Merge pull request #1081 from tiancheng2000/feature-layers-deconv-build

Feature: DeConv2d,3d build enabled

Author: Laicheng0830

tensorlayer/layers/convolution/simplified_deconv.py

@@ -82,10 +82,10 @@ def __init__(
         self.b_init = b_init
         self.in_channels = in_channels
 
-        # Attention: To build, we need not only the in_channels!
-        # if self.in_channels:
-        #     self.build(None)
-        #     self._built = True
+        # Attention: To build, we need not only the in_channels! Solved.
+        if self.in_channels is not None:
+            self.build(None)
+            self._built = True
 
         logging.info(
             "DeConv2d {}: n_filters: {} strides: {} padding: {} act: {} dilation: {}".format(
@@ -132,10 +132,13 @@ def build(self, inputs_shape):
             # dtype=tf.float32,
             name=self.name,
         )
-        if self.data_format == "channels_first":
-            self.in_channels = inputs_shape[1]
+        if inputs_shape is not None:
+            self.in_channels = inputs_shape[1 if self.data_format == "channels_first" else -1]
+        elif self.in_channels is not None:
+            inputs_shape = [1, self.in_channels, 1, 1
+                           ] if self.data_format == "channels_first" else [1, 1, 1, self.in_channels]
         else:
-            self.in_channels = inputs_shape[-1]
+            raise ValueError("Either inputs_shape or in_channels must be specified for build.")
         _out = self.layer(
             tf.convert_to_tensor(np.random.uniform(size=inputs_shape), dtype=np.float32)
         )  #np.random.uniform([1] + list(inputs_shape)))  # initialize weights
@@ -206,12 +209,12 @@ def __init__(
         self.data_format = data_format
         self.W_init = W_init
         self.b_init = b_init
-        self.in_channels = in_channels,
+        self.in_channels = in_channels
 
-        # Attention: To build, we need not only the in_channels!
-        # if self.in_channels:
-        #     self.build(None)
-        #     self._built = True
+        # Attention: To build, we need not only the in_channels! Solved.
+        if self.in_channels is not None:
+            self.build(None)
+            self._built = True
 
         logging.info(
             "DeConv3d %s: n_filters: %s strides: %s pad: %s act: %s" % (
@@ -252,16 +255,17 @@ def build(self, inputs_shape):
             bias_initializer=self.b_init,
             name=self.name,
         )
-        if self.data_format == "channels_first":
-            self.in_channels = inputs_shape[1]
+        if inputs_shape is not None:
+            self.in_channels = inputs_shape[1 if self.data_format == "channels_first" else -1]
+        elif self.in_channels is not None:
+            inputs_shape = [1, self.in_channels, 1, 1, 1
+                           ] if self.data_format == "channels_first" else [1, 1, 1, 1, self.in_channels]
         else:
-            self.in_channels = inputs_shape[-1]
-
+            raise ValueError("Either inputs_shape or in_channels must be specified for build.")
         _out = self.layer(
             tf.convert_to_tensor(np.random.uniform(size=inputs_shape), dtype=np.float32)
         )  #self.layer(np.random.uniform([1] + list(inputs_shape)))  # initialize weights
         outputs_shape = _out.shape
-        # self._add_weights(self.layer.weights)
         self._trainable_weights = self.layer.weights
 
     def forward(self, inputs):

tensorlayer/layers/normalization.py

@@ -291,9 +291,9 @@ def forward(self, inputs):
         if self.axes is None:
             self.axes = [i for i in range(len(inputs.shape)) if i != self.channel_axis]
 
-        mean, var = tf.nn.moments(inputs, self.axes, keepdims=False)
         if self.is_train:
             # update moving_mean and moving_var
+            mean, var = tf.nn.moments(inputs, self.axes, keepdims=False)
             self.moving_mean = moving_averages.assign_moving_average(
                 self.moving_mean, mean, self.decay, zero_debias=False
             )