Merge pull request PaddlePaddle#41 from LielinJiang/fix-bugs-from-qa

release some image classification models

examples/image_classification/README.MD

@@ -43,13 +43,13 @@ CUDA_VISIBLE_DEVICES=0,1,2,3 python -m paddle.distributed.launch main.py --arch
 ### 单卡预测
 执行如下命令进行预测
 ```bash
-python -u main.py --arch resnet50 -d --evaly-only /path/to/imagenet 
+python -u main.py --arch resnet50 -d --eval-only /path/to/imagenet 
 ```
 
 ### 多卡预测
 执行如下命令进行多卡预测
 ```bash
-CUDA_VISIBLE_DEVICES=0,1,2,3 python -m paddle.distributed.launch main.py --arch resnet50 --evaly-only /path/to/imagenet
+CUDA_VISIBLE_DEVICES=0,1,2,3 python -m paddle.distributed.launch main.py --arch resnet50 --eval-only /path/to/imagenet
 ```
 
 
@@ -71,12 +71,17 @@ CUDA_VISIBLE_DEVICES=0,1,2,3 python -m paddle.distributed.launch main.py --arch
 * **weight-decay**: 模型权重正则化系数，默认值：1e-4
 * **momentum**: SGD优化器的动量，默认值：0.9
 
+注意：使用```--resume```恢复训练时，假如你的模型路径为```./output/118.pdparams```，你输入的路径不需要带后缀，即```--resume ./output/118```即可。
 
 ## 模型
 
 | 模型 | top1 acc | top5 acc |
 | --- | --- | --- |
+| [ResNet18](https://paddle-hapi.bj.bcebos.com/models/resnet18.pdparams) | 71.72 | 90.60 |
+| [ResNet34](https://paddle-hapi.bj.bcebos.com/models/resnet34.pdparams) | 75.02 | 92.31 |
 | [ResNet50](https://paddle-hapi.bj.bcebos.com/models/resnet50.pdparams) | 76.27 | 93.03 |
+| [ResNet101](https://paddle-hapi.bj.bcebos.com/models/resnet101.pdparams) | 78.33 | 94.04 |
+| [ResNet152](https://paddle-hapi.bj.bcebos.com/models/resnet152.pdparams) | 78.78 | 94.40 |
 | [vgg16](https://paddle-hapi.bj.bcebos.com/models/vgg16.pdparams) | 71.92 | 90.65 | 
 | [mobilenet_v1](https://paddle-hapi.bj.bcebos.com/models/mobilenet_v1_x1.0.pdparams) | 71.16 | 89.89 | 
 | [mobilenet_v2](https://paddle-hapi.bj.bcebos.com/models/mobilenet_v2_x1.0.pdparams) | 72.30 | 90.74 | 

examples/image_classification/main.py

@@ -76,6 +76,9 @@ def main():
     device = set_device(FLAGS.device)
     fluid.enable_dygraph(device) if FLAGS.dynamic else None
 
+    model_list = [x for x in models.__dict__["__all__"]]
+    assert FLAGS.arch in model_list, "Expected FLAGS.arch in {}, but received {}".format(
+        model_list, FLAGS.arch)
     model = models.__dict__[FLAGS.arch](pretrained=FLAGS.eval_only and
                                         not FLAGS.resume)
 
@@ -94,7 +97,13 @@ def main():
             len(train_dataset) * 1. / FLAGS.batch_size / ParallelEnv().nranks),
         parameter_list=model.parameters())
 
-    model.prepare(optim, CrossEntropy(), Accuracy(topk=(1, 5)), inputs, labels)
+    model.prepare(
+        optim,
+        CrossEntropy(),
+        Accuracy(topk=(1, 5)),
+        inputs,
+        labels,
+        FLAGS.device)
 
     if FLAGS.eval_only:
         model.evaluate(
@@ -152,7 +161,7 @@ def main():
         type=str,
         help="checkpoint path to resume")
     parser.add_argument(
-        "--eval-only", action='store_true', help="enable dygraph mode")
+        "--eval-only", action='store_true', help="only evaluate the model")
     parser.add_argument(
         "--lr-scheduler",
         default='piecewise',

examples/image_classification/scripts/mobilenet_v1_x1.0.sh

@@ -0,0 +1,13 @@
+export CUDA_VISIBLE_DEVICES=0,1,2,3 
+
+# 默认imagenet数据存储在data/ILSVRC2012/下，去除-d便使用静态图模式运行
+python -m paddle.distributed.launch main.py \
+        --arch mobilenet_v1 \
+        --epoch 120 \
+        --batch-size 64 \
+        --learning-rate 0.1 \
+        --lr-scheduler piecewise \
+        --milestones 30 60 90 \
+        --weight-decay 3e-5 \
+        -d \
+        data/ILSVRC2012/

examples/image_classification/scripts/mobilenet_v2_x1.0.sh

@@ -0,0 +1,12 @@
+export CUDA_VISIBLE_DEVICES=0,1,2,3 
+
+# 默认imagenet数据存储在data/ILSVRC2012/下，去除-d便使用静态图模式运行
+python -m paddle.distributed.launch main.py \
+        --arch mobilenet_v2 \
+        --epoch 240 \
+        --batch-size 64 \
+        --learning-rate 0.1 \
+        --lr-scheduler cosine \
+        --weight-decay 4e-5 \
+        -d \
+        data/ILSVRC2012/

examples/image_classification/scripts/resnet101.sh

@@ -0,0 +1,10 @@
+export CUDA_VISIBLE_DEVICES=0,1,2,3 
+
+# 默认imagenet数据存储在data/ILSVRC2012/下，去除-d便使用静态图模式运行
+python -m paddle.distributed.launch main.py \
+        --arch resnet101 \
+        --epoch 90 \
+        --batch-size 64 \
+        --learning-rate 0.1 \
+        -d \
+        data/ILSVRC2012/

examples/image_classification/scripts/resnet152.sh

@@ -0,0 +1,10 @@
+export CUDA_VISIBLE_DEVICES=0,1,2,3 
+
+# 默认imagenet数据存储在data/ILSVRC2012/下，去除-d便使用静态图模式运行
+python -m paddle.distributed.launch main.py \
+        --arch resnet152 \
+        --epoch 90 \
+        --batch-size 64 \
+        --learning-rate 0.1 \
+        -d \
+        data/ILSVRC2012/

examples/image_classification/scripts/resnet18.sh

@@ -0,0 +1,11 @@
+export CUDA_VISIBLE_DEVICES=0,1,2,3 
+
+# 默认imagenet数据存储在data/ILSVRC2012/下，去除-d便使用静态图模式运行
+python -m paddle.distributed.launch main.py \
+        --arch resnet18 \
+        --epoch 120 \
+        --batch-size 64 \
+        --learning-rate 0.1 \
+        --lr-scheduler cosine \
+        -d \
+        data/ILSVRC2012/

examples/image_classification/scripts/resnet34.sh

@@ -0,0 +1,11 @@
+export CUDA_VISIBLE_DEVICES=0,1,2,3 
+
+# 默认imagenet数据存储在data/ILSVRC2012/下，去除-d便使用静态图模式运行
+python -m paddle.distributed.launch main.py \
+        --arch resnet34 \
+        --epoch 120 \
+        --batch-size 64 \
+        --learning-rate 0.1 \
+        --lr-scheduler cosine \
+        -d \
+        data/ILSVRC2012/

examples/image_classification/scripts/resnet50.sh

@@ -0,0 +1,10 @@
+export CUDA_VISIBLE_DEVICES=0,1,2,3 
+
+# 默认imagenet数据存储在data/ILSVRC2012/下，去除-d便使用静态图模式运行
+python -m paddle.distributed.launch main.py \
+        --arch resnet50 \
+        --epoch 90 \
+        --batch-size 64 \
+        --learning-rate 0.1 \
+        -d \
+        data/ILSVRC2012/

