clean vision

examples/bmn/eval.py

@@ -19,7 +19,8 @@ import logging
 import paddle.fluid as fluid
 
 from hapi.model import set_device, Input
-from hapi.vision.models import bmn, BmnLoss
+
+from modeling import bmn, BmnLoss
 from bmn_metric import BmnMetric
 from reader import BmnDataset
 from config_utils import *

hapi/vision/models/bmn_model.py → examples/bmn/modeling.py

@@ -26,7 +26,7 @@ DATATYPE = 'float32'
 
 pretrain_infos = {
     'bmn': ('https://paddlemodels.bj.bcebos.com/hapi/bmn.pdparams',
-            '9286c821acc4cad46d6613b931ba468c')
+            'aa84e3386e1fbd117fb96fa572feeb94')
 }
 
 
@@ -462,5 +462,5 @@ def bmn(tscale,
         weight_path = get_weights_path(*(pretrain_infos['bmn']))
         assert weight_path.endswith('.pdparams'), \
                 "suffix of weight must be .pdparams"
-        model.load(weight_path[:-9])
+        model.load(weight_path)
     return model

examples/bmn/predict.py

@@ -19,7 +19,8 @@ import logging
 import paddle.fluid as fluid
 
 from hapi.model import set_device, Input
-from hapi.vision.models import bmn, BmnLoss
+
+from modeling import bmn, BmnLoss
 from bmn_metric import BmnMetric
 from reader import BmnDataset
 from config_utils import *

examples/bmn/train.py

@@ -19,9 +19,10 @@ import sys
 import os
 
 from hapi.model import set_device, Input
-from hapi.vision.models import bmn, BmnLoss
+
 from reader import BmnDataset
 from config_utils import *
+from modeling import bmn, BmnLoss
 
 DATATYPE = 'float32'
 

examples/tsm/infer.py

@@ -20,9 +20,9 @@ import argparse
 import numpy as np
 
 from hapi.model import Input, set_device
-from hapi.vision.models import tsm_resnet50
 
 from check import check_gpu, check_version
+from modeling import tsm_resnet50
 from kinetics_dataset import KineticsDataset
 from transforms import *
 

examples/tsm/main.py

@@ -24,8 +24,8 @@ from paddle.fluid.dygraph.parallel import ParallelEnv
 
 from hapi.model import Model, CrossEntropy, Input, set_device
 from hapi.metrics import Accuracy
-from hapi.vision.models import tsm_resnet50
 
+from modeling import tsm_resnet50
 from check import check_gpu, check_version
 from kinetics_dataset import KineticsDataset
 from transforms import *

hapi/vision/models/tsm.py → examples/tsm/modeling.py

@@ -196,7 +196,7 @@ def _tsm_resnet(num_layers, seg_num=8, num_classes=400, pretrained=True):
         weight_path = get_weights_path(*(pretrain_infos[num_layers]))
         assert weight_path.endswith('.pdparams'), \
                 "suffix of weight must be .pdparams"
-        model.load(weight_path[:-9])
+        model.load(weight_path)
     return model
 
 

examples/yolov3/README.md

@@ -99,18 +99,12 @@ 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
 
 ```bash
@@ -122,7 +116,7 @@ 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=<path/to/dataset> --batch_size=16 --pretrain_weights=https://paddlemodels.bj.bcebos.com/hapi/darknet53_pretrained.pdparams
+CUDA_VISIBLE_DEVICES=0,1,2,3 python -m paddle.distributed.launch main.py --data=<path/to/dataset> --batch_size=16
 ```
 
