program: train.py
entity: dalle-mini
project: dalle-mini
method: random
metric:
  name: eval/loss
  goal: minimize
parameters:
  learning_rate:
    distribution: log_uniform
    # from exp(min) to exp(max)
    min: -6.9
    max: -3.5
  gradient_accumulation_steps:
    value: 8
  warmup_steps:
    value: 4000
#TODO: outdated command
command:
  - python3
  - ${program}
  - "--tokenizer_name"
  - "boris/dalle-mini-tokenizer"
  - "--config_name"
  - "facebook/bart-large-cnn"
  - "--dataset_repo_or_path"
  - "boris/gis_vqgan_f16_16384"
  - "--streaming"
  - "--use_auth_token"
  - "--image_vocab_size"
  - 16384
  - "--image_length"
  - 256
  - "--normalize_text"
  - True
  - "--per_device_train_batch_size"
  - 56
  - "--per_device_eval_batch_size"
  - 56
  - "--adafactor"
  - "--do_train"
  - "--do_eval"
  - "--num_train_epochs"
  - 1
  - "--logging_steps"
  - 40
  - "--eval_steps"
  - 800
  - "--output_dir"
  - "./output"
  - "--overwrite_output_dir"
  - "--max_train_samples"
  - 10000000
  - ${args}