This repository was first used for the paper A Fast Unified Model for Sentence Parsing and Understanding, adapted for several subsequent papers, and is under active development for related future projects. It contains code for sentence understanding models that use tree structure or dynamic graph structure.
Requirements:
- Python 3.6
- PyTorch 0.4
- Additional dependencies listed in python/requirements.txt
Install PyTorch based on instructions online: http://pytorch.org
Install the other Python dependencies using the command below.
python3 -m pip install -r python/requirements.txt
The main executable for the SNLI experiments in the paper is supervised_classifier.py, whose flags specify the hyperparameters of the model. You can specify gpu usage by setting --gpu flag greater than or equal to 0. Uses the CPU by default.
Here's a sample command that runs a fast, low-dimensional CPU training run, training and testing only on the dev set. It assumes that you have a copy of SNLI available locally.
PYTHONPATH=spinn/python \
python3 -m spinn.models.supervised_classifier --data_type nli \
--training_data_path ~/data/snli_1.0/snli_1.0_dev.jsonl \
--eval_data_path ~/data/snli_1.0/snli_1.0_dev.jsonl \
--embedding_data_path python/spinn/tests/test_embedding_matrix.5d.txt \
--word_embedding_dim 5 --model_dim 10 --model_type CBOW
For full runs, you'll also need a copy of the 840B word 300D GloVe word vectors.
You can train SPINN using only sentence-level labels. In this case, the integrated parser will randomly sample labels during training time, and will be optimized with the REINFORCE algorithm. The command to run this model looks slightly different:
python3 -m spinn.models.rl_classifier --data_type listops \
--training_data_path spinn/python/spinn/data/listops/train_d20a.tsv \
--eval_data_path spinn/python/spinn/data/listops/test_d20a.tsv \
--word_embedding_dim 32 --model_dim 32 --mlp_dim 16 --model_type RLSPINN \
--rl_baseline value --rl_reward standard --rl_weight 42.0
Note: This model does not yet work well on natural language data, although it does on the included synthetic dataset called listops. Please look at the [sweep file][10] for an idea of which hyperparameters to use.
This project contains a handful of tools for easier analysis of your model's performance.
For one, after a periodic number of batches, some useful statistics are printed to a file specified by --log_path. This is convenient for visual inspection, and the script parse_logs.py is an example of how to easily parse this log file.
If you're interested in proposing a change or fix to SPINN, please submit a Pull Request. In addition, ensure that existing tests pass, and add new tests as you see appropriate. To run tests, simply run this command from the root directory:
nosetests python/spinn/tests
SPINN outputs metrics and statistics into a text protocol buffer format. When adding new fields to the proto file, the generated proto code needs to be updated.
bash python/build.sh
Copyright 2018, New York University
Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.