timm config for training an nfnet, load with --config arg, override batch size, lr for your number of GPUs/dist nodes

effres-agc.yaml

aa: rand-m6-n5-inc1-mstd1.0
amp: false
apex_amp: false
aug_splits: 0
batch_size: 256
bn_eps: null
bn_momentum: null
bn_tf: false
channels_last: false
checkpoint_hist: 10
clip_grad: 0.02
clip_mode: agc
color_jitter: 0.4
cooldown_epochs: 10
crop_pct: 0.94
cutmix: 0.0
cutmix_minmax: null
data: /data/imagenet/
data_dir: /imagenet/
dataset: ''
decay_epochs: 1.0
decay_rate: 0.988
dist_bn: reduce
drop: 0.4
drop_block: null
drop_connect: null
drop_path: 0.25
epochs: 656
eval_metric: top1
gp: fast
hflip: 0.5
img_size: 320
initial_checkpoint: ''
input_size: null
interpolation: ''
jsd: false
local_rank: 0
log_interval: 50
lr: 0.5
lr_cycle_limit: 1
lr_cycle_mul: 1.0
lr_noise: null
lr_noise_pct: 0.67
lr_noise_std: 1.0
mean: null
min_lr: 1.0e-05
mixup: 0.2
mixup_mode: batch
mixup_off_epoch: 0
mixup_prob: 1.0
mixup_switch_prob: 0.5
model: efficientnet_v2m
model_ema: true
model_ema_decay: 0.999975
model_ema_force_cpu: false
momentum: 0.9
native_amp: true
no_aug: false
no_prefetcher: false
no_resume_opt: false
num_classes: 1000
opt: fusedsgd
opt_betas: null
opt_eps: 0.001
output: ''
patience_epochs: 10
pin_mem: false
pretrained: false
ratio:
- 0.67
- 1.5
recount: 3
recovery_interval: 0
remode: pixel
reprob: 0.5
resplit: false
resume: ''
save_images: false
scale:
- 0.08
- 1.0
sched: cosine
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_multiplier: 1
vflip: 0.0
warmup_epochs: 10
warmup_lr: 1.0e-06
weight_decay: 7.0e-06
workers: 5

nfnet.yaml

aa: rand-m6-n4-inc1-mstd1.0
amp: false
apex_amp: false
aug_splits: 0
batch_size: 256
bn_eps: null
bn_momentum: null
bn_tf: false
channels_last: false
checkpoint_hist: 10
clip_grad: 0.015
clip_mode: agc
color_jitter: 0.4
cooldown_epochs: 10
crop_pct: 0.94
cutmix: 0.0
cutmix_minmax: null
data: /data/imagenet/
data_dir: /imagenet/
dataset: ''
decay_epochs: 1.0
decay_rate: 0.988
dist_bn: reduce
drop: 0.375
drop_block: null
drop_connect: null
drop_path: 0.25
epochs: 656
eval_metric: top1
gp: fast
hflip: 0.5
img_size: 256
initial_checkpoint: ''
input_size: null
interpolation: ''
jsd: false
local_rank: 0
log_interval: 50
lr: 0.5
lr_cycle_limit: 1
lr_cycle_mul: 1.0
lr_noise: null
lr_noise_pct: 0.67
lr_noise_std: 1.0
mean: null
min_lr: 1.0e-05
mixup: 0.2
mixup_mode: batch
mixup_off_epoch: 0
mixup_prob: 1.0
mixup_switch_prob: 0.5
model: eca_nfnet_l1
model_ema: true
model_ema_decay: 0.999975
model_ema_force_cpu: false
momentum: 0.9
native_amp: true
no_aug: false
no_prefetcher: false
no_resume_opt: false
num_classes: 1000
opt: fusedsgd
opt_betas: null
opt_eps: 0.001
output: ''
patience_epochs: 10
pin_mem: false
pretrained: false
ratio:
- 0.67
- 1.5
recount: 3
recovery_interval: 0
remode: pixel
reprob: 0.5
resplit: false
resume: ''
save_images: false
scale:
- 0.08
- 1.0
sched: cosine
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_multiplier: 1
vflip: 0.0
warmup_epochs: 10
warmup_lr: 1.0e-06
weight_decay: 7.0e-06
workers: 5

This was used on 4x 32GB V100.

If you need to change the batch size and number of GPUs, rescale the LR linearly ...
old_global_batch = 4 * 256
new_lr = new_global_batch_size * 0.5 / 1024

Hi Ross, I have seen various sets of Hparams for your certain models. I know the sets come from your empirical, experimental and sensible abilities. Can you share or write any statement of your experience. I am sure It will be helpful for community to save our resources.
cheers
Linh