!15129 fix ctpn vgg16 backbone training

From: @qujianwei
Reviewed-by: @oacjiewen,@c_34
Signed-off-by: @c_34

model_zoo/official/cv/ctpn/README.md

@@ -169,11 +169,23 @@ The `pretrained_path` should be a checkpoint of vgg16 trained on Imagenet2012. T
 ...
 from src.vgg16 import VGG16
 ...
-network = VGG16()
+network = VGG16(num_classes=cfg.num_classes)
 ...
 
 ```
 
+To train a better model, you can modify some parameter in modelzoo/official/cv/vgg16/src/config.py, here we suggested you modify the "warmup_epochs" just like below, you can also try to adjust other parameter.
+
+```python
+
+imagenet_cfg = edict({
+    ...
+    "warmup_epochs": 5
+    ...
+})
+
+```
+
 Then you can train it with ImageNet2012.
 > Notes:
 > RANK_TABLE_FILE can refer to [Link](https://www.mindspore.cn/tutorial/training/en/master/advanced_use/distributed_training_ascend.html) , and the device_ip can be got as [Link](https://gitee.com/mindspore/mindspore/tree/master/model_zoo/utils/hccl_tools). For large models like InceptionV4, it's better to export an external environment variable `export HCCL_CONNECT_TIMEOUT=600` to extend hccl connection checking time from the default 120 seconds to 600 seconds. Otherwise, the connection could be timeout since compiling time increases with the growth of model size.

model_zoo/official/cv/ctpn/eval.py

@@ -56,6 +56,7 @@ def ctpn_infer_test(dataset_path='', ckpt_path='', img_dir=''):
         os.mkdir(output_dir)
     for file in os.listdir(img_dir):
         img_basenames.append(os.path.basename(file))
+    img_basenames = sorted(img_basenames)
     for data in ds.create_dict_iterator():
         img_data = data['image']
         img_metas = data['image_shape']

model_zoo/official/cv/ctpn/src/CTPN/vgg16.py

@@ -23,36 +23,22 @@ def _weight_variable(shape, factor=0.01):
     init_value = np.random.randn(*shape).astype(np.float32) * factor
     return Tensor(init_value)
 
-def _BatchNorm2dInit(out_chls, momentum=0.1, affine=True, use_batch_statistics=False):
-    """Batchnorm2D wrapper."""
-    gamma_init = Tensor(np.array(np.ones(out_chls)).astype(np.float32))
-    beta_init = Tensor(np.array(np.ones(out_chls) * 0).astype(np.float32))
-    moving_mean_init = Tensor(np.array(np.ones(out_chls) * 0).astype(np.float32))
-    moving_var_init = Tensor(np.array(np.ones(out_chls)).astype(np.float32))
-
-    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)
-
 def _conv(in_channels, out_channels, kernel_size=3, stride=1, padding=0, pad_mode='pad', weights_update=True):
     """Conv2D wrapper."""
-    weights = 'ones'
     layers = []
     conv = nn.Conv2d(in_channels, out_channels,
                      kernel_size=kernel_size, stride=stride, padding=padding,
-                     pad_mode=pad_mode, weight_init=weights, has_bias=False)
+                     pad_mode=pad_mode, has_bias=False)
     if not weights_update:
         conv.weight.requires_grad = False
     layers += [conv]
-    layers += [_BatchNorm2dInit(out_channels)]
+    layers += [nn.BatchNorm2d(out_channels)]
     return nn.SequentialCell(layers)
 
 
 def _fc(in_channels, out_channels):
     '''full connection layer'''
-    weight = _weight_variable((out_channels, in_channels))
-    bias = _weight_variable((out_channels,))
-    return nn.Dense(in_channels, out_channels, weight, bias)
+    return nn.Dense(in_channels, out_channels)
 
 
 class VGG16FeatureExtraction(nn.Cell):
@@ -141,36 +127,38 @@ class VGG16Classfier(nn.Cell):
         self.relu = nn.ReLU()
         self.fc1 = _fc(in_channels=7*7*512, out_channels=4096)
         self.fc2 = _fc(in_channels=4096, out_channels=4096)
-        self.batch_size = 32
         self.reshape = P.Reshape()
+        self.dropout = nn.Dropout(0.5)
 
     def construct(self, x):
         """
         :param x: shape=(B, 512, 7, 7)
         :return:
         """
-        x = self.reshape(x, (self.batch_size, 7*7*512))
+        x = self.reshape(x, (-1, 7*7*512))
         x = self.fc1(x)
         x = self.relu(x)
+        x = self.dropout(x)
         x = self.fc2(x)
         x = self.relu(x)
+        x = self.dropout(x)
         return x
 
 class VGG16(nn.Cell):
-    def __init__(self):
+    def __init__(self, num_classes):
         """VGG16 construct for training backbone"""
         super(VGG16, self).__init__()
-        self.feature_extraction = VGG16FeatureExtraction(weights_update=True)
+        self.vgg16_feature_extractor = VGG16FeatureExtraction(weights_update=True)
         self.max_pool = nn.MaxPool2d(kernel_size=2, stride=2)
         self.classifier = VGG16Classfier()
-        self.fc3 = _fc(in_channels=4096, out_channels=1000)
+        self.fc3 = _fc(in_channels=4096, out_channels=num_classes)
 
     def construct(self, x):
         """
         :param x: shape=(B, 3, 224, 224)
         :return: logits, shape=(B, 1000)
         """
-        feature_maps = self.feature_extraction(x)
+        feature_maps = self.vgg16_feature_extractor(x)
         x = self.max_pool(feature_maps)
         x = self.classifier(x)
         x = self.fc3(x)

model_zoo/official/cv/ctpn/src/create_dataset.py

@@ -145,7 +145,7 @@ def create_train_dataset(dataset_type):
         # test: icdar2013 test
         icdar_test_image_files, icdar_test_anno_dict = create_icdar_svt_label(config.icdar13_test_path[0],\
             config.icdar13_test_path[1], "")
-        image_files = icdar_test_image_files
+        image_files = sorted(icdar_test_image_files)
         image_anno_dict = icdar_test_anno_dict
         data_to_mindrecord_byte_image(image_files, image_anno_dict, config.test_dataset_path, \
             prefix="ctpn_test.mindrecord", file_num=1)