seed: 0
run_mode: 'train'
output_dir: './output' # path to save checkpoint/strategy
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
  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: 100
  batch_size: 128
  image_size: 224
  sink_mode: True
  sink_size: 2
  num_classes: 1000
runner_wrapper:
  type: MFTrainOneStepCell
  scale_sense:
    type: FixedLossScaleUpdateCell
    loss_scale_value: 1024
  use_clip_grad: True
  max_grad_norm: 5.0


# 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: image_classification
  model_name: 'swin_base_p4w7'
# if True, do evaluate during the training process. if false, do nothing.
# note that the task trainer should support _evaluate_in_training function.
do_eval: False

# train dataset
train_dataset: &train_dataset
  seed: 0
  batch_size: 128
  data_loader:
    type: ImageFolderDataset
    dataset_dir: "imageNet-1k/train"
    num_parallel_workers: 8
    shuffle: True
  transforms:
    - type: Decode
    - type: RandomResizedCrop
      size: 224
      scale: [0.08, 1.0]
      interpolation: cubic
    - type: RandomHorizontalFlip
      prob: 0.5
    - type: ToPIL
    - type: rand_augment_transform
      config_str: "rand-m9-mstd0.5-inc1"
      hparams:
        translate_const: 100
        img_mean: [124, 116, 104]
        interpolation: cubic
    - type: ToTensor
    - type: Normalize
      mean: [0.485, 0.456, 0.406] # [123.675, 118.575, 103.53]
      std: [0.229, 0.224, 0.225] #
      is_hwc: False
    - type: RandomErasing
      probability: 0.25
      mode: pixel
      max_count: 1
  mixup_op:
    mixup_alpha: 0.8
    cutmix_alpha: 1.0
    cutmix_minmax:
    prob: 1.0
    switch_prob: 0.5
    label_smoothing: 0.1
  input_columns: ["image", "label"]
  output_columns: ["image", "label"]
  column_order: ["image", "label"]
  num_parallel_workers: 8
  python_multiprocessing: False
  drop_remainder: True
  repeat: 1
  numa_enable: False
  prefetch_size: 30
train_dataset_task:
  type: ImageCLSDataset
  dataset_config: *train_dataset

# eval dataset
eval_dataset: &eval_dataset
  seed: 0
  batch_size: 128
  data_loader:
    type: ImageFolderDataset
    dataset_dir: "imageNet-1k/val"
    num_parallel_workers: 8
    shuffle: True
  transforms:
    - type: Decode
    - type: Resize
      size: 256
      interpolation: cubic
    - type: CenterCrop
      size: 224
    - type: ToTensor
    - type: Normalize
      mean: [0.485, 0.456, 0.406]
      std: [0.229, 0.224, 0.225]
      is_hwc: False
  input_columns: ["image", "label"]
  output_columns: ["image", "label"]
  column_order: ["image", "label"]
  num_parallel_workers: 8
  python_multiprocessing: False
  drop_remainder: True
  repeat: 1
  numa_enable: False
  prefetch_size: 30
  eval_interval: 1
  eval_offset: 0
eval_dataset_task:
  type: ImageCLSDataset
  dataset_config: *eval_dataset

# model
model:
  arch:
    type: SwinForImageClassification
  model_config:
    type: SwinConfig
    num_labels: 1000 # num classes
    image_size: 224 # input image size
    patch_size: 4 # patch size
    num_channels: 3 # channels of input images
    embed_dim: 128 # embedding dimension
    depths: [2, 2, 18, 2] # number of transformer blocks for each swin layer
    num_heads: [4, 8, 16, 32] # number of attention heads for each swin layer
    window_size: 7 # window size for swin
    mlp_ratio: 4 # ffn_hidden_size = mlp_ratio * embed_dim
    qkv_bias: True # has transformer qkv bias or not
    layer_norm_eps: 0.00001  # eps of layer_norm
    hidden_dropout_prob: 0. # drop rate of MLP
    attention_probs_dropout_prob: 0. # drop rate of Attention
    drop_path_rate: 0.1 # drop path rate of transformer blocks
    use_absolute_embeddings: False # if using absolute position embedding
    patch_norm: True # use norm in SwinPatchEmbeddings
    hidden_act: gelu # activation of MLP
    weight_init: normal # weight initialize type
    loss_type: SoftTargetCrossEntropy # loss type
    checkpoint_name_or_path: swin_base_p4w7

# lr schedule
lr_schedule:
  type: cosine
  learning_rate: 0.00125
  lr_end: 0.00000025
  warmup_lr_init: 0.00000025
  warmup_epochs: 20
  total_steps: -1 # -1 means it will load the total steps of the dataset
layer_scale: True
layer_decay: 0.9

# optimizer
optimizer:
  type: adamw
  beta1: 0.9
  beta2: 0.999
  eps: 0.00000001 # 1e-8
  weight_decay: 0.05
lr_scale: True
lr_scale_factor: 512

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

# metric
metric:
  type: Accuracy
  eval_type: classification

# processor
processor:
  type: SwinProcessor
  image_processor:
    type: SwinImageProcessor
    size: 224  # input image size