 #### 动态图训练
@@ -132,7 +126,7 @@ 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=<path/to/dataset> --batch_size=16 -d --pretrain_weights=https://paddlemodels.bj.bcebos.com/hapi/darknet53_pretrained.pdparams
+CUDA_VISIBLE_DEVICES=0,1,2,3 python main.py -m paddle.distributed.launch --data=<path/to/dataset> --batch_size=16 -d
 ```
 
 

examples/yolov3/infer.py

@@ -25,8 +25,9 @@ from paddle.fluid.optimizer import Momentum
 from paddle.io import DataLoader
 
 from hapi.model import Model, Input, set_device
-from hapi.vision.models import yolov3_darknet53, YoloLoss
-from hapi.vision.transforms import *
+
+from modeling import yolov3_darknet53, YoloLoss
+from transforms import *
 
 from visualizer import draw_bbox
 

examples/yolov3/main.py

@@ -27,12 +27,12 @@ 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
+from hapi.vision.transforms import Compose, BatchCompose
 
+from modeling import yolov3_darknet53, YoloLoss
+from coco import COCODataset
 from coco_metric import COCOMetric
+from transforms import *
 
 NUM_MAX_BOXES = 50
 
@@ -126,10 +126,7 @@ def main():
                    pretrained=pretrained)
 
     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)
+        model.load(FLAGS.pretrain_weights, skip_mismatch=True, reset_optimizer=True)
 
     optim = make_optimizer(len(batch_sampler), parameter_list=model.parameters())
 
@@ -168,7 +165,7 @@ def main():
               save_dir="yolo_checkpoint/mixup",
               save_freq=10)
 
-    # do not use image mixup transfrom in laste FLAGS.no_mixup_epoch epoches
+    # do not use image mixup transfrom in the last FLAGS.no_mixup_epoch epoches
     dataset.mixup = False
     model.fit(train_data=loader,
               epochs=FLAGS.no_mixup_epoch,
@@ -200,8 +197,7 @@ if __name__ == '__main__':
     parser.add_argument(
         "-j", "--num_workers", default=4, type=int, help="reader worker number")
     parser.add_argument(
-        "-p", "--pretrain_weights",
-        default="./pretrain_weights/darknet53_pretrained", type=str,
+        "-p", "--pretrain_weights", default=None, type=str,
         help="path to pretrained weights")
     parser.add_argument(
         "-r", "--resume", default=None, type=str,

hapi/vision/models/yolov3.py → examples/yolov3/modeling.py

@@ -16,13 +16,13 @@ from __future__ import division
 from __future__ import print_function
 
 import paddle.fluid as fluid
-from paddle.fluid.dygraph.nn import Conv2D
+from paddle.fluid.dygraph.nn import Conv2D, BatchNorm
 from paddle.fluid.param_attr import ParamAttr
 from paddle.fluid.regularizer import L2Decay
 
 from hapi.model import Model, Loss
 from hapi.download import get_weights_path
-from .darknet import darknet53, ConvBNLayer
+from hapi.vision.models import darknet53
 
 __all__ = ['YoloLoss', 'YOLOv3', 'yolov3_darknet53']
 
@@ -33,6 +33,46 @@ pretrain_infos = {
 }
 
 
+class ConvBNLayer(fluid.dygraph.Layer):
+    def __init__(self,
+                 ch_in,
+                 ch_out,
+                 filter_size=3,
+                 stride=1,
+                 groups=1,
+                 padding=0,
+                 act="leaky"):
+        super(ConvBNLayer, self).__init__()
+
+        self.conv = Conv2D(
+            num_channels=ch_in,
+            num_filters=ch_out,
+            filter_size=filter_size,
+            stride=stride,
+            padding=padding,
+            groups=groups,
+            param_attr=ParamAttr(
+                initializer=fluid.initializer.Normal(0., 0.02)),
+            bias_attr=False,
+            act=None)
+        self.batch_norm = BatchNorm(
+            num_channels=ch_out,
+            param_attr=ParamAttr(
+                initializer=fluid.initializer.Normal(0., 0.02),
+                regularizer=L2Decay(0.)),
