Environment

• Use enivornment.yml or requirements.txt.

Description

Step 1:- Get pretrained model (if want to skip Environment Cache)

• Download pretrained model and tokenizer (GPT-2) in a local folder. Store it in a folder and load it from this location.

Step 2:- Data processing

• Split data into train/val/test.
• Input to model: " + text + + summary + ". Truncate lengths of text and summary to fit in the design. Total sequence length can be 768 or 1024.
• Create Datalaoders of train and val.

Step 3:- GPT2 Tokenizer and Model

• Add special tokens to GPT-2 tokenizer.
• Resize model embeddings for new tokenizer length.
• Fine-tuning model by passing train data and evaluating it on val data during training.
• Store the tokenizer and fine-tuned model.

Step 4:- Generation/Inference

• Generate summaries for test set which is not used during fine tune.
• Simple topk and beam search are used for the generation.

Step 5:- Evaluation with Rouge

• Compute Rouge scores for test outputs and store it.

TODO:

• Add argparser (currently all hyperparams are stored in config.py)
• Batch processing (currently working on batch_size = 1)

Credit:

• This code contains some parts from the following code/post: https://blog.paperspace.com/generating-text-summaries-gpt-2/