conv(in_channels, out_channels, kernel_size=3, stride=1, padding=0, pad_mode='pad'): + """Conv2D wrapper.""" + shape = (out_channels, in_channels, kernel_size, kernel_size) + weights = weight_init_ones(shape) + return nn.Conv2d(in_channels, out_channels, + kernel_size=kernel_size, stride=stride, padding=padding, + pad_mode=pad_mode, weight_init=weights, has_bias=False) + + +def _BatchNorm2dInit(out_chls, momentum=0.1, affine=True, use_batch_statistics=True): + """Batchnorm2D wrapper.""" + dtype = np.float32 + gamma_init = Tensor(np.array(np.ones(out_chls)).astype(dtype)) + beta_init = Tensor(np.array(np.ones(out_chls) * 0).astype(dtype)) + moving_mean_init = Tensor(np.array(np.ones(out_chls) * 0).astype(dtype)) + moving_var_init = Tensor(np.array(np.ones(out_chls)).astype(dtype)) + return nn.BatchNorm2d(out_chls, momentum=momentum, affine=affine, gamma_init=gamma_init, + beta_init=beta_init, moving_mean_init=moving_mean_init, + moving_var_init=moving_var_init, use_batch_statistics=use_batch_statistics) + + +class ResNetFea(nn.Cell): + """ + ResNet architecture. + + Args: + block (Cell): Block for network. + layer_nums (list): Numbers of block in different layers. + in_channels (list): Input channel in each layer. + out_channels (list): Output channel in each layer. + weights_update (bool): Weight update flag. + Returns: + Tensor, output tensor. + + Examples: + >>> ResNet(ResidualBlock, + >>> [3, 4, 6, 3], + >>> [64, 256, 512, 1024], + >>> [256, 512, 1024, 2048], + >>> False) + """ + def __init__(self, + block, + layer_nums, + in_channels, + out_channels, + weights_update=False): + super(ResNetFea, self).__init__() + + if not len(layer_nums) == len(in_channels) == len(out_channels) == 4: + raise ValueError("the length of " + "layer_num, inchannel, outchannel list must be 4!") + + bn_training = False + self.conv1 = _conv(3, 64, kernel_size=7, stride=2, padding=3, pad_mode='pad') + self.bn1 = _BatchNorm2dInit(64, affine=bn_training, use_batch_statistics=bn_training) + self.relu = P.ReLU() + self.maxpool = P.MaxPool(kernel_size=3, strides=2, pad_mode="SAME") + self.weights_update = weights_update + + if not self.weights_update: + self.conv1.weight.requires_grad = False + + self.layer1 = self._make_layer(block, + layer_nums[0], + in_channel=in_channels[0], + out_channel=out_channels[0], + stride=1, + training=bn_training, + weights_update=self.weights_update) + self.layer2 = self._make_layer(block, + layer_nums[1], + in_channel=in_channels[1], + out_channel=out_channels[1], + stride=2, + training=bn_training, + weights_update=True) + self.layer3 = self._make_layer(block, + layer_nums[2], + in_channel=in_channels[2], + out_channel=out_channels[2], + stride=2, + training=bn_training, + weights_update=True) + self.layer4 = self._make_layer(block, + layer_nums[3], + in_channel=in_channels[3], + out_channel=out_channels[3], + stride=2, + training=bn_training, + weights_update=True) + + def _make_layer(self, block, layer_num, in_channel, out_channel, stride, training=False, weights_update=False): + """Make block layer.""" + layers = [] + down_sample = False + if stride != 1 or in_channel != out_channel: + down_sample = True + resblk = block(in_channel, + out_channel, + stride=stride, + down_sample=down_sample, + training=training, + weights_update=weights_update) + layers.append(resblk) + + for _ in range(1, layer_num): + resblk = block(out_channel, out_channel, stride=1, training=training, weights_update=weights_update) + layers.append(resblk) + + return nn.SequentialCell(layers) + + def construct(self, x): + """ + construct the ResNet Network + + Args: + x: input feature data. + + Returns: + Tensor, output tensor. + """ + x = self.conv1(x) + x = self.bn1(x) + x = self.relu(x) + c1 = self.maxpool(x) + + c2 = self.layer1(c1) + identity = c2 + if not self.weights_update: + identity = F.stop_gradient(c2) + c3 = self.layer2(identity) + c4 = self.layer3(c3) + c5 = self.layer4(c4) + + return identity, c3, c4, c5 + + +class ResidualBlockUsing(nn.Cell): + """ + ResNet V1 residual block definition. + + Args: + in_channels (int) - Input channel. + out_channels (int) - Output channel. + stride (int) - Stride size for the initial convolutional layer. Default: 1. + down_sample (bool) - If to do the downsample in block. Default: False. + momentum (float) - Momentum for batchnorm layer. Default: 0.1. + training (bool) - Training flag. Default: False. + weights_updata (bool) - Weights update flag. Default: False. + + Returns: + Tensor, output tensor. + + Examples: + ResidualBlock(3,256,stride=2,down_sample=True) + """ + expansion = 4 + + def __init__(self, + in_channels, + out_channels, + stride=1, + down_sample=False, + momentum=0.1, + training=False, + weights_update=False): + super(ResidualBlockUsing, self).__init__() + + self.affine = weights_update + + out_chls = out_channels // self.expansion + self.conv1 = _conv(in_channels, out_chls, kernel_size=1, stride=1, padding=0) + self.bn1 = _BatchNorm2dInit(out_chls, momentum=momentum, affine=self.affine, use_batch_statistics=training) + + self.conv2 = _conv(out_chls, out_chls, kernel_size=3, stride=stride, padding=1) + self.bn2 = _BatchNorm2dInit(out_chls, momentum=momentum, affine=self.affine, use_batch_statistics=training) + + self.conv3 = _conv(out_chls, out_channels, kernel_size=1, stride=1, padding=0) + self.bn3 = _BatchNorm2dInit(out_channels, momentum=momentum, affine=self.affine, use_batch_statistics=training) + + if training: + self.bn1 = self.bn1.set_train() + self.bn2 = self.bn2.set_train() + self.bn3 = self.bn3.set_train() + + if not weights_update: + self.conv1.weight.requires_grad = False + self.conv2.weight.requires_grad = False + self.conv3.weight.requires_grad = False + + self.relu = P.ReLU() + self.downsample = down_sample + if self.downsample: + self.conv_down_sample = _conv(in_channels, out_channels, kernel_size=1, stride=stride, padding=0) + self.bn_down_sample = _BatchNorm2dInit(out_channels, momentum=momentum, affine=self.affine, + use_batch_statistics=training) + if training: + self.bn_down_sample = self.bn_down_sample.set_train() + if not weights_update: + self.conv_down_sample.weight.requires_grad = False + self.add = P.Add() + + def construct(self, x): + """ + construct the ResNet V1 residual block + + Args: + x: input feature data. + + Returns: + Tensor, output tensor. + """ + identity = x + + out = self.conv1(x) + out = self.bn1(out) + out = self.relu(out) + + out = self.conv2(out) + out = self.bn2(out) + out = self.relu(out) + + out = self.conv3(out) + out = self.bn3(out) + + if self.downsample: + identity = self.conv_down_sample(identity) + identity = self.bn_down_sample(identity) + + out = self.add(out, identity) + out = self.relu(out) + + return out diff --git a/model_zoo/official/cv/faster_rcnn/src/FasterRcnn/resnet50v1.py b/model_zoo/official/cv/faster_rcnn/src/FasterRcnn/resnet50v1.py new file mode 100644 index 00000000000..58aca1b293a --- /dev/null +++ b/model_zoo/official/cv/faster_rcnn/src/FasterRcnn/resnet50v1.py @@ -0,0 +1,264 @@ +# Copyright 2020-2021 Huawei Technologies Co., Ltd +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# ============================================================================ +"""Resnet50v1.0 backbone.""" + +import numpy as np +import mindspore.nn as nn +from mindspore.ops import operations as P +from mindspore.common.tensor import Tensor +from mindspore.ops import functional as F + + +def weight_init_ones(shape): + """Weight init.""" + return Tensor(np.full(shape, 0.01).astype(np.float32)) + + +def _conv(in_channels, out_channels, kernel_size=3, stride=1, padding=0, pad_mode='pad'): + """Conv2D wrapper.""" + shape = (out_channels, in_channels, kernel_size, kernel_size) + weights = weight_init_ones(shape) + return nn.Conv2d(in_channels, out_channels, + kernel_size=kernel_size, stride=stride, padding=padding, + pad_mode=pad_mode, weight_init=weights, has_bias=False) + + +def _BatchNorm2dInit(out_chls, momentum=0.1, affine=True, use_batch_statistics=True): + """Batchnorm2D wrapper.""" + dtype = np.float32 + gamma_init = Tensor(np.array(np.ones(out_chls)).astype(dtype)) + beta_init = Tensor(np.array(np.ones(out_chls) * 0).astype(dtype)) + moving_mean_init = Tensor(np.array(np.ones(out_chls) * 0).astype(dtype)) + moving_var_init = Tensor(np.array(np.ones(out_chls)).astype(dtype)) + return nn.BatchNorm2d(out_chls, momentum=momentum, affine=affine, gamma_init=gamma_init, + beta_init=beta_init, moving_mean_init=moving_mean_init, + moving_var_init=moving_var_init, use_batch_statistics=use_batch_statistics) + + +class ResNetFea(nn.Cell): + """ + ResNet architecture. + + Args: + block (Cell): Block for network. + layer_nums (list): Numbers of block in different layers. + in_channels (list): Input channel in each layer. + out_channels (list): Output channel in each layer. + weights_update (bool): Weight update flag. + Returns: + Tensor, output tensor. + + Examples: + >>> ResNet(ResidualBlock, + >>> [3, 4, 6, 3], + >>> [64, 256, 512, 1024], + >>> [256, 512, 1024, 2048], + >>> False) + """ + def __init__(self, + block, + layer_nums, + in_channels, + out_channels, + weights_update=False): + super(ResNetFea, self).__init__() + + if not len(layer_nums) == len(in_channels) == len(out_channels) == 4: + raise ValueError("the length of " + "layer_num, inchannel, outchannel list must be 4!") + + bn_training = False + self.conv1 = _conv(3, 64, kernel_size=7, stride=2, padding=3, pad_mode='pad') + self.bn1 = _BatchNorm2dInit(64, affine=bn_training, use_batch_statistics=bn_training) + self.relu = P.ReLU() + self.maxpool = P.MaxPool(kernel_size=3, strides=2, pad_mode="SAME") + self.weights_update = weights_update + + if not self.weights_update: + self.conv1.weight.requires_grad = False + + self.layer1 = self._make_layer(block, + layer_nums[0], + in_channel=in_channels[0], + out_channel=out_channels[0], + stride=1, + training=bn_training, + weights_update=self.weights_update) + self.layer2 = self._make_layer(block, + layer_nums[1], + in_channel=in_channels[1], + out_channel=out_channels[1], + stride=2, + training=bn_training, + weights_update=True) + self.layer3 = self._make_layer(block, + layer_nums[2], + in_channel=in_channels[2], + out_channel=out_channels[2], + stride=2, + training=bn_training, + weights_update=True) + self.layer4 = self._make_layer(block, + layer_nums[3], + in_channel=in_channels[3], + out_channel=out_channels[3], + stride=2, + training=bn_training, + weights_update=True) + + def _make_layer(self, block, layer_num, in_channel, out_channel, stride, training=False, weights_update=False): + """Make block layer.""" + layers = [] + down_sample = False + if stride != 1 or in_channel != out_channel: + down_sample = True + resblk = block(in_channel, + out_channel, + stride=stride, + down_sample=down_sample, + training=training, + weights_update=weights_update) + layers.append(resblk) + + for _ in range(1, layer_num): + resblk = block(out_channel, out_channel, stride=1, training=training, weights_update=weights_update) + layers.append(resblk) + + return nn.SequentialCell(layers) + + def construct(self, x): + """ + construct the ResNet Network + + Args: + x: input feature data. + + Returns: + Tensor, output tensor. + """ + x = self.conv1(x) + x = self.bn1(x) + x = self.relu(x) + c1 = self.maxpool(x) + + c2 = self.layer1(c1) + identity = c2 + if not self.weights_update: + identity = F.stop_gradient(c2) + c3 = self.layer2(identity) + c4 = self.layer3(c3) + c5 = self.layer4(c4) + + return identity, c3, c4, c5 + + +class ResidualBlockUsing_V1(nn.Cell): + """ + ResNet V1 residual block definition. + + Args: + in_channels (int) - Input channel. + out_channels (int) - Output channel. + stride (int) - Stride size for the initial convolutional layer. Default: 1. + down_sample (bool) - If to do the downsample in block. Default: False. + momentum (float) - Momentum for batchnorm layer. Default: 0.1. + training (bool) - Training flag. Default: False. + weights_updata (bool) - Weights update flag. Default: False. + + Returns: + Tensor, output tensor. + + Examples: + ResidualBlock(3,256,stride=2,down_sample=True) + """ + expansion = 4 + + def __init__(self, + in_channels, + out_channels, + stride=1, + down_sample=False, + momentum=0.1, + training=False, + weights_update=False): + super(ResidualBlockUsing_V1, self).__init__() + + self.affine = weights_update + + out_chls = out_channels // self.expansion + # self.conv1 = _conv(in_channels, out_chls, kernel_size=1, stride=1, padding=0) + self.conv1 = _conv(in_channels, out_chls, kernel_size=1, stride=stride, padding=0) + self.bn1 = _BatchNorm2dInit(out_chls, momentum=momentum, affine=self.affine, use_batch_statistics=training) + + # self.conv2 = _conv(out_chls, out_chls, kernel_size=3, stride=stride, padding=1) + self.conv2 = _conv(out_chls, out_chls, kernel_size=3, stride=1, padding=1) + self.bn2 = _BatchNorm2dInit(out_chls, momentum=momentum, affine=self.affine, use_batch_statistics=training) + + self.conv3 = _conv(out_chls, out_channels, kernel_size=1, stride=1, padding=0) + self.bn3 = _BatchNorm2dInit(out_channels, momentum=momentum, affine=self.affine, use_batch_statistics=training) + + if training: + self.bn1 = self.bn1.set_train() + self.bn2 = self.bn2.set_train() + self.bn3 = self.bn3.set_train() + + if not weights_update: + self.conv1.weight.requires_grad = False + self.conv2.weight.requires_grad = False + self.conv3.weight.requires_grad = False + + self.relu = P.ReLU() + self.downsample = down_sample + if self.downsample: + self.conv_down_sample = _conv(in_channels, out_channels, kernel_size=1, stride=stride, padding=0) + self.bn_down_sample = _BatchNorm2dInit(out_channels, momentum=momentum, affine=self.affine, + use_batch_statistics=training) + if training: + self.bn_down_sample = self.bn_down_sample.set_train() + if not weights_update: + self.conv_down_sample.weight.requires_grad = False + self.add = P.Add() + + def construct(self, x): + """ + construct the ResNet V1 residual block + + Args: + x: input feature data. + + Returns: + Tensor, output tensor. + """ + identity = x + + out = self.conv1(x) + out = self.bn1(out) + out = self.relu(out) + + out = self.conv2(out) + out = self.bn2(out) + out = self.relu(out) + + out = self.conv3(out) + out = self.bn3(out) + + if self.downsample: + identity = self.conv_down_sample(identity) + identity = self.bn_down_sample(identity) + + out = self.add(out, identity) + out = self.relu(out) + + return out diff --git a/model_zoo/official/cv/faster_rcnn/src/FasterRcnn/roi_align.py b/model_zoo/official/cv/faster_rcnn/src/FasterRcnn/roi_align.py index f099de0e867..4876f8943e1 100644 --- a/model_zoo/official/cv/faster_rcnn/src/FasterRcnn/roi_align.py +++ b/model_zoo/official/cv/faster_rcnn/src/FasterRcnn/roi_align.py @@ -1,4 +1,4 @@ -# Copyright 2020 Huawei Technologies Co., Ltd +# Copyright 2020-2021 Huawei Technologies Co., Ltd # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. @@ -84,8 +84,8 @@ class SingleRoIExtractor(nn.Cell): self.out_channels = out_channels self.featmap_strides = featmap_strides self.num_levels = len(self.featmap_strides) - self.out_size = roi_layer['out_size'] - self.sample_num = roi_layer['sample_num'] + self.out_size = config.roi_layer.out_size + self.sample_num = config.roi_layer.sample_num self.roi_layers = self.build_roi_layers(self.featmap_strides) self.roi_layers = L.CellList(self.roi_layers) diff --git a/model_zoo/official/cv/faster_rcnn/src/config.py b/model_zoo/official/cv/faster_rcnn/src/config.py index fbbc14ada93..0826c7871b7 100644 --- a/model_zoo/official/cv/faster_rcnn/src/config.py +++ b/model_zoo/official/cv/faster_rcnn/src/config.py @@ -11,147 +11,51 @@ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. -# =========================================================================== -""" -network config setting, will be used in train.py and eval.py -""" -from easydict import EasyDict as ed +# ============================================================================ -config = ed({ - "img_width": 1280, - "img_height": 768, - "keep_ratio": True, - "flip_ratio": 0.5, - "expand_ratio": 1.0, +"""Parse arguments""" - # anchor - "feature_shapes": [(192, 320), (96, 160), (48, 80), (24, 40), (12, 20)], - "anchor_scales": [8], - "anchor_ratios": [0.5, 1.0, 2.0], - "anchor_strides": [4, 8, 16, 32, 64], - "num_anchors": 3, +from pprint import pprint, pformat +import yaml - # resnet - "resnet_block": [3, 4, 6, 3], - "resnet_in_channels": [64, 256, 512, 1024], - "resnet_out_channels": [256, 512, 1024, 2048], +class Config: + """ + Configuration namespace. Convert dictionary to members. + """ + def __init__(self, cfg_dict): + for k, v in cfg_dict.items(): + if isinstance(v, (list, tuple)): + setattr(self, k, [Config(x) if isinstance(x, dict) else x for x in v]) + else: + setattr(self, k, Config(v) if isinstance(v, dict) else v) - # fpn - "fpn_in_channels": [256, 512, 1024, 2048], - "fpn_out_channels": 256, - "fpn_num_outs": 5, + def __str__(self): + return pformat(self.