You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Hi,
Thanks for sharing the interesting work.
I tried to finetune the model on FMOW-RGB dataset by using the pretrained weights of 800th epoch provided at the repository. I used the default hyperparameters in the file main_linprobe.py and set
args.finetune=True
args.nb_classes=62
args.eval_path=./fmow-rgb/val
args.eval_dataset=fmow
args.batch_size=16
and as mentioned in paper
args.blr=5e-3
args.weight_decay=5e-3
But the I am not able to reproduce the results mentioned in paper. Instead I get maximum accuracy of 62.
Using the blr=5e-3 gives NaN loss issue, so I reduced its value until it runs fine.
Surprisingly, train loss decreases and test loss increases. after 2nd epoch of finetuning accuracy started decreasing from 62 to 58.
Could you please tell me the exact values of hyperparameters and settings? And what could be the reason of not reproducing the finetuning results?
Hi,
Thanks for sharing the interesting work.
I tried to finetune the model on FMOW-RGB dataset by using the pretrained weights of 800th epoch provided at the repository. I used the default hyperparameters in the file main_linprobe.py and set
args.finetune=True
args.nb_classes=62
args.eval_path=./fmow-rgb/val
args.eval_dataset=fmow
args.batch_size=16
and as mentioned in paper
args.blr=5e-3
args.weight_decay=5e-3
But the I am not able to reproduce the results mentioned in paper. Instead I get maximum accuracy of 62.
Using the blr=5e-3 gives NaN loss issue, so I reduced its value until it runs fine.
Surprisingly, train loss decreases and test loss increases. after 2nd epoch of finetuning accuracy started decreasing from 62 to 58.
Could you please tell me the exact values of hyperparameters and settings? And what could be the reason of not reproducing the finetuning results?
Here is param settings from log file
Namespace(accum_iter=1,
base_resolution=2.5,
batch_size=16,
blr=7e-05,
checkpoint_interval=10,
checkpoint_path='scalemae-vitlarge-800.pth',
config='./config/fmow.yaml',
device='cuda',
dist_backend='nccl',
dist_eval=False,
dist_on_itp=False,
dist_url='env://',
distributed=True,
drop_path=0.2,
epochs=50,
eval=False,
eval_base_resolution=1.0,
eval_dataset='fmow',
eval_gsd=False,
eval_only=False,
eval_path='./fmow-rgb/val',
eval_reference_resolution=224,
eval_scale=224,
finetune=True,
global_pool=False,
gpu=0,
input_size=224,
layer_decay=0.75,
linear_layer_scale=1.0,
local_rank=0,
log_dir='./finetune_dir',
lr=None,
mask_ratio=0.75,
min_lr=0.0,
model='vit_large_patch16',
name='',
nb_classes=62,
no_autoresume=False,
norm_pix_loss=False,
num_workers=10,
output_dir='./finetune_dir',
pin_mem=True,
print_freq=20,
rank=0,
restart=False,
resume='',
scale_max=1.0,
scale_min=0.5,
seed=0,
source_size=[224],
start_epoch=0,
target_size=[224],
wandb_id=None,
warmup_epochs=0,
weight_decay=0.005,
world_size=8)
The text was updated successfully, but these errors were encountered: