Loading the weights - Bias term about sigver_wiwd

[luizgh]

Hi Victor,

We don't use biases on the convolutions / fully connected layers when using batch normalization, since the biases would be removed during batch normalization (when the mean is subtracted). Instead, the biases are added as part of the batch normalization layer, but they are called "beta" in this context.

For your reference, I have added a tensorflow implementation of the model, that uses the same weights. See the file "tf_example.py". Note that I only implemented the "test" version of batch normalization. If you want re-train / finetune the model you would need to adapt the code (use the tf implementation of batch normalization, but you need to check the details. e.g. Lasagne stores the inverse of the standard deviation, which may not be the case for tensorflow (it may store the standard deviation itself)).

Let me know if that solves your problem.

[victormelo]

Hello Luiz,

Thank you very much for your attention and kindness in providing me a working test version for Tensorflow.

Best regards,
Victor.