__dict__) - # rpn - "rpn_in_channels": 256, - "rpn_feat_channels": 256, - "rpn_loss_cls_weight": 1.0, - "rpn_loss_reg_weight": 1.0, - "rpn_cls_out_channels": 1, - "rpn_target_means": [0., 0., 0., 0.], - "rpn_target_stds": [1.0, 1.0, 1.0, 1.0], + def __repr__(self): + return self.__str__() - # bbox_assign_sampler - "neg_iou_thr": 0.3, - "pos_iou_thr": 0.7, - "min_pos_iou": 0.3, - "num_bboxes": 245520, - "num_gts": 128, - "num_expected_neg": 256, - "num_expected_pos": 128, +def parse_yaml(yaml_path): + """ + Parse the yaml config file. - # proposal - "activate_num_classes": 2, - "use_sigmoid_cls": True, + Args: + yaml_path: Path to the yaml config. + """ + with open(yaml_path, 'r') as fin: + try: + cfgs = yaml.load_all(fin.read(), Loader=yaml.FullLoader) + cfgs = [x for x in cfgs] + if len(cfgs) == 1: + cfg = cfgs[0] + except: + raise ValueError("Failed to parse yaml") + return cfg - # roi_align - "roi_layer": dict(type='RoIAlign', out_size=7, sample_num=2), - "roi_align_out_channels": 256, - "roi_align_featmap_strides": [4, 8, 16, 32], - "roi_align_finest_scale": 56, - "roi_sample_num": 640, - - # bbox_assign_sampler_stage2 - "neg_iou_thr_stage2": 0.5, - "pos_iou_thr_stage2": 0.5, - "min_pos_iou_stage2": 0.5, - "num_bboxes_stage2": 2000, - "num_expected_pos_stage2": 128, - "num_expected_neg_stage2": 512, - "num_expected_total_stage2": 512, - - # rcnn - "rcnn_num_layers": 2, - "rcnn_in_channels": 256, - "rcnn_fc_out_channels": 1024, - "rcnn_loss_cls_weight": 1, - "rcnn_loss_reg_weight": 1, - "rcnn_target_means": [0., 0., 0., 0.], - "rcnn_target_stds": [0.1, 0.1, 0.2, 0.2], - - # train proposal - "rpn_proposal_nms_across_levels": False, - "rpn_proposal_nms_pre": 2000, - "rpn_proposal_nms_post": 2000, - "rpn_proposal_max_num": 2000, - "rpn_proposal_nms_thr": 0.7, - "rpn_proposal_min_bbox_size": 0, - - # test proposal - "rpn_nms_across_levels": False, - "rpn_nms_pre": 1000, - "rpn_nms_post": 1000, - "rpn_max_num": 1000, - "rpn_nms_thr": 0.7, - "rpn_min_bbox_min_size": 0, - "test_score_thr": 0.05, - "test_iou_thr": 0.5, - "test_max_per_img": 100, - "test_batch_size": 2, - - "rpn_head_use_sigmoid": True, - "rpn_head_weight": 1.0, - - # LR - "base_lr": 0.04, - "warmup_step": 500, - "warmup_ratio": 1/16.0, - "sgd_step": [8, 11], - "sgd_momentum": 0.9, - - # train - "batch_size": 2, - "loss_scale": 256, - "momentum": 0.91, - "weight_decay": 1e-5, - "epoch_size": 12, - "save_checkpoint": True, - "save_checkpoint_epochs": 1, - "keep_checkpoint_max": 10, - "save_checkpoint_path": "./", - - # Number of threads used to process the dataset in parallel - "num_parallel_workers": 8, - # Parallelize Python operations with multiple worker processes - "python_multiprocessing": True, - "mindrecord_dir": "../MindRecord_COCO_TRAIN", - "coco_root": "./cocodataset/", - "train_data_type": "train2017", - "val_data_type": "val2017", - "instance_set": "annotations/instances_{}.json", - "coco_classes": ('background', 'person', 'bicycle', 'car', 'motorcycle', 'airplane', 'bus', - 'train', 'truck', 'boat', 'traffic light', 'fire hydrant', - 'stop sign', 'parking meter', 'bench', 'bird', 'cat', 'dog', - 'horse', 'sheep', 'cow', 'elephant', 'bear', 'zebra', - 'giraffe', 'backpack', 'umbrella', 'handbag', 'tie', - 'suitcase', 'frisbee', 'skis', 'snowboard', 'sports ball', - 'kite', 'baseball bat', 'baseball glove', 'skateboard', - 'surfboard', 'tennis racket', 'bottle', 'wine glass', 'cup', - 'fork', 'knife', 'spoon', 'bowl', 'banana', 'apple', - 'sandwich', 'orange', 'broccoli', 'carrot', 'hot dog', 'pizza', - 'donut', 'cake', 'chair', 'couch', 'potted plant', 'bed', - 'dining table', 'toilet', 'tv', 'laptop', 'mouse', 'remote', - 'keyboard', 'cell phone', 'microwave', 'oven', 'toaster', 'sink', - 'refrigerator', 'book', 'clock', 'vase', 'scissors', - 'teddy bear', 'hair drier', 'toothbrush'), - "num_classes": 81 -}) +def get_config(config_path): + """ + Get Config according to the yaml file and cli arguments. + """ + default = parse_yaml(config_path) + pprint(default) + return Config(default) diff --git a/model_zoo/official/cv/faster_rcnn/src/config_101.yaml b/model_zoo/official/cv/faster_rcnn/src/config_101.yaml new file mode 100644 index 00000000000..433770248a7 --- /dev/null +++ b/model_zoo/official/cv/faster_rcnn/src/config_101.yaml @@ -0,0 +1,156 @@ +# Copyright 2020-2021 Huawei Technologies Co., Ltd +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# =========================================================================== + +img_width: 1280 +img_height: 768 +keep_ratio: True +flip_ratio: 0.5 +expand_ratio: 1.0 + +# anchor +feature_shapes: +- [192, 320] +- [96, 160] +- [48, 80] +- [24, 40] +- [12, 20] +anchor_scales: [8] +anchor_ratios: [0.5, 1.0, 2.0] +anchor_strides: [4, 8, 16, 32, 64] +num_anchors: 3 + +# resnet +resnet_block: [3, 4, 23, 3] +resnet_in_channels: [64, 256, 512, 1024] +resnet_out_channels: [256, 512, 1024, 2048] + +# fpn +fpn_in_channels: [256, 512, 1024, 2048] +fpn_out_channels: 256 +fpn_num_outs: 5 + +# rpn +rpn_in_channels: 256 +rpn_feat_channels: 256 +rpn_loss_cls_weight: 1.0 +rpn_loss_reg_weight: 1.0 +rpn_cls_out_channels: 1 +rpn_target_means: [0., 0., 0., 0.] +rpn_target_stds: [1.0, 1.0, 1.0, 1.0] + +# bbox_assign_sampler +neg_iou_thr: 0.3 +pos_iou_thr: 0.7 +min_pos_iou: 0.3 +num_bboxes: 245520 +num_gts: 128 +num_expected_neg: 256 +num_expected_pos: 128 + +# proposal +activate_num_classes: 2 +use_sigmoid_cls: True + +# roi_align +roi_layer: {type: 'RoIAlign', out_size: 7, sample_num: 2} +roi_align_out_channels: 