search

GateNLP / gateplugin-LearningFramework

A plugin for the GATE language technology framework for training and using machine learning models. Currently supports Mallet (MaxEnt, NaiveBayes, CRF and others), LibSVM, Scikit-Learn, Weka, and DNNs through Pytorch and Keras.
https://gatenlp.github.io/gateplugin-LearningFramework/
GNU Lesser General Public License v2.1
26 stars 6 forks source link

Xingyi's bug report #109

Closed deansong closed 5 years ago

deansong commented 5 years ago

I notice several bugs may need to remove in next version

  1. Copy model from GPU machine to CPU only cause cuda error: raise RuntimeError(‘Attempting to deserialize object on a CUDA ’ (detail in slack) -- current solution is del state['optimizer'] in the getstate seems removed the cuda error

  2. Performance difference when copy model from GPU machine to CPU only machine. copied file: FileJsonPyTorch.model.module.pytorch about 10% accuracy difference in same test corpus. GATE version is same but pytorch is different 0.4.1 (gs8) vs 1.0.1 (xingyi's mac)

  3. Path error when copy FileJsonPyTorch.model.wrapper.pickle

    • the meta file path is stored in FileJsonPyTorch.model.wrapper.pickle, this cause can't find file error when copy to different machine
johann-petrak commented 5 years ago

1) corresponds to https://github.com/GateNLP/gate-lf-pytorch-json/issues/38 2) corresponds to https://github.com/GateNLP/gate-lf-pytorch-json/issues/41 3) corresponds to https://github.com/GateNLP/gate-lf-pytorch-json/issues/39

johann-petrak commented 5 years ago

1) and 3) should be fixed in the latest snapshot but need testing.

johann-petrak commented 5 years ago

Tested and all three issues seem to be solved.