examples/image_classification/scripts/vgg16.sh

@@ -0,0 +1,11 @@
+export CUDA_VISIBLE_DEVICES=0,1,2,3 
+
+# 默认imagenet数据存储在data/ILSVRC2012/下，去除-d便使用静态图模式运行
+python -m paddle.distributed.launch main.py \
+        --arch vgg16 \
+        --epoch 90 \
+        --batch-size 64 \
+        --learning-rate 0.01 \
+        --lr-scheduler cosine \
+        -d \
+        data/ILSVRC2012/

hapi/model.py

@@ -798,7 +798,7 @@ def _check_match(key, param):
                     "{} receives a shape {}, but the expected shape is {}.".
                     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'], \
@@ -936,35 +936,35 @@ def fit(
         Args:
             train_data (Dataset|DataLoader): An iterable data loader is used for 
                 train. An instance of paddle paddle.io.Dataset or 
-                paddle.io.Dataloader is recomended.
+                paddle.io.Dataloader is recomended. Default: None.
             eval_data (Dataset|DataLoader): An iterable data loader is used for
                 evaluation at the end of epoch. If None, will not do evaluation. 
                 An instance of paddle.io.Dataset or paddle.io.Dataloader 
-                is recomended.
+                is recomended. Default: None.
             batch_size (int): Integer number. The batch size of train_data and eval_data. 
                 When train_data and eval_data are both the instance of Dataloader, this 
-                parameter will be ignored.
-            epochs (int): Integer number. The number of epochs to train the model.
+                parameter will be ignored. Default: 1.
+            epochs (int): Integer number. The number of epochs to train the model. Default: 1.
             eval_freq (int): The frequency, in number of epochs, an evalutation
-                is performed.
+                is performed. Default: 1.
             log_freq (int): The frequency, in number of steps, the training logs
-                are printed.
+                are printed. Default: 10.
             save_dir(str|None): The directory to save checkpoint during training.
-                If None, will not save checkpoint.
-            save_freq (int): The frequency, in number of epochs, to save checkpoint.
+                If None, will not save checkpoint. Default: None.
+            save_freq (int): The frequency, in number of epochs, to save checkpoint. Default: 1.
             verbose (int): The verbosity mode, should be 0, 1, or 2.
-                0 = silent, 1 = progress bar, 2 = one line per epoch.
+                0 = silent, 1 = progress bar, 2 = one line per epoch. Default: 2.
             drop_last (bool): whether drop the last incomplete batch of train_data 
                 when dataset size is not divisible by the batch size. When train_data 
-                is an instance of Dataloader, this parameter will be ignored.
+                is an instance of Dataloader, this parameter will be ignored. Default: False.
             shuffle (bool): whther to shuffle train_data. When train_data is an instance 
-                of Dataloader, this parameter will be ignored.
+                of Dataloader, this parameter will be ignored. Default: True.
             num_workers (int): the number of subprocess to load data, 0 for no subprocess 
                 used and loading data in main process. When train_data and eval_data are
-                both the instance of Dataloader, this parameter will be ignored.
+                both the instance of Dataloader, this parameter will be ignored. Default: 0.
             callbacks (Callback|None): A list of `Callback` instances to apply
                 during training. If None, `ProgBarLogger` and `ModelCheckpoint`
-                are automatically inserted.
+                are automatically inserted. Default: None.
         """
 
         assert train_data is not None, \
@@ -1066,18 +1066,20 @@ def evaluate(
                 evaluation. An instance of paddle.io.Dataset or 
                 paddle.io.Dataloader is recomended.
             batch_size (int): Integer number. The batch size of train_data and eval_data. 
-                When train_data and eval_data are both the instance of Dataloader, this 
-                parameter will be ignored.
+                When eval_data is the instance of Dataloader, this argument will be ignored.
+                Default: 1.
             log_freq (int): The frequency, in number of steps, the eval logs
-                are printed.
+                are printed. Default: 10.
             verbose (int): The verbosity mode, should be 0, 1, or 2.
-                0 = silent, 1 = progress bar, 2 = one line per epoch.
+                0 = silent, 1 = progress bar, 2 = one line per epoch. Default: 2.
             num_workers (int): The number of subprocess to load data, 0 for no subprocess 
                 used and loading data in main process. When train_data and eval_data are
-                both the instance of Dataloader, this parameter will be ignored.
+                both the instance of Dataloader, this parameter will be ignored. Default: 0.
             callbacks (Callback|None): A list of `Callback` instances to apply
                 during training. If None, `ProgBarLogger` and `ModelCheckpoint`
-                are automatically inserted.
+                are automatically inserted. Default: None.
+        Returns:
+            dict: Result of metric.
         """
 
