search

THUNLP-MT / THUMT

An open-source neural machine translation toolkit developed by Tsinghua Natural Language Processing Group
BSD 3-Clause "New" or "Revised" License
701 stars 197 forks source link

Different performance with RNNSearch vs. RNNSearch_LRP #94

Closed liushengzhong1023 closed 4 years ago

liushengzhong1023 commented 4 years ago

Hi,

I am using this repo in a text classification problem. However, two unexpected phenomenons were observed:

  1. When I run the inference, it seems that the RNNSearch and RNNSearch_LRP will return different results and get different performance. Specifically in my case, different accuracies. I use RNNSearch for training, while want to use RNNSearch_LRP for inference and do LRP. The accuracy I got on RNNSearch_LRP is ~10% lower than RNNSearch.
  2. When I run the LRP with get_relevance.py, there seems to be a memory leakage during the code running. My CPU memory keeps increasing during execution.

Have you met similar problems in your past testing, or have any hints that may help me debug the problems at my side?

Thanks, Shengzhong