256 +roi_align_featmap_strides: [4, 8, 16, 32] +roi_align_finest_scale: 56 +roi_sample_num: 640 + +# bbox_assign_sampler_stage2 +neg_iou_thr_stage2: 0.5 +pos_iou_thr_stage2: 0.5 +min_pos_iou_stage2: 0.5 +num_bboxes_stage2: 2000 +num_expected_pos_stage2: 128 +num_expected_neg_stage2: 512 +num_expected_total_stage2: 512 + +# rcnn +rcnn_num_layers: 2 +rcnn_in_channels: 256 +rcnn_fc_out_channels: 1024 +rcnn_loss_cls_weight: 1 +rcnn_loss_reg_weight: 1 +rcnn_target_means: [0., 0., 0., 0.] +rcnn_target_stds: [0.1, 0.1, 0.2, 0.2] + +# train proposal +rpn_proposal_nms_across_levels: False +rpn_proposal_nms_pre: 2000 +rpn_proposal_nms_post: 2000 +rpn_proposal_max_num: 2000 +rpn_proposal_nms_thr: 0.7 +rpn_proposal_min_bbox_size: 0 + +# test proposal +rpn_nms_across_levels: False +rpn_nms_pre: 1000 +rpn_nms_post: 1000 +rpn_max_num: 1000 +rpn_nms_thr: 0.7 +rpn_min_bbox_min_size: 0 +test_score_thr: 0.05 +test_iou_thr: 0.5 +test_max_per_img: 100 +test_batch_size: 2 + +rpn_head_use_sigmoid: True +rpn_head_weight: 1.0 + +# LR +base_lr: 0.02 +warmup_step: 500 +warmup_ratio: 0.0625 +sgd_step: [8, 11] +sgd_momentum: 0.9 + +# train +batch_size: 2 +loss_scale: 256 +momentum: 0.91 +weight_decay: 0.00001 +epoch_size: 20 +save_checkpoint: True +save_checkpoint_epochs: 1 +keep_checkpoint_max: 20 +save_checkpoint_path: "./" + +# Number of threads used to process the dataset in parallel +num_parallel_workers: 8 +# Parallelize Python operations with multiple worker processes +python_multiprocessing: True +mindrecord_dir: "/disk2/dataset/COCO2017/MindRecord_COCO_TRAIN" +coco_root: "/disk2/dataset/COCO2017" +train_data_type: "train2017" +val_data_type: "val2017" +instance_set: "annotations/instances_{}.json" +coco_classes: ['background', 'person', 'bicycle', 'car', 'motorcycle', 'airplane', 'bus', + 'train', 'truck', 'boat', 'traffic light', 'fire hydrant', + 'stop sign', 'parking meter', 'bench', 'bird', 'cat', 'dog', + 'horse', 'sheep', 'cow', 'elephant', 'bear', 'zebra', + 'giraffe', 'backpack', 'umbrella', 'handbag', 'tie', + 'suitcase', 'frisbee', 'skis', 'snowboard', 'sports ball', + 'kite', 'baseball bat', 'baseball glove', 'skateboard', + 'surfboard', 'tennis racket', 'bottle', 'wine glass', 'cup', + 'fork', 'knife', 'spoon', 'bowl', 'banana', 'apple', + 'sandwich', 'orange', 'broccoli', 'carrot', 'hot dog', 'pizza', + 'donut', 'cake', 'chair', 'couch', 'potted plant', 'bed', + 'dining table', 'toilet', 'tv', 'laptop', 'mouse', 'remote', + 'keyboard', 'cell phone', 'microwave', 'oven', 'toaster', 'sink', + 'refrigerator', 'book', 'clock', 'vase', 'scissors', + 'teddy bear', 'hair drier', 'toothbrush'] +num_classes: 81 diff --git a/model_zoo/official/cv/faster_rcnn/src/config_152.yaml b/model_zoo/official/cv/faster_rcnn/src/config_152.yaml new file mode 100644 index 00000000000..75c1a7e3c7f --- /dev/null +++ b/model_zoo/official/cv/faster_rcnn/src/config_152.yaml @@ -0,0 +1,156 @@ +# Copyright 2020-2021 Huawei Technologies Co., Ltd +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# =========================================================================== + +img_width: 1280 +img_height: 768 +keep_ratio: True +flip_ratio: 0.5 +expand_ratio: 1.0 + +# anchor +feature_shapes: +- [192, 320] +- [96, 160] +- [48, 80] +- [24, 40] +- [12, 20] +anchor_scales: [8] +anchor_ratios: [0.5, 1.0, 2.0] +anchor_strides: [4, 8, 16, 32, 64] +num_anchors: 3 + +# resnet +resnet_block: [3, 8, 36, 3] +resnet_in_channels: [64, 256, 512, 1024] +resnet_out_channels: [256, 512, 1024, 2048] + +# fpn +fpn_in_channels: [256, 512, 1024, 2048] +fpn_out_channels: 256 +fpn_num_outs: 5 + +# rpn +rpn_in_channels: 256 +rpn_feat_channels: 256 +rpn_loss_cls_weight: 1.0 +rpn_loss_reg_weight: 1.0 +rpn_cls_out_channels: 1 +rpn_target_means: [0., 0., 0., 0.] +rpn_target_stds: [1.0, 1.0, 1.0, 1.0] + +# bbox_assign_sampler +neg_iou_thr: 0.3 +pos_iou_thr: 0.7 +min_pos_iou: 0.3 +num_bboxes: 245520 +num_gts: 128 +num_expected_neg: 256 +num_expected_pos: 128 + +# proposal +activate_num_classes: 2 +use_sigmoid_cls: True + +# roi_align +roi_layer: {type: 'RoIAlign', out_size: 7, sample_num: 2} +roi_align_out_channels: 256 +roi_align_featmap_strides: [4, 8, 16, 32] +roi_align_finest_scale: 56 +roi_sample_num: 640 + +# bbox_assign_sampler_stage2 +neg_iou_thr_stage2: 0.5 +pos_iou_thr_stage2: 0.5 +min_pos_iou_stage2: 0.5 +num_bboxes_stage2: 2000 +num_expected_pos_stage2: 128 +num_expected_neg_stage2: 512 +num_expected_total_stage2: 512 + +# rcnn +rcnn_num_layers: 2 +rcnn_in_channels: 256 +rcnn_fc_out_channels: 1024 +rcnn_loss_cls_weight: 1 +rcnn_loss_reg_weight: 1 +rcnn_target_means: [0., 0., 0., 0.] +rcnn_target_stds: [0.1, 0.1, 0.2, 0.2] + +# train proposal +rpn_proposal_nms_across_levels: False +rpn_proposal_nms_pre: 2000 +rpn_proposal_nms_post: 2000 +rpn_proposal_max_num: 2000 +rpn_proposal_nms_thr: 0.7 +rpn_proposal_min_bbox_size: 0 + +# test proposal +rpn_nms_across_levels: False +rpn_nms_pre: 1000 +rpn_nms_post: 1000 +rpn_max_num: 1000 +rpn_nms_thr: 0.7 +rpn_min_bbox_min_size: 0 +test_score_thr: 0.05 +test_iou_thr: 0.5 +test_max_per_img: 100 +test_batch_size: 2 + +rpn_head_use_sigmoid: True +rpn_head_weight: 1.0 + +# LR +base_lr: 0.02 +warmup_step: 500 +warmup_ratio: 0.0625 +sgd_step: [8, 11] +sgd_momentum: 0.9 + +# train +batch_size: 2 +loss_scale: 256 +momentum: 0.91 +weight_decay: 0.00001 +epoch_size: 20 +save_checkpoint: True +save_checkpoint_epochs: 1 +keep_checkpoint_max: 20 +save_checkpoint_path: "./" + +# Number of threads used to process the dataset in parallel +num_parallel_workers: 8 +# Parallelize Python operations with multiple worker processes +python_multiprocessing: