configs/peptune_config.yaml

noise:
  type: loglinear
  sigma_min: 1e-4
  sigma_max: 20
  state_dependent: True

mode: ppl_eval  # train / ppl_eval / sample_eval
diffusion: absorbing_state
vocab: old_smiles # old_smiles / new_smiles / selfies / helm
backbone: roformer  # peptideclm / helmgpt / dit / roformer / finetune_roformer
parameterization: subs  # subs
time_conditioning: False
T: 0  # 0 (continuous time) / 1000 
subs_masking: False

seed: 42

mcts: 
  num_children: 50
  num_objectives: 5
  topk: 100
  mask_token: 4
  num_iter: 128
  sampling: 0 # 0 is gumbel sampling / > 0 samples children from top k probs
  invalid_penalty: 0.5
  sample_prob: 1.0
  perm: True
  dual: False
  single: False
  time_dependent: True

lr_scheduler:
  _target_: transformers.get_constant_schedule_with_warmup
  num_warmup_steps: 2500

data:
  train: To Be Added
  valid: To Be Added
  batchinohup ng: wrapping # padding / wrapping

loader:
  global_batch_size: 64
  eval_global_batch_size: ${.global_batch_size}
  # Note: batch_size and eval_batch_size are **per machine**
  batch_size: ${div_up:${.global_batch_size}, ${eval:${trainer.devices} * ${trainer.num_nodes}}}
  eval_batch_size: ${div_up:${.eval_global_batch_size}, ${eval:${trainer.devices} * ${trainer.num_nodes}}}
  num_workers: ${eval:"len(__import__('os').sched_getaffinity(0))"}
  pin_memory: True

sampling:
  predictor: ddpm_cache  # analytic, ddpm, ddpm_cache
  num_sequences: 100
  sampling_eps: 1e-3
  steps: 128
  seq_length: 100
  noise_removal: True
  num_sample_batches: 2  # Total samples: `num_gpus` * `loader.eval_batch_size` * num_sample_batches
  num_sample_log: 2
  stride_length: 1
  num_strides: 1

training:
  antithetic_sampling: True
  sampling_eps: 1e-3
  focus_mask: False
  #dynamic_batching: True
  accumulator: False

eval:
  checkpoint_path: 
  disable_ema: False
  compute_generative_perplexity: False
  perplexity_batch_size: 8
  compute_perplexity_on_sanity: False
  gen_ppl_eval_model_name_or_path: gpt2-large  # gpt2-large, meta-llama/Llama-2-7b-hf
  generate_samples: True
  generation_model: 
  
optim:
  weight_decay: 0.075
  lr: 3e-4
  beta1: 0.9
  beta2: 0.999
  eps: 1e-8

pepclm:
  hidden_size: 768
  cond_dim: 256
  n_heads: 20
  n_blocks: 4
  dropout: 0.5
  length: 512
  #scale_by_sigma: True

model:
  type: ddit
  hidden_size: 768
  cond_dim: 128
  length: 512
  n_blocks: 12
  n_heads: 12
  scale_by_sigma: True
  dropout: 0.1

roformer:
  hidden_size: 768
  n_layers: 8
  n_heads: 8
  max_position_embeddings: 1035

helmgpt:
  hidden_size: 256
  embd_pdrop: 0.1
  resid_pdrop: 0.1
  attn_pdrop: 0.1
  ff_dropout: 0.
  block_size: 140
  n_layer: 8
  n_heads: 8


trainer:
  _target_: lightning.Trainer
  accelerator: cuda
  num_nodes: 1
  devices: ${device_count:}
  accumulate_grad_batches: ${div_up:${loader.global_batch_size}, ${eval:${trainer.devices} * ${loader.batch_size} * ${trainer.num_nodes}}}
  gradient_clip_val: 1.0
  precision: 64-true
  num_sanity_val_steps: 2
  max_epochs: 100
  max_steps: 1_000_000
  log_every_n_steps: 10
  limit_train_batches: 1.0   # train on full dataset, can be used to toggle quick run
  limit_val_batches: 1.0     # validate on full dataset, can be used to toggle quick run
  #val_check_interval: 40 #954
  check_val_every_n_epoch: 1

hydra:
  run:
    dir: ./${now:%Y.%m.%d}/
  job:
    chdir: True

checkpointing:
  # Use custom `save_dir` if, e.g., saving to S3 bucket, otherwise leave this parameter as is
  save_dir: ${cwd:}
  # Note: `checkpoints` path should correspond to `checkpoint_every_n_steps.dirpath`
  resume_from_ckpt: True
  resume_ckpt_path:

callbacks:
  model_checkpoint:
    _target_: pytorch_lightning.callbacks.ModelCheckpoint
    every_n_epochs: 1
    monitor: "val/nll"
    save_top_k: 10
    mode: "min"
    dirpath: