patch_size/patch8/pretrain/config.yaml

name: patch_size/patch8/pretrain
notes: patch_size ablations patch8 (patch_size=8 mask_patch_size=16)
output_dir: experiments/patch_size/output/patch_size/patch8/pretrain
input_space: flat
patch_size: 8
num_frames: 16
t_patch_size: 4
mask_ratio: 0.9
pred_mask_ratio: null
masking: tube
masking_kwargs: {}
mask_patch_size: 16
model: mae_vit_base
model_kwargs:
  decoding: attn
  pos_embed: sep
  target_norm: null
  pca_norm_nc: 2
  t_pred_stride: 2
  no_decode_pos: true
  mask_drop_scale: false
  pred_edge_pad: 0
  gauss_sigma: null
  class_token: true
  reg_tokens: 0
  no_embed_class: true
  head_init_scale: 0.0
  decoder_depth: 4
  drop_path_rate: 0.0
datasets:
  hcp-train:
    type: wds
    url: /data/fmri-datasets/pretrain/hcpya-all.${input_space}.wds/hcpya-all-${input_space}-{00000..01799}.tar
    clipping: random
    clipping_kwargs:
      oversample: 4.0
    shuffle: true
    buffer_size: 2000
    samples_per_epoch: 200000
  hcp-train-subset:
    type: arrow
    root: s3://medarc/fmri-datasets/eval/hcpya-clips.${input_space}.arrow/validation
    split_range:
    - 0
    - 2000
    shuffle: false
  hcp-val:
    type: arrow
    root: s3://medarc/fmri-datasets/eval/hcpya-clips.${input_space}.arrow/test
    split_range:
    - 0
    - 2000
    shuffle: false
train_dataset: hcp-train
eval_datasets:
- hcp-train-subset
- hcp-val
val_dataset: null
clip_vmax: 3.0
normalize: frame
tr_scale: null
crop_scale: null
crop_aspect: null
gray_jitter: null
num_workers: 16
epochs: 100
batch_size: 32
accum_iter: 1
base_lr: 0.001
min_lr: 0.0
warmup_epochs: 5
weight_decay: 0.05
betas:
- 0.9
- 0.95
clip_grad: 1.0
amp: true
amp_dtype: float16
ckpt: null
resume: true
auto_resume: true
start_epoch: 0
max_checkpoints: 0
checkpoint_period: null
plot_period: 1
device: cuda
presend_cuda: false
seed: 6500
debug: false
wandb: true
wandb_entity: null
wandb_project: fMRI-foundation-model
rank: 0
world_size: 1
gpu: 0
distributed: true
dist_backend: nccl
in_chans: 1
img_size:
- 224
- 560