+            bias_attr=ParamAttr(
+                initializer=fluid.initializer.Constant(0.0),
+                regularizer=L2Decay(0.)))
+
+        self.act = act
+
+    def forward(self, inputs):
+        out = self.conv(inputs)
+        out = self.batch_norm(out)
+        if self.act == 'leaky':
+            out = fluid.layers.leaky_relu(x=out, alpha=0.1)
+        return out
+
 class YoloDetectionBlock(fluid.dygraph.Layer):
     def __init__(self, ch_in, channel):
         super(YoloDetectionBlock, self).__init__()
@@ -118,7 +158,7 @@ class YOLOv3(Model):
         self.nms_posk = 100
         self.draw_thresh = 0.5
 
-        self.backbone = darknet53(pretrained=False)
+        self.backbone = darknet53(pretrained=(model_mode=='train'))
         self.block_outputs = []
         self.yolo_blocks = []
         self.route_blocks = []
@@ -254,7 +294,7 @@ def _yolov3_darknet(num_layers=53, num_classes=80,
         weight_path = get_weights_path(*(pretrain_infos[num_layers]))
         assert weight_path.endswith('.pdparams'), \
                 "suffix of weight must be .pdparams"
-        model.load(weight_path[:-9])
+        model.load(weight_path)
     return model
 
 

hapi/datasets/__init__.py

@@ -12,7 +12,14 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+from . import folder
+from . import mnist
+from . import flowers
+
 from .folder import *
 from .mnist import *
 from .flowers import *
-from .coco import *
+
+__all__ = folder.__all__ \
+        + mnist.__all__ \
+        + flowers.__all__

hapi/model.py

@@ -816,7 +816,7 @@ class Model(fluid.dygraph.Layer):
             except ValueError as err:
                 if skip_mismatch:
                     warnings.warn(
-                        ("Skip loading for {}. ".format(key) + err.message))
+                        ("Skip loading for {}. ".format(key) + str(err)))
                     # reset optimizer when mismatch happens
                     reset_optimizer = True
                 else:

hapi/vision/models/__init__.py

@@ -17,24 +17,15 @@ from . import vgg
 from . import mobilenetv1
 from . import mobilenetv2
 from . import darknet
-from . import yolov3
-from . import tsm
-from . import bmn_model
 
 from .resnet import *
 from .mobilenetv1 import *
 from .mobilenetv2 import *
 from .vgg import *
 from .darknet import *
-from .yolov3 import *
-from .tsm import *
-from .bmn_model import *
 
 __all__ = resnet.__all__ \
         + vgg.__all__ \
         + mobilenetv1.__all__ \
         + mobilenetv2.__all__ \
-        + darknet.__all__ \
-        + yolov3.__all__ \
-        + tsm.__all__ \
-        + bmn_model.__all__
+        + darknet.__all__

hapi/vision/models/darknet.py

@@ -21,7 +21,7 @@ from paddle.fluid.dygraph.nn import Conv2D, BatchNorm
 from hapi.model import Model
 from hapi.download import get_weights_path
 
-__all__ = ['DarkNet', 'ConvBNLayer', 'darknet53']
+__all__ = ['DarkNet', 'darknet53']
 
 # {num_layers: (url, md5)}
 pretrain_infos = {
@@ -136,7 +136,7 @@ class LayerWarp(fluid.dygraph.Layer):
 DarkNet_cfg = {53: ([1, 2, 8, 8, 4])}
 
 
-class DarkNet(fluid.dygraph.Layer):
+class DarkNet(Model):
     """DarkNet model from
     `"YOLOv3: An Incremental Improvement" <https://arxiv.org/abs/1804.02767>`_
 

hapi/vision/transforms/__init__.py

@@ -12,6 +12,11 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+from . import transforms
+from . import functional
+
 from .transforms import *
 from .functional import *
-from .detection_transforms import *
+
+__all__ = transforms.__all__ \
+        + functional.__all__

hapi/vision/transforms/functional.py

@@ -26,6 +26,8 @@ else:
     Sequence = collections.abc.Sequence
     Iterable = collections.abc.Iterable
 
+__all__ = ['flip', 'resize']
+
 
 def flip(image, code):
     """