halvardbariller / molecule-retrieval-nl-queries

Molecule Retrieval with Natural Language Queries. Challenge for the MVA Class "Advanced Learning for Text and Graph Data", 2023.

molecule-retrieval-nl-queries

Molecule Retrieval with Natural Language Queries. Challenge for the MVA Class "Advanced Learning for Text and Graph Data", 2023.

Description of runnable files

In all these runnable files, hyperparameters, model definitions and all the code that needs to be changed according to your needs are indicated with a comment header, such as:

###########################
#      GRAPH ENCODER      #
###########################
# Define your graph encoder here. Below are some of the encoders that we tried.

main.py

Trains a model. The model to train including hyperparameters is defined in this file (but the actual implementation is split into several files). The results are logged on Weights and Biases. Can also resume training from a checkpoint file.

Usage: python main.py [-n run_name_for_wandb] [-m path/to/pretrained_model.pt] (make sure that the right model is defined in the code before using a pretrained model)

eval.py

Computes predictions on the test set for a single pre-trained model.

Usage: after defining the model corresponding to your checkpoint in eval.py, run: python eval.py path/to/model.pt

average_predictions.py

Computes predictions by aggregating several models in two ways:

• by summing the normalized average similarity scores of each model,
• by summing the ranks computed with the similarity scores of each model (hard voting).

Usage: define the models and set the checkpoint paths directly in the code in the form of a dictionary, add it to the list of models from which the predictions are computed, then run python average_predictions.py.