Something wrong with output_shape in merge layer

[henry0312]

Sample code: https://gist.github.com/henry0312/9c1aee0a4330c860ff7b2bb139acbb15

Run: python gistfile1.py

Then:

Traceback (most recent call last):
  File "gistfile1.py", line 58, in <module>
    model, net = create_model()
  File "gistfile1.py", line 32, in create_model
    output = Dense(10, activation='sigmoid')(merged_vector)
  File "/usr/local/var/pyenv/versions/3.5.1/lib/python3.5/site-packages/Keras-1.0.3-py3.5.egg/keras/engine/topology.py", line 458, in __call__
    self.build(input_shapes[0])
  File "/usr/local/var/pyenv/versions/3.5.1/lib/python3.5/site-packages/Keras-1.0.3-py3.5.egg/keras/layers/core.py", line 589, in build
    assert len(input_shape) == 2
AssertionError

[henry0312]

maybe related to #2854 and #2818

[fchollet]

There was an issue with the Merge layer, which is that output_shape was expecting the full shape (including the batch size). This was inconsistent with the Lambda layer, so I change the Merge layer to have the right behavior. This is pushed.

There is another issue with your snippet: you are sharing a BatchNormalization layer, which is impossible. As a fix, you could use mode=2 in BatchNorm.

[fchollet]

Your snippet works with me after the latest fix (and changing the BatchNorm mode to 2).

[henry0312]

thanks!
I'll check later.

2016年5月31日火曜日、François [email protected]さんは書きました:

Your snippet works with me after the latest fix (and changing the
BatchNorm mode to 2).

—
You are receiving this because you authored the thread.
Reply to this email directly, view it on GitHub
#2855 (comment),
or mute the thread
https://github.com/notifications/unsubscribe/AAadGtNBtj2LToUE1ZQChO6NheuU5BMKks5qG7QXgaJpZM4IqKUr
.

[henry0312]

91b9302 fixed the issue, but another issue occurs 😢

sample code: https://gist.github.com/henry0312/0007239173035d8bc5725647a2f2a552
Run: python keras_import_bug.py
(Note that https://gist.github.com/henry0312/9c1aee0a4330c860ff7b2bb139acbb15 will create my_model_architecture.json)

Then:

Traceback (most recent call last):
  File "keras_import_bug.py", line 2, in <module>
    model = model_from_json(open('my_model_architecture.json').read())
  File "/usr/local/var/pyenv/versions/3.5.1/lib/python3.5/site-packages/Keras-1.0.3-py3.5.egg/keras/models.py", line 36, in model_from_json
    return layer_from_config(config, custom_objects=custom_objects)
  File "/usr/local/var/pyenv/versions/3.5.1/lib/python3.5/site-packages/Keras-1.0.3-py3.5.egg/keras/utils/layer_utils.py", line 35, in layer_from_config
    return layer_class.from_config(config['config'])
  File "/usr/local/var/pyenv/versions/3.5.1/lib/python3.5/site-packages/Keras-1.0.3-py3.5.egg/keras/engine/topology.py", line 2238, in from_config
    process_layer(layer_data)
  File "/usr/local/var/pyenv/versions/3.5.1/lib/python3.5/site-packages/Keras-1.0.3-py3.5.egg/keras/engine/topology.py", line 2235, in process_layer
    layer(input_tensors)
  File "/usr/local/var/pyenv/versions/3.5.1/lib/python3.5/site-packages/Keras-1.0.3-py3.5.egg/keras/engine/topology.py", line 1267, in __call__
    self.add_inbound_node(layers, node_indices, tensor_indices)
  File "/usr/local/var/pyenv/versions/3.5.1/lib/python3.5/site-packages/Keras-1.0.3-py3.5.egg/keras/engine/topology.py", line 543, in add_inbound_node
    Node.create_node(self, inbound_layers, node_indices, tensor_indices)
  File "/usr/local/var/pyenv/versions/3.5.1/lib/python3.5/site-packages/Keras-1.0.3-py3.5.egg/keras/engine/topology.py", line 155, in create_node
    output_shapes = to_list(outbound_layer.get_output_shape_for(input_shapes))
  File "/usr/local/var/pyenv/versions/3.5.1/lib/python3.5/site-packages/Keras-1.0.3-py3.5.egg/keras/engine/topology.py", line 1282, in get_output_shape_for
    return (input_shape[0],) + self._output_shape
TypeError: can only concatenate tuple (not "list") to tuple

[henry0312]

Maybe, this will fix.

diff --git a/keras/engine/topology.py b/keras/engine/topology.py
index a5e147d..82bed08 100644
--- a/keras/engine/topology.py
+++ b/keras/engine/topology.py
@@ -1279,7 +1279,7 @@ class Merge(Layer):
                 output_shape = self._output_shape(input_shape)
                 return output_shape
             elif self._output_shape is not None:
-                return (input_shape[0],) + self._output_shape
+                return (input_shape[0],) + tuple(self._output_shape)
             else:
                 # TODO: consider shape auto-inference with TF
                 raise Exception('The Merge layer ' + self.name +

[fchollet]

Right. I fixed it. This is because JSON deserialization converts all tuples to lists.

[henry0312]

This is because JSON deserialization converts all tuples to lists.

I see.
I confirmed 80bfec7 fixed the issue.

Thank you!