True +mindrecord_dir: "/disk2/dataset/COCO2017/MindRecord_COCO_TRAIN" +coco_root: "/disk2/dataset/COCO2017" +train_data_type: "train2017" +val_data_type: "val2017" +instance_set: "annotations/instances_{}.json" +coco_classes: ['background', 'person', 'bicycle', 'car', 'motorcycle', 'airplane', 'bus', + 'train', 'truck', 'boat', 'traffic light', 'fire hydrant', + 'stop sign', 'parking meter', 'bench', 'bird', 'cat', 'dog', + 'horse', 'sheep', 'cow', 'elephant', 'bear', 'zebra', + 'giraffe', 'backpack', 'umbrella', 'handbag', 'tie', + 'suitcase', 'frisbee', 'skis', 'snowboard', 'sports ball', + 'kite', 'baseball bat', 'baseball glove', 'skateboard', + 'surfboard', 'tennis racket', 'bottle', 'wine glass', 'cup', + 'fork', 'knife', 'spoon', 'bowl', 'banana', 'apple', + 'sandwich', 'orange', 'broccoli', 'carrot', 'hot dog', 'pizza', + 'donut', 'cake', 'chair', 'couch', 'potted plant', 'bed', + 'dining table', 'toilet', 'tv', 'laptop', 'mouse', 'remote', + 'keyboard', 'cell phone', 'microwave', 'oven', 'toaster', 'sink', + 'refrigerator', 'book', 'clock', 'vase', 'scissors', + 'teddy bear', 'hair drier', 'toothbrush'] +num_classes: 81 \ No newline at end of file diff --git a/model_zoo/official/cv/faster_rcnn/src/config_50.yaml b/model_zoo/official/cv/faster_rcnn/src/config_50.yaml new file mode 100644 index 00000000000..72221474d1c --- /dev/null +++ b/model_zoo/official/cv/faster_rcnn/src/config_50.yaml @@ -0,0 +1,156 @@ +# Copyright 2020-2021 Huawei Technologies Co., Ltd +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# =========================================================================== + +img_width: 1280 +img_height: 768 +keep_ratio: True +flip_ratio: 0.5 +expand_ratio: 1.0 + +# anchor +feature_shapes: +- [192, 320] +- [96, 160] +- [48, 80] +- [24, 40] +- [12, 20] +anchor_scales: [8] +anchor_ratios: [0.5, 1.0, 2.0] +anchor_strides: [4, 8, 16, 32, 64] +num_anchors: 3 + +# resnet +resnet_block: [3, 4, 6, 3] +resnet_in_channels: [64, 256, 512, 1024] +resnet_out_channels: [256, 512, 1024, 2048] + +# fpn +fpn_in_channels: [256, 512, 1024, 2048] +fpn_out_channels: 256 +fpn_num_outs: 5 + +# rpn +rpn_in_channels: 256 +rpn_feat_channels: 256 +rpn_loss_cls_weight: 1.0 +rpn_loss_reg_weight: 1.0 +rpn_cls_out_channels: 1 +rpn_target_means: [0., 0., 0., 0.] +rpn_target_stds: [1.0, 1.0, 1.0, 1.0] + +# bbox_assign_sampler +neg_iou_thr: 0.3 +pos_iou_thr: 0.7 +min_pos_iou: 0.3 +num_bboxes: 245520 +num_gts: 128 +num_expected_neg: 256 +num_expected_pos: 128 + +# proposal +activate_num_classes: 2 +use_sigmoid_cls: True + +# roi_align +roi_layer: {type: 'RoIAlign', out_size: 7, sample_num: 2} +roi_align_out_channels: 256 +roi_align_featmap_strides: [4, 8, 16, 32] +roi_align_finest_scale: 56 +roi_sample_num: 640 + +# bbox_assign_sampler_stage2 +neg_iou_thr_stage2: 0.5 +pos_iou_thr_stage2: 0.5 +min_pos_iou_stage2: 0.5 +num_bboxes_stage2: 2000 +num_expected_pos_stage2: 128 +num_expected_neg_stage2: 512 +num_expected_total_stage2: 512 + +# rcnn +rcnn_num_layers: 2 +rcnn_in_channels: 256 +rcnn_fc_out_channels: 1024 +rcnn_loss_cls_weight: 1 +rcnn_loss_reg_weight: 1 +rcnn_target_means: [0., 0., 0., 0.] +rcnn_target_stds: [0.1, 0.1, 0.2, 0.2] + +# train proposal +rpn_proposal_nms_across_levels: False +rpn_proposal_nms_pre: 2000 +rpn_proposal_nms_post: 2000 +rpn_proposal_max_num: 2000 +rpn_proposal_nms_thr: 0.7 +rpn_proposal_min_bbox_size: 0 + +# test proposal +rpn_nms_across_levels: False +rpn_nms_pre: 1000 +rpn_nms_post: 1000 +rpn_max_num: 1000 +rpn_nms_thr: 0.7 +rpn_min_bbox_min_size: 0 +test_score_thr: 0.05 +test_iou_thr: 0.5 +test_max_per_img: 100 +test_batch_size: 2 + +rpn_head_use_sigmoid: True +rpn_head_weight: 1.0 + +# LR +base_lr: 0.04 +warmup_step: 500 +warmup_ratio: 0.0625 +sgd_step: [8, 11] +sgd_momentum: 0.9 + +# train +batch_size: 2 +loss_scale: 256 +momentum: 0.91 +weight_decay: 0.00001 +epoch_size: 20 +save_checkpoint: True +save_checkpoint_epochs: 1 +keep_checkpoint_max: 20 +save_checkpoint_path: "./" + +# Number of threads used to process the dataset in parallel +num_parallel_workers: 8 +# Parallelize Python operations with multiple worker processes +python_multiprocessing: True +mindrecord_dir: "/disk2/dataset/COCO2017/MindRecord_COCO_TRAIN" +coco_root: "/disk2/dataset/COCO2017" +train_data_type: "train2017" +val_data_type: "val2017" +instance_set: "annotations/instances_{}.json" +coco_classes: ['background', 'person', 'bicycle', 'car', 'motorcycle', 'airplane', 'bus', + 'train', 'truck', 'boat', 'traffic light', 'fire hydrant', + 'stop sign', 'parking meter', 'bench', 'bird', 'cat', 'dog', + 'horse', 'sheep', 'cow', 'elephant', 'bear', 'zebra', + 'giraffe', 'backpack', 'umbrella', 'handbag', 'tie', + 'suitcase', 'frisbee', 'skis', 'snowboard', 'sports ball', + 'kite', 'baseball bat', 'baseball glove', 'skateboard', + 'surfboard', 'tennis racket', 'bottle', 'wine glass', 'cup', + 'fork', 'knife', 'spoon', 'bowl', 'banana', 'apple', + 'sandwich', 'orange', 'broccoli', 'carrot', 'hot dog', 'pizza', + 'donut', 'cake', 'chair', 'couch', 'potted plant', 'bed', + 'dining table', 'toilet', 'tv', 'laptop', 'mouse', 'remote', + 'keyboard', 'cell phone', 'microwave', 'oven', 'toaster', 'sink', + 'refrigerator', 'book', 'clock', 'vase', 'scissors', + 'teddy bear', 'hair drier', 'toothbrush'] +num_classes: 81 \ No newline at end of file diff --git a/model_zoo/official/cv/faster_rcnn/src/convert_checkpoint.py b/model_zoo/official/cv/faster_rcnn/src/convert_checkpoint.py index dab0974e2c6..f260686d4c3 100644 --- a/model_zoo/official/cv/faster_rcnn/src/convert_checkpoint.py +++ b/model_zoo/official/cv/faster_rcnn/src/convert_checkpoint.py @@ -1,4 +1,4 @@ -# Copyright 2021 Huawei Technologies Co., Ltd +# Copyright 2020-2021 Huawei Technologies Co., Ltd # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. @@ -13,7 +13,7 @@ # limitations under the License. # =========================================================================== """ -convert resnet50 pretrain model to faster_rcnn backbone pretrain model +convert resnet pretrain model to faster_rcnn backbone pretrain model """ import argparse from mindspore.train.serialization import load_checkpoint, save_checkpoint @@ -26,10 +26,10 @@ parser.add_argument('--ckpt_file', type=str, default='', help='ckpt file path') args_opt = parser.parse_args() def load_weights(model_path, use_fp16_weight): """ - load resnet50 pretrain checkpoint file. + load resnet pretrain checkpoint file. Args: - model_path (str): resnet50 pretrain checkpoint file . + model_path (str): resnet pretrain checkpoint file . use_fp16_weight(bool): whether save weight into float16. Returns: @@ -61,4 +61,4 @@ def load_weights(model_path, use_fp16_weight): if __name__ == "__main__": parameter_list = load_weights(args_opt.ckpt_file, use_fp16_weight=False) - save_checkpoint(parameter_list, "resnet50_backbone.ckpt") + save_checkpoint(parameter_list, "resnet_backbone.ckpt") diff --git a/model_zoo/official/cv/faster_rcnn/src/dataset.py b/model_zoo/official/cv/faster_rcnn/src/dataset.py index acea7efe2b8..58940b89a2a 100644 --- a/model_zoo/official/cv/faster_rcnn/src/dataset.py +++ b/model_zoo/official/cv/faster_rcnn/src/dataset.py @@ -25,8 +25,6 @@ import mmcv import mindspore.dataset as de import mindspore.dataset.vision.c_transforms as C from mindspore.mindrecord import FileWriter -from src.config import config - def bbox_overlaps(bboxes1, bboxes2, mode='iou'): """Calculate the ious between each bbox of bboxes1 and bboxes2. @@ -160,7 +158,7 @@ class Expand: return img, boxes, labels -def rescale_column(img, img_shape, gt_bboxes, gt_label, gt_num): +def rescale_column(img, img_shape, gt_bboxes, gt_label, gt_num, config): """rescale operation for image""" img_data, scale_factor = mmcv.imrescale(img, (config.img_width, config.img_height), return_scale=True) if img_data.shape[0] > config.img_height: @@ -183,7 +181,7 @@ def rescale_column(img, img_shape, gt_bboxes, gt_label, gt_num): return (pad_img_data, img_shape, gt_bboxes, gt_label, gt_num) -def rescale_column_test(img, img_shape, gt_bboxes, gt_label, gt_num): +def rescale_column_test(img, img_shape, gt_bboxes, gt_label, gt_num, config): """rescale operation for image of eval""" img_data, scale_factor = mmcv.imrescale(img, (config.img_width, config.img_height), return_scale=True) if img_data.shape[0] > config.img_height: @@ -203,7 +201,7 @@ def rescale_column_test(img, img_shape, gt_bboxes, gt_label, gt_num): return (pad_img_data, img_shape, gt_bboxes, gt_label, gt_num) -def resize_column(img, img_shape, gt_bboxes, gt_label, gt_num): +def resize_column(img, img_shape, gt_bboxes, gt_label, gt_num, config): """resize operation for image""" img_data = img img_data, w_scale, h_scale = mmcv.imresize( @@ -221,7 +219,7 @@ def resize_column(img, img_shape, gt_bboxes, gt_label, gt_num): return (img_data, img_shape, gt_bboxes, gt_label, gt_num) -def resize_column_test(img, img_shape, gt_bboxes, gt_label, gt_num): +def resize_column_test(img, img_shape, gt_bboxes, gt_label, gt_num, config): """resize operation for image of eval""" img_data = img img_data, w_scale, h_scale = mmcv.imresize( @@ -239,7 +237,7 @@ def resize_column_test(img, img_shape, gt_bboxes, gt_label, gt_num): return (img_data, img_shape, gt_bboxes, gt_label, gt_num) -def impad_to_multiple_column(img, img_shape, gt_bboxes, gt_label, gt_num): +def impad_to_multiple_column(img, img_shape, gt_bboxes, gt_label, gt_num, config): """impad operation for image""" img_data = mmcv.impad(img, (config.img_height, config.img_width)) img_data = img_data.astype(np.float32) @@ -294,16 +292,16 @@ def expand_column(img, img_shape, gt_bboxes, gt_label, gt_num): return (img, img_shape, gt_bboxes, gt_label, gt_num) -def preprocess_fn(image, box, is_training): +def preprocess_fn(image, box, is_training, config): """Preprocess function for dataset.""" def _infer_data(image_bgr, image_shape, gt_box_new, gt_label_new, gt_iscrowd_new_revert): image_shape = image_shape[:2] input_data = image_bgr, image_shape, gt_box_new, gt_label_new, gt_iscrowd_new_revert if config.keep_ratio: - input_data = rescale_column_test(*input_data) + input_data = rescale_column_test(*input_data, config=config) else: - input_data = resize_column_test(*input_data) + input_data = resize_column_test(*input_data, config=config) input_data = imnormalize_column(*input_data) output_data = transpose_column(*input_data) @@ -336,9 +334,9 @@ def preprocess_fn(image, box, is_training): if expand: input_data = expand_column(*input_data) if config.keep_ratio: - input_data = rescale_column(*input_data) + input_data = rescale_column(*input_data, config=config) else: - input_data = resize_column(*input_data) + input_data = resize_column(*input_data, config=config) input_data = imnormalize_column(*input_data) if flip: input_data = flip_column(*input_data) @@ -349,7 +347,7 @@ def preprocess_fn(image, box, is_training): return _data_aug(image, box, is_training) -def create_coco_label(is_training): +def create_coco_label(is_training, config): """Get image path and annotation from COCO.""" from pycocotools.coco import COCO @@ -431,13 +429,13 @@ def filter_valid_data(image_dir, anno_path): return image_files, image_anno_dict -def data_to_mindrecord_byte_image(dataset="coco", is_training=True, prefix="fasterrcnn.mindrecord", file_num=8): +def