         if fluid.in_dygraph_mode():
@@ -1142,16 +1144,18 @@ def predict(self,
                 is recomended.
             batch_size (int): Integer number. The batch size of train_data and eval_data. 
                 When train_data and eval_data are both the instance of Dataloader, this 
-                parameter will be ignored.
+                argument will be ignored. Default: 1.
             num_workers (int): the number of subprocess to load data, 0 for no subprocess 
                 used and loading data in main process. When train_data and eval_data are
-                both the instance of Dataloader, this parameter will be ignored.
+                both the instance of Dataloader, this argument will be ignored. Default: 0.
             stack_output (bool): whether stack output field like a batch, as for an output
                 filed of a sample is in shape [X, Y], test_data contains N samples, predict
                 output field will be in shape [N, X, Y] if stack_output is True, and will
                 be a length N list in shape [[X, Y], [X, Y], ....[X, Y]] if stack_outputs
                 is False. stack_outputs as False is used for LoDTensor output situation,
-                it is recommended set as True if outputs contains no LoDTensor. Default False
+                it is recommended set as True if outputs contains no LoDTensor. Default: False.
+        Returns:
+            list: output of models.
         """
 
         if fluid.in_dygraph_mode():

hapi/vision/models/mobilenetv1.py

@@ -263,7 +263,7 @@ def forward(self, inputs):
 
 
 def _mobilenet(arch, pretrained=False, **kwargs):
-    model = MobileNetV1(num_classes=1000, with_pool=True, **kwargs)
+    model = MobileNetV1(**kwargs)
     if pretrained:
         assert arch in model_urls, "{} model do not have a pretrained model now, you should set pretrained=False".format(
             arch)
@@ -276,12 +276,13 @@ def _mobilenet(arch, pretrained=False, **kwargs):
     return model
 
 
-def mobilenet_v1(pretrained=False, scale=1.0):
+def mobilenet_v1(pretrained=False, scale=1.0, **kwargs):
     """MobileNetV1
     
     Args:
         pretrained (bool): If True, returns a model pre-trained on ImageNet. Default: False.
         scale: (float): scale of channels in each layer. Default: 1.0.
     """
-    model = _mobilenet('mobilenetv1_' + str(scale), pretrained, scale=scale)
+    model = _mobilenet(
+        'mobilenetv1_' + str(scale), pretrained, scale=scale, **kwargs)
     return model

hapi/vision/models/mobilenetv2.py

@@ -237,7 +237,7 @@ def forward(self, inputs):
 
 
 def _mobilenet(arch, pretrained=False, **kwargs):
-    model = MobileNetV2(num_classes=1000, with_pool=True, **kwargs)
+    model = MobileNetV2(**kwargs)
     if pretrained:
         assert arch in model_urls, "{} model do not have a pretrained model now, you should set pretrained=False".format(
             arch)
@@ -250,12 +250,13 @@ def _mobilenet(arch, pretrained=False, **kwargs):
     return model
 
 
-def mobilenet_v2(pretrained=False, scale=1.0):
+def mobilenet_v2(pretrained=False, scale=1.0, **kwargs):
     """MobileNetV2
     
     Args:
         pretrained (bool): If True, returns a model pre-trained on ImageNet. Default: False.
         scale: (float): scale of channels in each layer. Default: 1.0.
     """
-    model = _mobilenet('mobilenetv2_' + str(scale), pretrained, scale=scale)
+    model = _mobilenet(
+        'mobilenetv2_' + str(scale), pretrained, scale=scale, **kwargs)
     return model