Error in model load weights

[lenlen]

System information

• Have I written custom code (as opposed to using a stock example script provided in TensorFlow): N/A
• OS Platform and Distribution (e.g., Linux Ubuntu 16.04): from nightly-devel-gpu-py3 and nightly-gpu-py3 docker images
• TensorFlow installed from (source or binary): from nightly-devel-gpu-py3 and nightly-gpu-py3 docker images
• TensorFlow version (use command below): from nightly-devel-gpu-py3 and nightly-gpu-py3 docker images
• Python version: from nightly-devel-gpu-py3 and nightly-gpu-py3 docker images
• Bazel version (if compiling from source): from nightly-devel-gpu-py3 and nightly-gpu-py3 docker images
• GCC/Compiler version (if compiling from source): from nightly-devel-gpu-py3 and nightly-gpu-py3 docker images
• CUDA/cuDNN version: 9.0
• GPU model and memory: tesla k80
• Exact command to reproduce: N/A

Describe the problem

Running with nightly (both nightly-devel-gpu-py3 and nightly-gpu-py3) is not possible load weights on the keras model.
In the example belowe I have saved the weights, they are saved correctly in the file system and then I have tryed to load them, but I've got an error.

Source code / logs

Here the snippet:

import tensorflow as tf
from tensorflow import keras as ks


def main():
    input_rgb = ks.layers.Input(shape=(1, 5, 5, 3), name="input_rgb")
    x = ks.layers.Dense(1, activation='relu', name="Dense_1")(input_rgb)
    x = ks.layers.Dense(1, activation='sigmoid', name="sigmoid")(x)
    model = ks.models.Model(inputs=[input_rgb], outputs=[x])
    model.compile(
        loss={'sigmoid': 'binary_crossentropy'},
        optimizer=tf.keras.optimizers.Adam())
    model.save_weights('/tmp/test_weights.h5')

    model.load_weights('/tmp/test_weights.h5', by_name=True)


if __name__ == "__main__":
    main()

And here the output:

2018-05-15 09:56:21.404382: W tensorflow/core/util/tensor_slice_reader.cc:95] Could not open /tmp/test_weights.h5: Data loss: not an sstable (bad magic number): perhaps your file is in a different file format and you need to use a different restore operator?

[allenlavoie]

You can safely ignore the warning. It does a test to see if the file is in TensorFlow format, sees that it isn't, and loads using HDF5.

I'll put together a change to get rid of the warning anyway. Thank you for the report!