p4zaa / otto-multi-objective-recommender-system Goto Github PK

@torch.no_grad()
def recommendation(user_id, model, x_dict, edge_index_dict):
  # Get model decoder
  #model = Model(**model_params)
  with torch.no_grad():
    encoder = model.encoder(data.x_dict, data.edge_index_dict)

  # Get node representations for users and movies
  user_representations = encoder['session']
  movie_representations = encoder['aid']

  # Compute the dot product between user and movie representations to get edge weights
  edge_weights = user_representations.mm(movie_representations.T)

  # Make predictions for each user by taking the top k largest edge weights
  k = 20  # number of recommendations to make
  _, top_k_indices = edge_weights.topk(k, dim=1)
  recommendations = top_k_indices.numpy()

  # Print recommendations for the first user
  print(f'Recommendations for user {user_id}: {recommendations[user_id]}')

This line of code below cause memory crash ;(

session = 1
recommendation(session, model, Rtest_data.x_dict, Rtest_data.edge_index_dict)

Adding this layer consume high GPU RAM
No CUDA GPU will cause a very slow training. 🥲

Data Loader and Preprocessing

• Using Polars library for better memory and time efficient

  Use Polars for load and sample dataset and then convert to Pandas

• Create new DataFrame sample and store in drive

  trained_model cannot evaluate in new test set where it wasn't trained
  Created with 10% samples dataset

• Create separate source_idx and target_idx files with all nodes for later use (load in parquet when used)

  Use all nodes dict index will cause IndexError: index out of range in self

Modeling

• Create node-level features (which do not have from start) using torch.nn.Embedding

  REF

Training by mini-batch

• Make train function return model
• Save model after finished train: torch.save(model, 'path')

Encoding by mini-batch

• Load saved model: model = torch.load('path')

Validation

• Create new validation function

  See example here

Memory improvement

• Add gc.collect() to the code for clear up memory
• Add torch.cuda.empty_cache() to clear up GPU RAM

Additional

• Considering using Node2Vec for initialize node features and edge features

  Example

• Considering using Word2Vec

  Example

• ts can be used as edge features (depends on task)