tensorflow / nmt

TensorFlow Neural Machine Translation Tutorial
Apache License 2.0
6.39k stars 1.96k forks source link

host memory problem #155

Closed T-B-F closed 6 years ago

T-B-F commented 7 years ago

Hi guys and thanks for the tutorial/great API.

I have experienced a "strange" behavior when playing with the code. Using my own dataset, not very large, and different architectures/parameter sets: "classic LSTM", "LSTM+attention", etc / num_layer=2, 4, ... num_units=256, 512, ... etc I noticed that the RAM of my computer increased continuously when using any attention mechanism.

However this behavior does not show up when using setting "pass_hidden_state" to False. Is it somehow linked? and "expected"? I would be grateful to anyone with an explanation.

(Using tf-1.2)