data_to_mindrecord_byte_image(config, dataset="coco", is_training=True, prefix="fasterrcnn.mindrecord", file_num=8): """Create MindRecord file.""" mindrecord_dir = config.mindrecord_dir mindrecord_path = os.path.join(mindrecord_dir, prefix) writer = FileWriter(mindrecord_path, file_num) if dataset == "coco": - image_files, image_anno_dict = create_coco_label(is_training) + image_files, image_anno_dict = create_coco_label(is_training, config=config) else: image_files, image_anno_dict = filter_valid_data(config.IMAGE_DIR, config.ANNO_PATH) @@ -456,7 +454,7 @@ def data_to_mindrecord_byte_image(dataset="coco", is_training=True, prefix="fast writer.commit() -def create_fasterrcnn_dataset(mindrecord_file, batch_size=2, device_num=1, rank_id=0, is_training=True, +def create_fasterrcnn_dataset(config, mindrecord_file, batch_size=2, device_num=1, rank_id=0, is_training=True, num_parallel_workers=8, python_multiprocessing=False): """Create FasterRcnn dataset with MindDataset.""" cv2.setNumThreads(0) @@ -465,7 +463,7 @@ def create_fasterrcnn_dataset(mindrecord_file, batch_size=2, device_num=1, rank_ num_parallel_workers=4, shuffle=is_training) decode = C.Decode() ds = ds.map(input_columns=["image"], operations=decode) - compose_map_func = (lambda image, annotation: preprocess_fn(image, annotation, is_training)) + compose_map_func = (lambda image, annotation: preprocess_fn(image, annotation, is_training, config=config)) if is_training: ds = ds.map(input_columns=["image", "annotation"], diff --git a/model_zoo/official/cv/faster_rcnn/src/util.py b/model_zoo/official/cv/faster_rcnn/src/util.py index fdd1ab04895..b53315093f2 100644 --- a/model_zoo/official/cv/faster_rcnn/src/util.py +++ b/model_zoo/official/cv/faster_rcnn/src/util.py @@ -1,4 +1,4 @@ -# Copyright 2020 Huawei Technologies Co., Ltd +# Copyright 2020-2021 Huawei Technologies Co., Ltd # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. diff --git a/model_zoo/official/cv/faster_rcnn/train.py b/model_zoo/official/cv/faster_rcnn/train.py index 4856f3e6e95..003ca7dd268 100644 --- a/model_zoo/official/cv/faster_rcnn/train.py +++ b/model_zoo/official/cv/faster_rcnn/train.py @@ -31,11 +31,10 @@ from mindspore.train.serialization import load_checkpoint, load_param_into_net from mindspore.nn import SGD from mindspore.common import set_seed -from src.FasterRcnn.faster_rcnn_r50 import Faster_Rcnn_Resnet50 from src.network_define import LossCallBack, WithLossCell, TrainOneStepCell, LossNet -from src.config import config from src.dataset import data_to_mindrecord_byte_image, create_fasterrcnn_dataset from src.lr_schedule import dynamic_lr +import src.config as cfg set_seed(1) @@ -48,10 +47,25 @@ parser.add_argument("--device_target", type=str, default="Ascend", parser.add_argument("--device_id", type=int, default=0, help="Device id, default: 0.") parser.add_argument("--device_num", type=int, default=1, help="Use device nums, default: 1.") parser.add_argument("--rank_id", type=int, default=0, help="Rank id, default: 0.") +parser.add_argument("--backbone", type=str, required=True, \ + help="backbone network name, options:resnet_v1_50, resnet_v1.5_50, resnet_v1_101, resnet_v1_152") args_opt = parser.parse_args() context.set_context(mode=context.GRAPH_MODE, device_target=args_opt.device_target, device_id=args_opt.device_id) +if args_opt.backbone in ("resnet_v1.5_50", "resnet_v1_101", "resnet_v1_152"): + from src.FasterRcnn.faster_rcnn_resnet import Faster_Rcnn_Resnet + if args_opt.backbone == "resnet_v1.5_50": + config = cfg.get_config("./src/config_50.yaml") + elif args_opt.backbone == "resnet_v1_101": + config = cfg.get_config("./src/config_101.yaml") + elif args_opt.backbone == "resnet_v1_152": + config = cfg.get_config("./src/config_152.yaml") + +elif args_opt.backbone == "resnet_v1_50": + config = cfg.get_config("./src/config_50.yaml") + from src.FasterRcnn.faster_rcnn_resnet50v1 import Faster_Rcnn_Resnet + if __name__ == '__main__': if args_opt.device_target == "GPU": context.set_context(enable_graph_kernel=True) @@ -91,7 +105,7 @@ if __name__ == '__main__': print("Please make sure config:coco_root is valid.") raise ValueError(config.coco_root) print("Create Mindrecord. It may take some time.") - data_to_mindrecord_byte_image("coco", True, prefix) + data_to_mindrecord_byte_image(config, "coco", True, prefix) print("Create Mindrecord Done, at {}".format(mindrecord_dir)) else: print("coco_root not exits.") @@ -101,7 +115,7 @@ if __name__ == '__main__': print("Please make sure config:image_dir is valid.") raise ValueError(config.image_dir) print("Create Mindrecord. It may take some time.") - data_to_mindrecord_byte_image("other", True, prefix) + data_to_mindrecord_byte_image(config, "other", True, prefix) print("Create Mindrecord Done, at {}".format(mindrecord_dir)) else: print("image_dir or anno_path not exits.") @@ -114,7 +128,7 @@ if __name__ == '__main__': loss_scale = float(config.loss_scale) # When create MindDataset, using the fitst mindrecord file, such as FasterRcnn.mindrecord0. - dataset = create_fasterrcnn_dataset(mindrecord_file, batch_size=config.batch_size, + dataset = create_fasterrcnn_dataset(config, mindrecord_file, batch_size=config.batch_size, device_num=device_num, rank_id=rank, num_parallel_workers=config.num_parallel_workers, python_multiprocessing=config.python_multiprocessing) @@ -122,7 +136,7 @@ if __name__ == '__main__': dataset_size = dataset.get_dataset_size() print("Create dataset done!") - net = Faster_Rcnn_Resnet50(config=config) + net = Faster_Rcnn_Resnet(config=config) net = net.set_train() load_path = args_opt.pre_trained