InferSent

In this repo I reproduce the sentence encoders from the paper "Supervised Learning of Universal Sentence Representations from Natural Language Inference Data" by Conneau et al. (2017) and evaluate resulting models on SentEval testing framework. Please find summary of the findings in Report.pdf.

Environment

In env.yml please find the acts environment containing all the packages needed for training and evaluation.

Pretrained models

All pretrained models are avalaiable here.

Structure

InferSent
| env.yml
| models.py       % includes the definition of the encoders
| dataset.py      % fetches the dataset related objects from torchtext
| task.py         % contains PyTorchLightining trainer InferSent
| train.py        % training logic
| utils.py        % utiltiy functions
| demoNotebook.ipnyb

| eval_iters.py   % evaluate the models on SNLI 
| eval.py         % evalaue the models on SentEval

| SentEval
|   ...       % content of the SentEval repo

Training

To train the model, run:

python train.py --args

See train.py for supported arguments.

Evaluation

Evaluate on SNLI

Run:

python eval_iters.py

In this case the path to the models need to be changed in the .py script.

Evaluate on SentEval

To evaluate on SentEval, copy the SentEval repo into InferSent and then run:

python eval.py --model_path --usepytorch

Evaluate the Short/long sentences on SNLI