Keras: Bug: Input layer with an Input layer as input_tensor results in empty _feed_input_names when building a Model, which makes the model unable to train

[turtleizzy]

Environment:

Ubuntu 16.04,
Python 3.5.3, 
Keras 2.0.2 from pypi, 
Backend: Tensorflow 1.1.0rc0 (with CUDA, self-compiled)

Code to reproduce this issue:

from keras.engine.topology import get_source_inputs
from keras.layers import Input, Dense
from keras.models import Model

input_shape = (512, 512, 3)

input_tensor = Input(shape = input_shape)
print (input_tensor._keras_history[0].is_placeholder) # True
x_1 = Dense(16)(input_tensor)
model_1 = Model(input_tensor, x_1)
print (len(model_1. _feed_input_names)>0)  #  True

img_input = Input(tensor = input_tensor, shape = input_shape) # <-- Bug here, digested from ./applications/inception_v3.py Line 160

print (img_input._keras_history[0].is_placeholder) # False 
print (input_tensor._keras_history[0].is_placeholder) # False <-- Somehow input_tensor is also modified
x_2 = Dense(16)(img_input)
inputs = get_source_inputs(img_input) # digested from ./applications/inception_v3.py Line 353
model_2 = Model(inputs, x_2) 
print (len(model_2. _feed_input_names)>0)  #  False

# model_2.fit(x, y) will result in 'ValueError: The model expects 0 input arrays', which renders this model untrainable.

This bug was originally found when I tried mitigating my code to Keras 2.x utilizing built-in inception-v3 framework. fit_on_generator wouldn't start and complained ValueError: The model expects 0 input arrays, but only received one array. A little bit digging into inception-v3 code revealed this bug.

A temporary workaround is to change Line 160 in inception_v3.py into img_input = input_tensor.

[jonilaserson]

Just discovered this bug right now as well.

[jonilaserson]

If you look at the analogous part in resnet50.py, this is how this case was handled:

    if input_tensor is None:
        img_input = Input(shape=input_shape)
    else:
        if not K.is_keras_tensor(input_tensor):
            img_input = Input(tensor=input_tensor, shape=input_shape)
        else:
            img_input = input_tensor

[adambielski]

Fixed in https://github.com/fchollet/keras/pull/6186

[turtleizzy]

Thanks.

I am still wondering about one thing: creating an Input object b with an Input object a as input_tensor changes the property of the original Input object a, is it supposed to work this way?

input_tensor = Input(shape = input_shape)
print (input_tensor._keras_history[0].is_placeholder) # True

img_input = Input(tensor = input_tensor, shape = input_shape) # <-- Bug here, digested from ./applications/inception_v3.py Line 160

print (img_input._keras_history[0].is_placeholder) # False 
print (input_tensor._keras_history[0].is_placeholder) # False <-- Somehow input_tensor is also modified

[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.

[raviranjan-innoplexus]

Hi, I am getting the following error while running a pre trained keras model. AttributeError("'Model' object has no attribute '_feed_input_names'",). I can't find any solution to this on the web so far. How do resolve this issue?