Minor fixes #66
Minor fixes #66
Conversation
| @@ -95,9 +95,6 @@ def train(args): | |||
| start = time.time() | |||
| x, y = data_loader.next_batch() | |||
| feed = {model.input_data: x, model.targets: y} | |||
| for i, (c, h) in enumerate(model.initial_state): | |||
There was a problem hiding this comment.
Hey @lns, are you sure about this? Could you point me to a resource which says passing the state directly is permitted?
There was a problem hiding this comment.
Hi, I take a look at the recent sources of tensorflow( rnn_cell_impl.py and seq2seq.py(?)) and the code seems changed a lot. In my experiments with r11(?) it's safe to remove these lines to have reasonable results. I'm not 100% sure about the way of passing states to a lstm cell, as we are using self.initial_state in rnn_decoder in model.py which is not(?) updated across the batches. If we want to pass the states from the last batch to the next batch, maybe there are more lines to be changed?
There was a problem hiding this comment.
I agree with you, filed an issue tensorflow/models#774
|
@lns This PR has merge conflicts. |
Add support for cell other than lstm, which uses state_is_tuple=True. Remove of unused state 'c' and 'h' in feed.