---
exp_name: default
seed: 1
num_workers: 8
debug: false
use_gpu: true
torch_dist:
  use: true
  amp: false
  gpus_per_model: 1
  backend: nccl
  init_method: env://
batch_size_per_gpu: 
  continuous: 64
  seq: 64
  generate: 64
exp_dir: 
  home: ''
data:
  name: pmc
  path: 
    home: ''
  dataset:
    tasks: ['categorical','series_name','axis','caption','data']
    chart_text:
      tgt_token: <mask_1>
      window_size: 16
      widen_rate: 0.1
      max_widen_len: 1
      min_sent_len: 10
    chart_data:
      chart_tasks: ['task6']
      active_charts: []
      norm_mode: minmax
      scale_mode: log10
      scale_eps: [1.100001, 1.100001]
      scale_floor: [0.000001, 0.000001]
      sep_token: <SEP>
model:
  active: ['continuous']
  seperate_data_task: True
  seq:
    opt_mode: 1 #Stage 1: text only, Stage 2: Data only, Stage 0: Both
    freeze_encoder: False
    decoder2_num_layers: 8
    hf_model: 
      name: google/t5-v1_1-large
      tgt_token: <mask_1>
      model_max_length: 512
      max_source_len: 1024 
      max_target_len: 256 
      pad_to_max_len: True
      ignore_pad_token_for_loss: True
      use_fast: True
  continuous_data:
    active: ['scale', 'continuous'] 
    use_pos_embs: true
    encoder:
      chart_type_conditional: False
      max_blocks:
        series: 64
        points: 256
      conv:
        channels: [256, 128, 64]
        kernels: [3,3,3,2]
        stride: 1
        padding: 1
        use_bn: false
        n_res_block: 2
        n_res_channel: 32
        res_kernels: [3,1]
        res_padding: 1
      transformer:
        use: true
        name: gpt
        n_layer: 4
        n_head: 4
    decoder:
      chart_type_conditional: True
      conv:
        channels: [32, 64, 128] 
        kernels: [3,3,3,2]
        stride: 1
        padding: 1
        use_bn: false
        n_res_block: 2
        n_res_channel: 32
        res_kernels: [3,1]
        res_padding: 1
        use_proj: false
      transformer:
        use: true
        name: t5_decoder 
        n_layer: 4
        n_head: 4
        d_kv: 8
        num_buckets: 8
        max_distance: 32
      scale:
        n_head: 1 
    disc:
      use: True
      disc_start: 100
      disc_loss: hinge
      disc_factor: 1.0
      disc_weight: 1.0
      disc_conditional: True
      use_pos_embs: True
      conv:
        channels: [256, 128, 64]
        kernels: [3,3,3,2]
        stride: 1
        padding: 1
        use_bn: false
        n_res_block: 2
        n_res_channel: 32
        res_kernels: [3,1]
        res_padding: 1
      transformer:
        use: true
        name: gpt
        n_layer: 4
        n_head: 4
    vq:
      name: vq
      emb_dim1: 32
      emb_dim2: 32
      n_emb1: 128
      n_emb2: 128
      emb_len1: 16
      emb_len2: 12
      beta: 0.25
      ema_update: true
      random_restart: true
      tiled: true
    mhd:
      use: true
      name: mhd1d
      bottleneck: False
      bottleneck_dim: 64
      hypothese_count: 256 
      hypothese_bsz: 256
      residual: True
      loss_reduce: mean
      dist_reduce: mean
      loss_reduce_dims: [-2, -1]
      norm: True
      act: relu
      decoder_loss: winner 
      dist_loss: mse 
      gamma: 0.25
      dropout_rate: 0.5
      decoder:
        act: leakyrelu
        n_res_block: 2
        res_kernels: [3,1]
        n_res_channel: 8
    fid:
      use_pos_embs: True
      enc_out_dim: 32
      conv:
        channels: [256, 128, 64]
        kernels: [3,3,3,2]
        stride: 1
        padding: 1
        use_bn: false
        n_res_block: 2
        n_res_channel: 32
        res_kernels: [3,1]
        res_padding: 1
      transformer:
        use: true
        name: gpt
        n_layer: 4
        n_head: 4
    loss_fn:
      scale: l1
      continuous: mse
train:
  epochs:
    warmup: 0
    continuous: 150
    seq: 50
    total: 100
  gradient_accum_steps: 1 
  max_grad_norm: 1.0
  loss_weights:
    text: 1.0
    code: 1.0  
    cb1: 1.0
    cb2: 1.0
    wta: 1.0
    continuous: 1.0
    scale: 1.0
    categorical: 1.0
    series_name: 1.0
    ct: 1.0
    row: 1.0
    col: 1.0
    gan: 1.0
    disc: 1.0
  optim:
    type: AdamW
    learning_rate: 0.0005
    betas: [0.9, 0.999]
    eps: 1e-8
    weight_decay: 3e-7
  label_smoothing_factor: 0.1
  intervals:
    eval: 1
    snapshot: 20
    display: 20000
    gen: 20000
    val: 20000
eval:
  fid: False
  val_step_count: 50
  write_samples: False
  gen_steps: 2
  max_steps: 1000000
  sample_interval: 50
  display_interval: 50
  sample_epoch: 50
  include_inputs_for_metrics: False
  eval_accumulation_steps: 10
  num_beams: 3
  max_length: 256
  repetition_penalty: 1.0
  gen_temperature: 1.0
  hypo_count: 1
  hypo_bsz: 1
  ksm:
    active: true
    name: google/pegasus-pubmed
    use_fast: true
fp16: 
  use: false
  eval: false
  loss_scale: 1.0
  initial_scale_power: 64
  loss_scale_window: 1000
  hysteresis: 2
  min_loss_scale: 1000
  opt_level: O3
timeout: 3600