Keras 2 bug in `concatenate()` in a loop appending to a tensor list (ex: densenet)

[ahundt]

I've got code that works in the keras-1 compatibility layer of keras-2, but fails when I use the new concatenate() API.

Key lines:

    concat_axis = 1 if K.image_data_format() == 'channels_first' else -1

    x_list = [x]

    for i in range(nb_layers):
        x = __conv_block(x, growth_rate, bottleneck,
                         dropout_rate, weight_decay)
        x_list.append(x)

        # commenting merge() and uncommenting concatenate() fails
        x = merge(x_list, mode='concat', concat_axis=concat_axis)
        # x = concatenate(x_list, concat_axis)

        if grow_nb_filters:
            nb_filter += growth_rate

    if return_concat_list:
        return x, nb_filter, x_list
    else:
        return x, nb_filter

Error:

Traceback (most recent call last):
  File "/Users/athundt/source/keras-contrib/keras_contrib/applications/densenet.py", line 712, in <module>
    dn = DenseNet()
  File "/Users/athundt/source/keras-contrib/keras_contrib/applications/densenet.py", line 136, in DenseNet
    model = Model(inputs, x, name='densenet')
  File "/usr/local/lib/python2.7/site-packages/Keras-2.0.1-py2.7.egg/keras/legacy/interfaces.py", line 87, in wrapper
    return func(*args, **kwargs)
  File "/usr/local/lib/python2.7/site-packages/Keras-2.0.1-py2.7.egg/keras/engine/topology.py", line 1704, in __init__
    str(layers_with_complete_input))
RuntimeError: Graph disconnected: cannot obtain value for tensor Tensor("conv2d_2/convolution:0", shape=(?, 32, 32, 12), dtype=float32) at layer "concatenate_1". The following previous layers were accessed without issue: ['input_1', 'initial_conv2D', 'batch_normalization_1', 'activation_1', 'conv2d_1']

The full code is in keras-team/keras-contrib@bf95474

Thank you!

• Check that you are up-to-date with the master branch of Keras. You can update with:
  pip install git+git://github.com/fchollet/keras.git --upgrade --no-deps

• If running on TensorFlow, check that you are up-to-date with the latest version. The installation instructions can be found here.

• Provide a link to a GitHub Gist of a Python script that can reproduce your issue (or just copy the script here if it is short).

[dbysll]

Similar thing happened to me.... Well, it used to work for me using
merge([up6, conv4], mode='concat', concat_axis=1)

but after updating keras, it fails using
concatenate([up6, conv4], axis=1)

The error message is:
File "unet_2d.py", line 119, in get_2D_unet
up6 = concatenate([up6, conv4], axis=1)
File "/data/deep/anaconda/anaconda2/envs/tensorflow/lib/python2.7/site-packages/keras/layers/merge.py", line 391, in concatenate
return Concatenate(axis=axis, **kwargs)(inputs)
File "/data/deep/anaconda/anaconda2/envs/tensorflow/lib/python2.7/site-packages/keras/engine/topology.py", line 554, in call
output = self.call(inputs, **kwargs)
File "/data/deep/anaconda/anaconda2/envs/tensorflow/lib/python2.7/site-packages/keras/layers/merge.py", line 166, in call
return K.concatenate(inputs, axis=self.axis)
File "/data/deep/anaconda/anaconda2/envs/tensorflow/lib/python2.7/site-packages/keras/backend/tensorflow_backend.py", line 1525, in concatenate
return tf.concat([to_dense(x) for x in tensors], axis)
File "/data/deep/anaconda/anaconda2/envs/tensorflow/lib/python2.7/site-packages/tensorflow/python/ops/array_ops.py", line 1075, in concat
dtype=dtypes.int32).get_shape(
File "/data/deep/anaconda/anaconda2/envs/tensorflow/lib/python2.7/site-packages/tensorflow/python/framework/ops.py", line 669, in convert_to_tensor
ret = conversion_func(value, dtype=dtype, name=name, as_ref=as_ref)
File "/data/deep/anaconda/anaconda2/envs/tensorflow/lib/python2.7/site-packages/tensorflow/python/framework/constant_op.py", line 176, in _constant_tensor_conversion_function
return constant(v, dtype=dtype, name=name)
File "/data/deep/anaconda/anaconda2/envs/tensorflow/lib/python2.7/site-packages/tensorflow/python/framework/constant_op.py", line 165, in constant
tensor_util.make_tensor_proto(value, dtype=dtype, shape=shape, verify_shape=verify_shape))
File "/data/deep/anaconda/anaconda2/envs/tensorflow/lib/python2.7/site-packages/tensorflow/python/framework/tensor_util.py", line 367, in make_tensor_proto
_AssertCompatible(values, dtype)
File "/data/deep/anaconda/anaconda2/envs/tensorflow/lib/python2.7/site-packages/tensorflow/python/framework/tensor_util.py", line 302, in _AssertCompatible
(dtype.name, repr(mismatch), type(mismatch).name))
TypeError: Expected int32, got list containing Tensors of type '_Message' instead.

Please help me!!!!

[junwei-pan]

@dbysll Please update your pip, tensorflow and keras respectively:
pip install pip --upgrade
pip install tensorflow --upgrade
pip install keras --upgrade

[dbysll]

@kemaswill Thanks for your comments. But I believe I am already using the latest version of Keras and tensorflow.

[ahundt]

@farizrahman4u any thoughts on this? I know you were recently working on merge.py in #5812

I can also duplicate this problem in keras-contrib cifar10_densenet.py, if any the merge() calls are appropriately changed to concatenate(). It runs if left alone which is with the outdated keras-1 APIs.

[farizrahman4u]

Please provide a minimal standalone script which I can run locally.

[ahundt]

@farizrahman4u I added #6034 which reproduces the bug.

[farizrahman4u]

@ahundt Again, I don't understand how it is reproducible.. (what is __conv_block ?). Can you provide a minimal standalone script that I can quickly copy-paste to my terminal? Thanks!

[ahundt]

@farizrahman4u sorry there was a typo which is now fixed.

[ahundt]

Here is the code:

import numpy as np
from numpy.testing import assert_allclose
from keras import layers
from keras import models
from keras import backend as K
from keras.utils.test_utils import layer_test
from keras.utils.test_utils import keras_test
from keras.layers import merge

x3 = np.random.random((1, 1, 1))
nb_layers = 4
x_i = layers.Input(shape=(None, None))
x_list = [x_i]
x = x_i
for i in range(nb_layers):
    x_list.append(x)
    # The concatenate line fails if uncommented + comment merge line
    # x = layers.concatenate(x_list)
    x = layers.merge(x_list, mode='concat', concat_axis=1)
concat_model = models.Model(x_i, x)
concat_out = concat_model.predict([x3])
x3 = np.repeat(x3, 16, axis=1)
assert concat_out.shape == (1, 16, 1)
assert_allclose(concat_out, x3)