A Dynet Framework (Python)

A Dynet framework written by Johnny Wei ([email protected]). The goal is to make it easy to experiment with new models by providing basic handling of input/output, training, testing, and validation. A few baselines for language and sequence to sequence modelling are also provided. Don't be afraid to look into the source - take a look at the baselines as examples! If it's not obvious on how to extend this for your needs, let me know through github or email.

The framework is still largely a work in progress. I'd like to add more examples (e.g. classification, sequence tagging etc.) to the framework.

Training an LSTM Language Model

To train an LSTM language model on the Mikolov data with the specified hyperparameters in lm.py, first specify the framework directory and run the example script with:

export SEQ2SEQROOT=/directory/of/repository
bash runs/example/lm_ptb_train.py

The hyperparmeters specify a very large language model (512 input dim, 1024 hidden dim, 3 layer LSTM)! I trained using a Titan X (12GB), and tuned the memory accordingly. With it, training was possible at 383s per epoch, and after epoch 31 the validation shows:

Epoch 31. Time elapsed: 383s, 42068/42068. Total Loss: 3759063.9351. Average sequence loss: 89.3568. Average Token Loss: 4.0438.
Done. Total loss: 3759063.935143
[lr=0.00763076 clips=2630 updates=2630]
Validating...
Done. Validation loss: 338222.015259. Average Token loss: 4.585439. Perplexity: 98.046266.
Monitored quantity improved.

The validation results and checkpointed models should be saved in runs/example/ (specified in the script). Test the checkpointed model with:

bash runs/example/lm_ptb_test.sh

I was able to achieve these results on the test set:

Testing model on metric (perplexity).
Testing...
Done. Validation loss: 375382.022034. Average Token loss: 4.553949. Perplexity: 95.006857.

For reference, Zaremba et al., 2014 achieves 82.2 perplexity on validation, and 78.4 on test with a 2 layer, 1500 hidden dim LSTM. This difference in performance needs to be investigated.

Training a Seq2seq Parser

In this example, we will use the silver parse trees in data/parse/ generated from the BerkeleyParser on the Mikolov language modeling data as our dataset (more on this in data/parse/README.md). First specify the framework directory and run the preprocessing script:

export SEQ2SEQROOT=/directory/of/repository
python data/parse/preprocess.py

To train a seq2seq parser with the specified hyperparameters in seq2seq.py, run:

bash runs/example/seq2seq_ptb_train.sh

The hyperparmeters specify a very large seq2seq model (512 input dim, 512 hidden dim, 2 layer encoder/decoder)! I trained using a Titan X (12GB), and tuned the memory accordingly. With it, training was possible at 349s per epoch, and after epoch 45 the validation shows:

Epoch 45. Time elapsed: 349s, 41556/41556. Total Loss: 496500.1872. Average sequence loss: 11.9477. Average Token Loss: 0.5266.
Done. Total loss: 496500.187222
[lr=0.0732718 clips=1072 updates=1299]
Validating...
Done. Validation loss: 5964783.725300. Token-level accuracy: 0.320665. Sequence-level accuracy: 0.105638.
Monitored quantity improved.

The results of the validation should be written to runs/examples/seq2seq/validation. To evaluate, you should postprocess the sequences and evaluate using EVALB (work in progress).