Reinforced-Mnemonic-Reader

Reinforced-Mnemonic-Reader copied to clipboard

Reinforced Mnemonic Reader in tensorflow

RMR: https://arxiv.org/abs/1705.02798

Pipline

1. Run the preprocess.ipynb to get the input datasets.
2. Run train_h5py.py to start training. Now elmo and cove is not useable.

notes

1. conv1d in tensor2tensor is used to instead of the matrix matmul (full connection) operation in RMR model.
2. Welcome to discuss any problem about this project (especially the RL loss).
3. The reinforcement loss should be used after the convergence of cross-entropy.
4. RMR_modelV3 is based on the version 3, and RMR_modelV6 is based on the version 6 of [https://arxiv.org/abs/1705.02798v3]. It seems that v3 performs better than v6.

Updates

• [x] Init for the RMR model (without dynamic-critical reinforcement learning DCRL)
• [x] Add the self-critical sequence training (SCST) (no test)
• [x] Update cudnn LSTM and squad 2.0
• [x] Update v3 in modelV0
• [ ] Test the RL loss

Results

Result on dev set of squad 1.1

EM:71.17% F1:79.56% (no elmo, no cove paper v6)

EM:74.37% F1:82.67% (hidden size=256 +elmo v6)

EM:72.08% F1:80.51% (no elmo, no cove paper v3)

EM:72.87% F1:81.47% (pytorch version, based on paper v3)

Result on dev set of squad 2.0

EM:64.89% F1:67.81% (+elmo+cove v3)