volo_d1_ascend.yaml

# system
mode: 0
distribute: True
num_parallel_workers: 8
val_while_train: True
val_interval: 1

# dataset
dataset: 'imagenet'
data_dir: '/path/to/imagenet'
shuffle: True
dataset_download: False
batch_size: 128
drop_remainder: True

# augmentation
image_resize: 224
scale: [0.08, 1.0]
ratio: [0.75, 1.33]
hflip: 0.5
vflip: 0.0
interpolation: 'bicubic'
auto_augment: 'randaug-m9-mstd0.5-inc1'
re_prob: 0.25
mixup: 0.2
cutmix: 1.0
cutmix_prob: 1.0
crop_pct: 0.96
color_jitter: [0.4, 0.4, 0.4]

# model
model: 'volo_d1'
num_classes: 1000
pretrained: False
ckpt_path: ''
keep_checkpoint_max: 10
ckpt_save_dir: './ckpt'
ckpt_save_policy: 'top_k'
drop_path_rate: 0.1
dataset_sink_mode: True
amp_level: 'O3'
val_amp_level: 'O3'
ema: True
ema_decay: 0.9995

# loss
loss: 'CE'
label_smoothing: 0.1

# lr scheduler
scheduler: 'cosine_decay'
lr: 0.0019
min_lr: 0.00001
epoch_size: 300
warmup_epochs: 20
decay_epochs: 280
decay_rate: 0.1

# optimizer
opt: 'adamw'
weight_decay: 0.05
momentum: 0.9
loss_scale_type: 'dynamic'
loss_scale: 1024
use_nesterov: False
drop_overflow_update: True