make BN shape compatible to previous version

Author: ChrisWu1997

tensorlayer/files/utils.py

@@ -2666,6 +2666,10 @@ def _load_weights_from_hdf5_group(f, layers, skip=False):
             elif isinstance(layer, tl.layers.Layer):
                 weight_names = [n.decode('utf8') for n in g.attrs['weight_names']]
                 for iid, w_name in enumerate(weight_names):
+                    # FIXME : this is only for compatibility
+                    if isinstance(layer, tl.layers.BatchNorm) and np.asarray(g[w_name]).ndim > 1:
+                        assign_tf_variable(layer.all_weights[iid], np.asarray(g[w_name]).squeeze())
+                        continue
                     assign_tf_variable(layer.all_weights[iid], np.asarray(g[w_name]))
             else:
                 raise Exception("Only layer or model can be saved into hdf5.")

tensorlayer/layers/normalization.py

@@ -315,6 +315,7 @@ class BatchNorm1d(BatchNorm):
     >>> bn = tl.layers.BatchNorm1d(num_features=32)
 
     """
+
     def _check_input_shape(self, inputs):
         if inputs.ndim != 2 and inputs.ndim != 3:
             raise ValueError('expected input to be 2D or 3D, but got {}D input'.format(inputs.ndim))
@@ -337,6 +338,7 @@ class BatchNorm2d(BatchNorm):
     >>> bn = tl.layers.BatchNorm2d(num_features=32)
 
     """
+
     def _check_input_shape(self, inputs):
         if inputs.ndim != 4:
             raise ValueError('expected input to be 4D, but got {}D input'.format(inputs.ndim))
@@ -359,6 +361,7 @@ class BatchNorm3d(BatchNorm):
     >>> bn = tl.layers.BatchNorm3d(num_features=32)
 
     """
+
     def _check_input_shape(self, inputs):
         if inputs.ndim != 5:
             raise ValueError('expected input to be 5D, but got {}D input'.format(inputs.ndim))

tensorlayer/models/mobilenetv1.py

@@ -43,9 +43,9 @@ def restore_params(network, path='models'):
         expected_bytes=25600116
     )  # ls -al
     params = load_npz(name=os.path.join(path, 'mobilenet.npz'))
-    for idx, net_weight in enumerate(network.all_weights):
-        if 'batchnorm' in net_weight.name:
-            params[idx] = params[idx].reshape(1, 1, 1, -1)
+    # for idx, net_weight in enumerate(network.all_weights):
+    #     if 'batchnorm' in net_weight.name:
+    #         params[idx] = params[idx].reshape(1, 1, 1, -1)
     assign_weights(params[:len(network.all_weights)], network)
     del params
 

tensorlayer/models/resnet.py

@@ -194,8 +194,8 @@ def restore_params(network, path='models'):
             continue
         w_names = list(f[layer.name])
         params = [f[layer.name][n][:] for n in w_names]
-        if 'bn' in layer.name:
-            params = [x.reshape(1, 1, 1, -1) for x in params]
+        # if 'bn' in layer.name:
+        #     params = [x.reshape(1, 1, 1, -1) for x in params]
         assign_weights(params, layer)
         del params
 

tests/layers/test_layers_normalization.py

@@ -261,6 +261,7 @@ def test_input_shape(self):
             self.assertIsInstance(e, ValueError)
             print(e)
 
+
 if __name__ == '__main__':
 
     tl.logging.set_verbosity(tl.logging.DEBUG)