# MemSum Setup and Usage Guide

## Step 1: Set Up Environment
First, you can install the entire codebase from this google drive link containing embeddings and the pretrained model. 
If you install from the gdrive you can skip installing GloVe embeddings, and go straight to testing the trained model. 
https://drive.google.com/file/d/1NWwM06eLVWsf7XZYFcXuBmXBSEFM5V3q/view?usp=sharing

Create an Anaconda environment named 'memsum' and activate it:

conda create -n memsum python=3.7
conda activate memsum

Install Jupyter Notebook and IPython within the environment:

conda install ipython
conda install -c anaconda jupyter

Step 2: Install Dependencies and Pretrained Model

Install required Python packages:

pip install -r requirements.txt

Install PyTorch with the appropriate CUDA toolkit version (replace <CUDA_VERSION> with your actual CUDA version):

conda install pytorch cudatoolkit=<CUDA_VERSION> -c pytorch -y

Download and set up pretrained GloVe word embeddings for the MemSum model:

python download_and_load_word_embedding.py

Step 3: Test Trained Model

Evaluate the full MemSum model's performance on the Pubmed's test set using the following command. The output includes ROUGE 1/2/L scores:

python my_test.py -model_type MemSum_Final \
-summarizer_model_path model/MemSum_Final/pubmed_full/200dim/best/model.pt \
-vocabulary_path model/glove/vocabulary_200dim.pkl \
-corpus_path data/pubmed_full/test_PUBMED.jsonl -gpu 0 \
-max_extracted_sentences_per_document 7 -p_stop_thres 0.6 \
-output_file results/MemSum_Final/pubmed_full/200dim/test_results.txt \
-max_doc_len 500 -max_seq_len 100

Step 4: Use Pretrained Summarizer Module

Load the full MemSum model:

from my_summarizers import ExtractiveSummarizer_MemSum_Final
memsum_model = ExtractiveSummarizer_MemSum_Final(
    "model/MemSum_Final/pubmed_full/200dim/best/model.pt",
    "model/glove/vocabulary_200dim.pkl",
    gpu=0, embed_dim=200, max_doc_len=500, max_seq_len=100
)

Prepare a document for summarization (The document should be a list of sentences):

import json
# Load the document
database = [json.loads(line) for line in open("data/pubmed_full/test_PUBMED.jsonl").readlines()]
pos = 6  # Example position in the database
document = database[pos]["text"]
gold_summary = database[pos]["summary"]

Generate a summary using MemSum:

extracted_summary = memsum_model.extract(
    [document], p_stop_thres=0.6, max_extracted_sentences_per_document=7, 
    return_sentence_position=False
)[0]

Evaluate the summary with ROUGE scores:

from rouge_score import rouge_scorer
rouge_cal = rouge_scorer.RougeScorer(['rouge1', 'rouge2', 'rougeLsum'], use_stemmer=True)
print(rouge_cal.score("\n".join(gold_summary), "\n".join(extracted_summary)))

Step 5: Train Model

Train the full MemSum model on the PubMed dataset by running the train.py script in the appropriate directory:

cd src/MemSum_Final/; python train.py -config_file_path config/pubmed_full/200dim/run0/training.config