Keras: Setting trainable to false change the order in which the layers of a sub-model are serialized

[javiercorrea]

When using a model within another model and setting some layers with trainable=False change the way the model is serialized.

Then, it is not possible to use load_weights.

Here's an example which results in error:

from keras.models import Model
from keras.layers import Dense, Input, Add 

def create_model(frozen=True):
    x = Input(shape=(100,))
    y = Dense(20)(x)
    y = Dense(5, trainable=frozen)(y)
    y = Dense(10)(y)
    model1 = Model(x, y)

    X = Input(shape=(100,))
    Y = model1(X)
    model2 = Model(X, Y)

    return model2

m1 = create_model(True)
m1.save_weights('test.h5')

m2 = create_model(False)
m2.load_weights('test.h5')

which results in this error:

Traceback (most recent call last):
  File "test.py", line 21, in <module>
    m2.load_weights('test.h5')
  File "/home/javier/.local/lib/python3.5/site-packages/keras/engine/topology.py", line 2538, in load_weights
    load_weights_from_hdf5_group(f, self.layers)
  File "/home/javier/.local/lib/python3.5/site-packages/keras/engine/topology.py", line 2970, in load_weights_from_hdf5_group
    K.batch_set_value(weight_value_tuples)
  File "/home/javier/.local/lib/python3.5/site-packages/keras/backend/tensorflow_backend.py", line 2148, in batch_set_value
    assign_op = x.assign(assign_placeholder)
  File "/home/javier/.local/lib/python3.5/site-packages/tensorflow/python/ops/variables.py", line 512, in assign
    return state_ops.assign(self._variable, value, use_locking=use_locking)
  File "/home/javier/.local/lib/python3.5/site-packages/tensorflow/python/ops/state_ops.py", line 271, in assign
    validate_shape=validate_shape)
  File "/home/javier/.local/lib/python3.5/site-packages/tensorflow/python/ops/gen_state_ops.py", line 47, in assign
    use_locking=use_locking, name=name)
  File "/home/javier/.local/lib/python3.5/site-packages/tensorflow/python/framework/op_def_library.py", line 768, in apply_op
    op_def=op_def)
  File "/home/javier/.local/lib/python3.5/site-packages/tensorflow/python/framework/ops.py", line 2338, in create_op
    set_shapes_for_outputs(ret)
  File "/home/javier/.local/lib/python3.5/site-packages/tensorflow/python/framework/ops.py", line 1719, in set_shapes_for_outputs
    shapes = shape_func(op)
  File "/home/javier/.local/lib/python3.5/site-packages/tensorflow/python/framework/ops.py", line 1669, in call_with_requiring
    return call_cpp_shape_fn(op, require_shape_fn=True)
  File "/home/javier/.local/lib/python3.5/site-packages/tensorflow/python/framework/common_shapes.py", line 610, in call_cpp_shape_fn
    debug_python_shape_fn, require_shape_fn)
  File "/home/javier/.local/lib/python3.5/site-packages/tensorflow/python/framework/common_shapes.py", line 676, in _call_cpp_shape_fn_impl
    raise ValueError(err.message)
ValueError: Dimension 0 in both shapes must be equal, but are 5 and 20 for 'Assign_2' (op: 'Assign') with input shapes: [5,10], [20,5].

Even when using load_weights(..., by_name=True) the problem persist.

[javiercorrea]

This is a fix when the sub-model name's are the same as the ones saved. The problem persist if no layer names are specified.

diff --git a/keras/engine/topology.py b/keras/engine/topology.py
index 7b27a9b..36019ac 100644
--- a/keras/engine/topology.py
+++ b/keras/engine/topology.py
@@ -3021,6 +3021,7 @@ def load_weights_from_hdf5_group_by_name(f, layers):
         g = f[name]
         weight_names = [n.decode('utf8') for n in g.attrs['weight_names']]
         weight_values = [g[weight_name] for weight_name in weight_names]
+        weight_indices = dict((name, i) for i, name in enumerate(weight_names))

         for layer in index.get(name, []):
             symbolic_weights = layer.weights
@@ -3039,6 +3040,7 @@ def load_weights_from_hdf5_group_by_name(f, layers):
                                  ' element(s).')
             # Set values.
             for i in range(len(weight_values)):
+                weight_index = weight_indices[symbolic_weights[i].name]
                 weight_value_tuples.append((symbolic_weights[i],
-                                            weight_values[i]))
+                                            weight_values[weight_index]))
     K.batch_set_value(weight_value_tuples)

[minbinL]

if 2 models' trainable parameter is different, the weights can't be loaded.

[stale[bot]]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed after 30 days if no further activity occurs, but feel free to re-open a closed issue if needed.

[ClementWalter]

Still an issue with keras 2.1.5, copy/paste the code results in the same error