keras_configuration.md

In the latest version of Keras, the default backend is tensorflow and image_dim_ordering is tf. You need to change them.

1. Keras 2.x + TensorFlow configuration file of Keras: ~/.keras/keras.json

{
    "epsilon": 1e-07,
    "floatx": "float32",
    "image_data_format": "channels_first",
    "backend": "tensorflow"
}

2. Keras 1.2 + Theano configuration file of Keras: ~/.keras/keras.json

{
    "image_dim_ordering": "th",
    "epsilon": 1e-07,
    "floatx": "float32",
    "backend": "theano"
}

Ref: https://keras.io/backend/

Troubleshooting

In our experiment, we mainly tested on Keras 1.2 and Theano. If you use Keras 2.x or TensorFlow, please see the following discussions:

1. ValueError: Only layers of same output shape can be merged using sum mode
2. Only layers of same output shape can be merged using sum mode. Layer shapes: [(None, 6, 32, 2), (None, 2, 32, 2), (None, 2, 32, 2)]