haoqing-wang / localmim Goto Github PK

I cannot reproduce ADE20k performance (49.5 mIoU) shown in Table 2 of the paper.

Using your released checkpoint and semantic segmentation code, I got 48.65 mIoU (see the log in onedrive).

Any suggestions for the reproduction?

Where is the paper link? Cannot find it.

I guess there is a minor typo in the pretrain code:

This should be rather "MIM_vit_large_patch16" I guess?

I want to be sure about the learning rates you have used for fine-tuning.
In the paper you state:

And, in this repo, it is written that for 100 epoch pre-trained ViT-B you use 4e^-3 learning rate (but use 2*e^-3 for 1600 epoch pre-trained model):

I want to be sure because your paper says that the MAE fine-tune setting was mostly used, but MAE only uses the learning rate of 4*e^-3 (base learning rate of e^-3 multiplied by total_batchsize/256.).

Also, when is the 3*e^-3 learning rate used for fine-tuning?

there is the pre training code
OMP_NUM_THREADS=1 python -m torch.distributed.launch --nproc_per_node=8 run_pretrain.py --output_dir /output_dir/ --batch_size 256 --model mim_swin_base_patch4_win7 --target HOG --hog_nbins 18 --hog_bias --mask_ratio 0.75 --epochs 400 --warmup_epochs 40 --blr 1e-4 --weight_decay 0.05 --data_path /path/to/imagenet/

but I can't see where is the setup

Hello, how to convert the output of mim_swin_base_patch4_win7 to an image