Decode generating same summery

Question

Decode generating same summery

Rachnas opened this issue 6 years ago · comments

Hi,
Thanks for providing Pytorch implementation for summerization. I trained the model on custom data set for ~15000 iterations. On decoding, same summery is generated for all the stories. I am not sure where exactly is the problem in code, still debugging. Any help/suggestion is appreciated.
In beam_search code, there is a comment '#batch should have only one example' where as in BeamSearch constructor, bacther takes batch size same as beam_size(4). Is it correct ?

Thanks

Atul Kumar · Answer 1 · Sat May 26 2018 02:18:13 GMT+0800 (China Standard Time)

Yes the first iteration of beam search replicate the same prediction for all 4 steps.
https://github.com/atulkum/pointer_summarizer/blob/master/training_ptr_gen/decode.py#L124

The best way to debug is to step through each line by putting breakpoint in code. An editor like pycharm would help.

Atul Kumar · Answer 2 · Mon Dec 09 2019 05:48:23 GMT+0800 (China Standard Time)

In decode beam search batch should have only one example
https://github.com/atulkum/pointer_summarizer/blob/master/training_ptr_gen/decode.py#L109
https://github.com/atulkum/pointer_summarizer/blob/master/data_util/batcher.py#L226

f · Answer 3 · Thu Jan 02 2020 15:40:01 GMT+0800 (China Standard Time)

i have 2 questions:
why beam_size must equals batch_size?why sample in one batch must same?

Atul Kumar · Answer 4 · Sun Jan 05 2020 01:29:54 GMT+0800 (China Standard Time)

The whole batch is used as beam for one instance. This way the decoding is made parallel on gpu.