timm

/

vit_base_patch16_224.augreg2_in21k_ft_in1k

like 7

aa: rand-m8-inc1-mstd101
amp: true
amp_dtype: float16
amp_impl: native
aot_autograd: false
apex_amp: false
aug_repeats: 0
aug_splits: 0
batch_size: 512
bce_loss: false
bce_target_thresh: null
bn_eps: null
bn_momentum: null
channels_last: false
checkpoint_hist: 10
class_map: ''
clip_grad: 2.0
clip_mode: norm
color_jitter: 0.4
cooldown_epochs: 10
crop_pct: 1.0
cutmix: 1.0
cutmix_minmax: null
data_dir: /data/imagenet/
dataset: ''
dataset_download: false
decay_epochs: 100
decay_milestones:
- 30
- 60
decay_rate: 0.1
dist_bn: reduce
drop: 0.0
drop_block: null
drop_connect: null
drop_path: 0.1
epoch_repeats: 0.0
epochs: 50
eval_metric: top1
experiment: ''
fast_norm: false
fuser: ''
gp: null
grad_checkpointing: true
hflip: 0.5
img_size: null
in_chans: null
initial_checkpoint: ''
input_size: null
interpolation: ''
jsd_loss: false
layer_decay: 0.7
local_rank: 0
log_interval: 50
log_wandb: false
lr: 0.0002
lr_base: 0.1
lr_base_scale: ''
lr_base_size: 256
lr_cycle_decay: 0.5
lr_cycle_limit: 1
lr_cycle_mul: 1.0
lr_k_decay: 1.0
lr_noise:
- 0.1
- 0.9
lr_noise_pct: 0.67
lr_noise_std: 1.0
mean: null
min_lr: 5.0e-07
mixup: 0.8
mixup_mode: batch
mixup_off_epoch: 0
mixup_prob: 1.0
mixup_switch_prob: 0.5
model: vit_base_patch16_224.augreg_in21k
model_ema: true
model_ema_decay: 0.9998
model_ema_force_cpu: false
momentum: 0.9
native_amp: false
no_aug: false
no_ddp_bb: false
no_prefetcher: false
no_resume_opt: false
num_classes: 1000
opt: adamw
opt_betas: null
opt_eps: null
output: ''
patience_epochs: 10
pin_mem: false
pretrained: true
ratio:
- 0.75
- 1.3333333333333333
recount: 1
recovery_interval: 0
remode: pixel
reprob: 0.3
resplit: false
resume: ''
save_images: false
scale:
- 0.08
- 1.0
sched: cosine
sched_on_updates: true
seed: 42
smoothing: 0.1
split_bn: false
start_epoch: null
std: null
sync_bn: false
torchscript: false
train_interpolation: random
train_split: train
tta: 0
use_multi_epochs_loader: false
val_split: validation
validation_batch_size: null
vflip: 0.0
warmup_epochs: 10
warmup_lr: 0.0
warmup_prefix: true
weight_decay: 0.05
worker_seeding: all
workers: 8

aa: rand-m8-inc1-mstd101
amp: true
amp_dtype: float16
amp_impl: native
aot_autograd: false
apex_amp: false
aug_repeats: 0
aug_splits: 0
batch_size: 512
bce_loss: false
bce_target_thresh: null
bn_eps: null
bn_momentum: null
channels_last: false
checkpoint_hist: 10
class_map: ''
clip_grad: 3.0
clip_mode: norm
color_jitter: 0.4
cooldown_epochs: 10
crop_pct: 1.0
cutmix: 0.0
cutmix_minmax: null
data: /data/imagenet/
data_dir: /data/imagenet/
dataset: ''
dataset_download: false
decay_epochs: 100
decay_milestones:
- 30
- 60
decay_rate: 0.1
dist_bn: reduce
drop: 0.0
drop_block: null
drop_connect: null
drop_path: 0.1
dynamo: false
dynamo_backend: null
epoch_repeats: 0.0
epochs: 50
eval_metric: top1
experiment: ''
fast_norm: false
fuser: ''
gp: null
grad_checkpointing: true
hflip: 0.5
img_size: null
in_chans: null
initial_checkpoint: ''
input_size: null
interpolation: ''
jsd_loss: false
layer_decay: 0.75
local_rank: 0
log_interval: 50
log_wandb: false
lr: 0.0001
lr_base: 0.1
lr_base_scale: ''
lr_base_size: 256
lr_cycle_decay: 0.5
lr_cycle_limit: 1
lr_cycle_mul: 1.0
lr_k_decay: 1.0
lr_noise:
- 0.1
- 1.0
lr_noise_pct: 0.67
lr_noise_std: 1.0
mean: null
min_lr: 5.0e-07
mixup: 0.3
mixup_mode: batch
mixup_off_epoch: 0
mixup_prob: 1.0
mixup_switch_prob: 0.5
model: convnext_small.in12k
model_ema: true
model_ema_decay: 0.9998
model_ema_force_cpu: false
momentum: 0.9
native_amp: false
no_aug: false
no_ddp_bb: false
no_prefetcher: false
no_resume_opt: false
num_classes: 1000
opt: adamw
opt_betas: null
opt_eps: null
output: ''
patience_epochs: 10
pin_mem: false
pretrained: true
ratio:
- 0.75
- 1.3333333333333333
recount: 1
recovery_interval: 0
remode: pixel
reprob: 0.3
resplit: false
resume: ''
save_images: false
scale:
- 0.08
- 1.0
sched: cosine
sched_on_updates: false
seed: 42
smoothing: 0.1
split_bn: false
start_epoch: null
std: null
sync_bn: false
torchcompile: null
torchscript: false
train_interpolation: random
train_split: train
tta: 0
use_multi_epochs_loader: false
val_split: validation
validation_batch_size: null
vflip: 0.0
warmup_epochs: 10
warmup_lr: 1.0e-06
warmup_prefix: false
weight_decay: 0.05
worker_seeding: all
workers: 8

@pyone those are two sets of hparams, a bit diff but same theme that ended up similar result for my 'augreg2' runs, basically re-finetuning the in21k models from 'How to train your vit' with better params, key ingredient is the layer-wise lr decay (lr_decay arg)

that was either 4 or 8 gpus (to calc global batch size), I think it was 4...

Hey, how to get rid of Unknown model error
it seems that we can only upload models in the range that timm supports

How about uploading a customized model? U c, many researchers may come up with new model structure with new name in cv domain, how about them?

@rwightman Thank you very much. It is clear.