Add simple wrapper for beam search. #675
Conversation
Provide a simple seq2seq model.
pkuyym
changed the title
|
It seems that this PR may potentially conflict with this #729 . Has work in this PR already finished so we cant begin reviewing it now? Thank you for this work. |
|
@lcy-seso This work is almost finished except some refinement and cleaning. However, the api for beam search is ready to review, please feel free to do it. |
|
I see. Thank you. |
pkuyym
changed the title
| "(default: %(default)d)") | ||
|
|
||
|
|
||
| def lstm_step(x_t, hidden_t_prev, cell_t_prev, size): |
There was a problem hiding this comment.
Do you need implement a separate lstm here? If true, make it private?
|
|
||
| forget_gate = fluid.layers.sigmoid(x=linear([hidden_t_prev, x_t])) | ||
| input_gate = fluid.layers.sigmoid(x=linear([hidden_t_prev, x_t])) | ||
| output_gate = fluid.layers.sigmoid(x=linear([hidden_t_prev, x_t])) |
There was a problem hiding this comment.
are all these gate the same? I remember in TF's LSTM implementation, you don't need to do 3 separate fc.
| return translation_ids, translation_scores, feeding_list | ||
|
|
||
|
|
||
| def to_lodtensor(data, place, dtype='int64'): |
There was a problem hiding this comment.
Can this become a general library?
| return lod_t, lod[-1] | ||
|
|
||
|
|
||
| def lodtensor_to_ndarray(lod_tensor): |
There was a problem hiding this comment.
Can this become a general library? Need comments?
| @@ -0,0 +1,458 @@ | |||
| """seq2seq model for fluid.""" | |||
There was a problem hiding this comment.
To me, this seems to be a example? Perhaps the main file should live in model zoo or examples or tests? And some general utility methods can live here?
|
|
||
|
|
||
| if __name__ == '__main__': | ||
| #train_main() |
| @@ -0,0 +1,245 @@ | |||
| # Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve. | |||
There was a problem hiding this comment.
should this file live in example or tests or model zoo?
| return self._need_reorder | ||
|
|
||
|
|
||
| class MemoryState(object): |
There was a problem hiding this comment.
should this and many others be private? _MemoryState
| self._counter.stop_gradient = True | ||
|
|
||
| # write initial state | ||
| block.append_op( |
There was a problem hiding this comment.
why is this a append_op not a layers.write_arrray?
| self._switched_decoder = False | ||
|
|
||
|
|
||
| class TrainingDecoder(object): |
|
|
Provide a simple seq2seq model.
Fix part of #674.