run_mae_vit_base_p16_224_800ep.yaml

seed: 0
run_mode: 'train'
output_dir: './output'  # 当前不支持自定义修改，请勿修改该默认值
load_checkpoint: ''
src_strategy_path_or_dir: ''
auto_trans_ckpt: False  # If true, auto transform load_checkpoint to load in distributed model
only_save_strategy: False
resume_training: False

# context
context:
  mode: 0 #0--Graph Mode; 1--Pynative Mode
  device_target: "Ascend"
  enable_graph_kernel: False
  graph_kernel_flags: "--opt_level=0"
  max_call_depth: 10000
  save_graphs: False
  device_id: 0

# aicc
remote_save_url: "Please input obs url on AICC platform."

# runner
runner_config:
  epochs: 10
  batch_size: 64
  image_size: 224
  sink_mode: True
  sink_size: 2
runner_wrapper:
  type: MFTrainOneStepCell
  scale_sense:
    type: FixedLossScaleUpdateCell
    loss_scale_value: 1024

# parallel
use_parallel: False
parallel:
  parallel_mode: 0 # 0-data parallel, 1-semi-auto parallel, 2-auto parallel, 3-hybrid parallel
  gradients_mean: True
  enable_alltoall: False
  full_batch: False
  search_mode: "sharding_propagation"
  enable_parallel_optimizer: False
  strategy_ckpt_save_file: "./ckpt_strategy.ckpt"
parallel_config:
  data_parallel: 1
  model_parallel: 1
  expert_parallel: 1
  pipeline_stage: 1
  micro_batch_num: 1
  gradient_aggregation_group: 4
micro_batch_interleave_num: 1

# moe
moe_config:
  expert_num: 1
  capacity_factor: 1.05
  aux_loss_factor: 0.05
  num_experts_chosen: 1

# recompute
recompute_config:
  recompute: False
  parallel_optimizer_comm_recompute: False
  mp_comm_recompute: True
  recompute_slice_activation: False

# autotune
auto_tune: False
filepath_prefix: './autotune'
autotune_per_step: 10

# profile
profile: False
profile_start_step: 1
profile_stop_step: 10
init_start_profile: False
profile_communication: False
profile_memory: True

# Trainer
trainer:
  type: masked_image_modeling
  model_name: 'mae_vit_base_p16'

# train dataset
train_dataset: &train_dataset
  seed: 2022
  batch_size: 64
  data_loader:
    type: ImageFolderDataset
    dataset_dir: "imageNet-1k/train"
    num_parallel_workers: 8
    shuffle: True
  transforms:
    - type: RandomCropDecodeResize
      size: 224
      scale: [0.2, 1.0]
      interpolation: cubic
    - type: RandomHorizontalFlip
      prob: 0.5
    - type: Normalize
      mean: [123.675, 118.575, 103.53]
      std: [58.395, 62.22, 57.375]
    - type: HWC2CHW
  mask_policy:
    type: MaeMask
    input_size: 224
    patch_size: 16
    mask_ratio: 0.75
  input_columns: ["image"]
  output_columns: ["image", "mask", "ids_restore", "unmask_index"]
  column_order: ["image", "mask", "ids_restore", "unmask_index"]
  num_parallel_workers: 8
  python_multiprocessing: False
  drop_remainder: True
  repeat: 1
  numa_enable: False
  prefetch_size: 30
train_dataset_task:
  type: MIMDataset
  dataset_config: *train_dataset

# model
model:
  arch:
    type: ViTMAEForPreTraining
  model_config:
    type: ViTMAEConfig
    mask_ratio: 0.75  # mask ratio
    image_size: 224  # input image size
    patch_size: 16  # patch size
    num_channels: 3  # channels of input images
    initializer_range: 0.02  # initial std of cls_tokens and mask_tokens
    hidden_size: 768  # embedding dimension
    num_hidden_layers: 12  # number of transformer blocks
    num_attention_heads: 12  # number of attention heads
    intermediate_size: 3072  # ffn_hidden_size of encoder
    qkv_bias: True  # if has bias in qkv dense
    hidden_act: gelu  # activation of MLP
    post_layernorm_residual: False  # if using post layernorm residual
    layer_norm_eps: 0.000001  # eps of layer_norm
    attention_probs_dropout_prob: 0.  # drop rate of Attention
    hidden_dropout_prob: 0.  # drop rate of MLP
    decoder_hidden_size: 512  # decoder embedding dim
    decoder_num_hidden_layers: 8  # number of decoder transformer blocks
    decoder_num_attention_heads: 16  # number of decoder attention heads
    decoder_intermediate_size: 2048  # ffn_hidden_size of decoder
    norm_pix_loss: True  # if using norm pixel loss
    checkpoint_name_or_path: 'mae_vit_base_p16'

# lr sechdule
lr_schedule:
  type: cosine
  learning_rate: 0.00015
  lr_end: 0.
  warmup_lr_init: 0.
  warmup_epochs: 40
  total_steps: -1 # -1 means it will load the total steps of the dataset
layer_scale: False
layer_decay: 0.65

# optimizer
optimizer:
  type: adamw
  beta1: 0.9
  beta2: 0.95
  eps: 0.00000001 # 1e-8
  weight_decay: 0.05
lr_scale: True
lr_scale_factor: 256

# callbacks
callbacks:
  - type: MFLossMonitor
  - type: CheckpointMointor
    prefix: "mindformers"
    save_checkpoint_steps: 100
    integrated_save: True
    async_save: False
  - type: ObsMonitor
eval_callbacks:
  - type: ObsMonitor

# processor
processor:
  type: ViTMAEProcessor
  image_processor:
    type: ViTMAEImageProcessor
    size: 224  # input image size