config.yaml

defaults:
  - model: nrms
  - embedding_layer: word_embedding
  - hparams_model: ${model}
  - hparams_embedding: ${embedding_layer}
  - optim: ${embedding_layer}
  - data_path: ${embedding_layer}
  - dataset: ${embedding_layer}
  - deterministic: disabled
  - optional debug_run: disabled
  - _self_
default_choices: ${hydra:runtime.choices}
dataset_type: small
num_workers: 4
data_path:
  train:
    samples: ${oc.env:WORKDIR}/dataset/mind/${dataset_type}/processed/train/behaviors.pkl
    articles:  ${oc.env:WORKDIR}/dataset/mind/${dataset_type}/processed/train/articles.pkl
  valid:
    samples: ${oc.env:WORKDIR}/dataset/mind/${dataset_type}/processed/valid/behaviors.pkl
    articles:  ${oc.env:WORKDIR}/dataset/mind/${dataset_type}/processed/valid/articles.pkl
  category: ${oc.env:WORKDIR}/dataset/mind/${dataset_type}/processed/categories.txt
  subcategory: ${oc.env:WORKDIR}/dataset/mind/${dataset_type}/processed/subcategories.txt
hparams:
  # value of ${default_choices} becomes invalid under multiran setting (hydra v1.1):  https://github.com/facebookresearch/hydra/issues/1882
  model: ${default_choices.model}
  embedding_layer: ${default_choices.embedding_layer}
  article_attributes:
    - title
    - body
    - category
    - subcategory
  embedding_dim: 400
  attn_hidden_dim: 200
  category_hidden_dim: 100
  p_dropout: 0.2
  n_categories: ${line_count:${data_path.category}}
  n_subcategories: ${line_count:${data_path.subcategory}}
  data_shuffle_seed: 0
  train_seed: 0
  n_negatives: 4
  n_epochs: 3
  max_history_length: 50
  optim: ${optim}
data_loader:
  train:
    _target_: torch.utils.data.DataLoader
    batch_size: ${hparams.batch_size.train}
    shuffle: True
    num_workers: ${num_workers}
    collate_fn:
      _target_: mind_recommenders_pytorch.train.data.input_collator.InputCollator
  valid:
    _target_: torch.utils.data.DataLoader
    batch_size: ${hparams.batch_size.valid}
    shuffle: False
    num_workers: ${num_workers}
    collate_fn:
      _target_: mind_recommenders_pytorch.train.data.input_collator.InputCollator
pl_model:
  _target_: mind_recommenders_pytorch.train.pl_model.PlModel
trainer:
  _target_: pytorch_lightning.Trainer
  default_root_dir: "./"
  max_epochs: ${hparams.n_epochs}
  gpus: 
    - 0
  detect_anomaly: False
  precision: ${hparams.precision}
  callbacks:
    - _target_: pytorch_lightning.callbacks.ModelCheckpoint
      monitor: val_auc
      filename: '{epoch:04d}-{step:04d}-{val_auc:.3f}'
      save_top_k: 10
  logger:
    _target_: pytorch_lightning.loggers.TensorBoardLogger
    save_dir: ${trainer.default_root_dir} 
    name: train_logs


hydra:
  run:
    dir: ./logs/${hydra.job.override_dirname}
  sweep:
    dir: ./logs/
    subdir: ${hydra.job.override_dirname}