diff --git a/examples/yolov3/.gitignore b/examples/yolov3/.gitignore index 01cdc1cbde3507576115b3a9832be30158e56fe5..19ac4b2c67adf15aaa730c8b9823a2b5c2da8780 100644 --- a/examples/yolov3/.gitignore +++ b/examples/yolov3/.gitignore @@ -1 +1,2 @@ dataset/voc* +pretrain_weights/darknet53_pretrained.pdparams diff --git a/examples/yolov3/README.md b/examples/yolov3/README.md index cc6d302a544f9b8d9cd06fc363b90d053919a5e9..9a0d2cc15c4390f1fe666b776dfeadd9c1af558a 100644 --- a/examples/yolov3/README.md +++ b/examples/yolov3/README.md @@ -99,12 +99,17 @@ YOLOv3 的网络结构由基础特征提取网络、multi-scale特征融合层 | ... ``` +```bash +sh pretrain_weights/download.sh +``` + ### 模型训练 -数据准备完毕后，可使用`main.py`脚本启动训练和评估，如下脚本会自动每epoch交替进行训练和模型评估，并将checkpoint默认保存在`yolo_checkpoint`目录下。 +数据准备完成后，可使用`main.py`脚本启动训练和评估，如下脚本会自动每epoch交替进行训练和模型评估，并将checkpoint默认保存在`yolo_checkpoint`目录下。 YOLOv3模型训练总batch_size为64训练，以下以使用4卡Tesla P40每卡batch_size为16训练介绍训练方式。对于静态图和动态图，多卡训练中`--batch_size`为每卡上的batch_size，即总batch_size为`--batch_size`乘以卡数。 +YOLOv3模型训练须加载骨干网络[DarkNet53]()的预训练权重，可在训练时通过`--pretrain_weights`指定，若指定为URL，将自动下载权重至`~/.cache/paddle/weights`目录并加载。 `main.py`脚本参数可通过如下命令查询 @@ -117,7 +122,7 @@ python main.py --help 使用如下方式进行多卡训练: ```bash -CUDA_VISIBLE_DEVICES=0,1,2,3 python -m paddle.distributed.launch main.py --data= --batch_size=16 +CUDA_VISIBLE_DEVICES=0,1,2,3 python -m paddle.distributed.launch main.py --data= --batch_size=16 --pretrain_weights=https://paddlemodels.bj.bcebos.com/hapi/darknet53_pretrained.pdparams ``` #### 动态图训练 @@ -127,7 +132,7 @@ CUDA_VISIBLE_DEVICES=0,1,2,3 python -m paddle.distributed.launch main.py --data= 使用如下方式进行多卡训练: ```bash -CUDA_VISIBLE_DEVICES=0,1,2,3 python main.py -m paddle.distributed.launch --data= --batch_size=16 -d +CUDA_VISIBLE_DEVICES=0,1,2,3 python main.py -m paddle.distributed.launch --data= --batch_size=16 -d --pretrain_weights=https://paddlemodels.bj.bcebos.com/hapi/darknet53_pretrained.pdparams ``` diff --git a/examples/yolov3/main.py b/examples/yolov3/main.py index ebe85543712d82267eaded26da4c1db8800b735f..7203329c7f75c97c1646b25320a44adadbcfc5ac 100644 --- a/examples/yolov3/main.py +++ b/examples/yolov3/main.py @@ -27,6 +27,7 @@ from paddle.io import DataLoader from hapi.model import Model, Input, set_device from hapi.distributed import DistributedBatchSampler +from hapi.download import is_url, get_weights_path from hapi.datasets import COCODataset from hapi.vision.transforms import * from hapi.vision.models import yolov3_darknet53, YoloLoss @@ -124,8 +125,11 @@ def main(): model_mode='eval' if FLAGS.eval_only else 'train', pretrained=pretrained) - if FLAGS.pretrain_weights is not None: - model.load(FLAGS.pretrain_weights, skip_mismatch=True, reset_optimizer=True) + if FLAGS.pretrain_weights and not FLAGS.eval_only: + pretrain_weights = FLAGS.pretrain_weights + if is_url(pretrain_weights): + pretrain_weights = get_weights_path(pretrain_weights) + model.load(pretrain_weights, skip_mismatch=True, reset_optimizer=True) optim = make_optimizer(len(batch_sampler), parameter_list=model.parameters()) @@ -196,7 +200,8 @@ if __name__ == '__main__': parser.add_argument( "-j", "--num_workers", default=4, type=int, help="reader worker number") parser.add_argument( - "-p", "--pretrain_weights", default=None, type=str, + "-p", "--pretrain_weights", + default="./pretrain_weights/darknet53_pretrained", type=str, help="path to pretrained weights") parser.add_argument( "-r", "--resume", default=None, type=str, diff --git a/hapi/download.py b/hapi/download.py index 10d3fba390647c494448b83295901a8973d2aba8..e9a89ba53bc3bc74f03977659156121bce6db577 100644 --- a/hapi/download.py +++ b/hapi/download.py @@ -29,13 +29,22 @@ from paddle.fluid.dygraph.parallel import ParallelEnv import logging logger = logging.getLogger(__name__) -__all__ = ['get_weights_path'] +__all__ = ['get_weights_path', 'is_url'] WEIGHTS_HOME = osp.expanduser("~/.cache/paddle/hapi/weights") DOWNLOAD_RETRY_LIMIT = 3 +def is_url(path): + """ + Whether path is URL. + Args: + path (string): URL string or not. + """ + return path.startswith('http://') or path.startswith('https://') + + def get_weights_path(url, md5sum=None): """Get weights path from WEIGHT_HOME, if not exists, download it from url. @@ -62,6 +71,7 @@ def get_path(url, root_dir, md5sum=None, check_exist=True): WEIGHTS_HOME or DATASET_HOME md5sum (str): md5 sum of download package """ + assert is_url(url), "downloading from {} not a url".format(url) # parse path after download to decompress under root_dir fullpath = map_path(url, root_dir) diff --git a/hapi/model.py b/hapi/model.py index 3255e614fd80529cdd7ac17ca31604c6815a11c4..e2fa2d6e3ba900e783421d0e8b29fa9a3aad5813 100644 --- a/hapi/model.py +++ b/hapi/model.py @@ -798,6 +798,13 @@ class Model(fluid.dygraph.Layer): format(key, list(state.shape), list(param.shape))) return param, state + def _strip_postfix(path): + path, ext = os.path.splitext(path) + assert ext in ['', '.pdparams', '.pdopt', '.pdmodel'], \ + "Unknown postfix {} from weights".format(ext) + return path + + path = _strip_postfix(path) param_state = _load_state_from_path(path + ".pdparams") assert param_state, "Failed to load parameters, please check path." diff --git a/hapi/vision/models/darknet.py b/hapi/vision/models/darknet.py index 85e25f4e1205ea62ec878409d640ba42e7335ee2..027cb4f2094e37ede347c090c5849cfcbf00611e 100755 --- a/hapi/vision/models/darknet.py +++ b/hapi/vision/models/darknet.py @@ -136,7 +136,7 @@ class LayerWarp(fluid.dygraph.Layer): DarkNet_cfg = {53: ([1, 2, 8, 8, 4])} -class DarkNet(Model): +class DarkNet(fluid.dygraph.Layer): """DarkNet model from `"YOLOv3: An Incremental Improvement" `_ diff --git a/hapi/vision/models/yolov3.py b/hapi/vision/models/yolov3.py index 840a402e3e1b473cc3c3deffc1b2d0fe6e7a2307..bafe148b14b7e1baf8124c61969238e049e2d6f0 100644 --- a/hapi/vision/models/yolov3.py +++ b/hapi/vision/models/yolov3.py @@ -118,7 +118,7 @@ class YOLOv3(Model): self.nms_posk = 100 self.draw_thresh = 0.5 - self.backbone = darknet53(pretrained=(model_mode=='train')) + self.backbone = darknet53(pretrained=False) self.block_outputs = [] self.yolo_blocks = [] self.route_blocks = []