anurudhp / smai-project

Using CNNs for (binary) sentiment analysis.

SMAI Project: Sentence Classification using deep learning

Directory Structure:

.
├── datasets/
│   ├── <dataset_name>/
│   └── ...
├── doc/
├── src/
│   ├── models/
│   │   ├── <model_name.py>
│   │   └── ...
│   ├── parsers/
│   │   ├── <parser_name.py>
│   │   └── ...
│   ├── train.py
│   └── test.py
└── var/ (not committed)
    ├── train/
    │   └── <model_name>/
    │       └── <dataset_name>/
    └── wordvec/
        └── <dataset_name>/

Usage:

make.py [-h] [--model MODEL] [--dataset DATASET] [--parser PARSER] [--log-level LOG_LEVEL] TASK

• TASK: Task to perform {preprocess, train, test}
• --model MODEL: Model to use (src/models/MODEL.py)
• --dataset DATASET: Dataset to use (dataset/DATASET/)
• --parser PARSER: Data parser to use (src/parsers/PARSER.py)
• --log-level LOG_LEVEL: Logging level (10 for testing, and 40 for production. Default: 30)

Workflow

• preprocess.py:
  • Learns word vectors, and saves them to var/wordvec/...
• train.py:
  • Trains the model, and saves weights to var/train/...
  • Logs training reports to var/log/train/...
• test.py:
  • Loads weights from var/train/..., and predicts the labels for test data
  • Logs training reports to var/log/test/...
  • If labels are known (validation), then scores are reported.
• models/MODEL.py
  • class Model implements the CNN.
• parsers/PARSER.py
  • Implements classes to lazy-load data, and for generating word vectors using the trained word vector vocab.

Adding Models:

Check Readme in src/models/

Adding Parsers:

Adding a parser for new datasets - to handle preprocessing and conversion to word vectors. Check Readme in src/parsers/

Adding Datasets:

Add datasets to the datasets folder. Parsers should know the corresponding internal directory structure.

Contributors

Team: \sigma \sqrt -1

• Arjun P
• Gokul B Nair
• Soumya Vadlamannati
• Sai Anurudh Reddy Peduri

smai-project's Issues

Implement Model

Implement a basic model based on the paper.

Parser for IMDB dataset

Implement the data parser for the IMDB dataset, with preprocessing.

Training

Implement the train function in src/train.py.
Trains a model, and saves weights to var/train/... (exact folder depending on config)

Testing

Implement the test function in src/test.py.
Load weights from the checkpoint, and run the test.
If labels are known, report scores.

Word2Vec

Implement the word vector generator in src/preprocess.py
Learn the word vectors, and store them in var/wordvec/...

Extra: Add support for minimal user validation (random sampling etc.)