config_sample.yaml

mode: train   # train/test
save_dir: scratch/checkpoints/26_sgd_diff_large_gru_mlp   # directory to save checkpoint, write logs/results, and load checkpoints from
seed: 23443521      # random seed
use_gpu: true       # whether to use gpu/cuda (default=false)
data:
  dir: scratch      # path to data directory to be used as prefix vocabulary files and data jsons
  review_vocab: review_vocab.pkl
  summary_vocab: summary_vocab.pkl
  train: {
    batch_size: 32,
    jsonfile: audio_train.json.preprocessed     # preprocessed train data
  }
  val: {
    batch_size: 32,
    jsonfile: audio_dev.json.preprocessed       # preprocessed train data
  }
  weights: [4.0,4.0,1.0]      # may be used to specify class weights to tackle class imbalance
                              # possible values:
                              # - list of floats - directly used as class weights
                              # - 'weighted' - use torch.utils.data.sampler.WeightedRandomSampler
                              # - false - no class reweighting (default)
model:
  params: {
    combined_lookup: true,    # use same embeddings for both review words and summary words (default=false)
    word_emb_dim: 200,        # word embedding dimension
    rnn_type: gru,            # rnn type used in HAN (lstm/gru)
    rnn_hidden_dim: 100,
    emb_dim: 200,             # dimension of embeddings computed by each layer in HAN - both sentence emb and document emb
    use_summary: true,        # whether to use summary features (default=true)
    use_summ_mlp: true,       # whether to pass summary emb through mlp before concatenation with review emb (default=false)
  }
  reload: checkpoint.pth.tar  # name of checkpoint (in save_dir) to load (use fresh model is not provided)
optim:
  class: sgd                # class of optmizer (sgd/adam)
  params: {
    lr: 0.01,               # learning rate of optimizer (0.001 for adam, 0.01 for sgd)
    weight_decay: 0.0005    # weight decay to use
  }
  scheduler: {
    factor: 0.4,       # factor to use for torch.optim.lr_scheduler.ReduceLROnPlateau
    patience: 3        # patience to use for torch.optim.lr_scheduler.ReduceLROnPlateau
  }
training: {
  n_epochs: 5,
  start_from_checkpoint: true   # whether to start from loaded checkpoint
}