Igor Shalyminov

Thanks everyone! Although disabling TPM seems to be unbricking the device, it doesn't feel like a legit solution, especially for Surface. I remember I kept getting bitlocked every single time...

@timgrossmann could you please review - specifically, is this a supported Firefox version?

Here it is: https://github.com/ishalyminov/seq2seq_dialogue/blob/master/training_utils.py#L18

Wow, didn't know about the TimeDistributedSoftmax, thanks! Although I tried without the Softmax layer at all (and MSE objective) - the same basic behavior.. My parameters are here: https://github.com/ishalyminov/seq2seq_dialogue/blob/master/sanity_check.json

I actually don't see time_distributed_softmax among the Keras activations: https://github.com/fchollet/keras/blob/master/keras/activations.py#L6 - maybe softmax already works with time distributed data (it's mentioned there that it can handle 3D data)?

Could you describe how you pass data into the network and what you input/output tensors look like (also, is there any difference if I set batch_input_shape or input_shape, output_shape individually)?...

@farizrahman4u I've just managed to: 1) create a model with explicitly specified input_length, output_length for one bucket 2) save its weights 3) create a model for another bucket (again, setting...