swin的预训练权重

[LiBingyu01]

请问您，swin的预训练权重链接是什么？我在官网上下载的不能匹配进去。、

[peter-wang321]

https://github.com/microsoft/Swin-Transformer#main-results-on-imagenet-with-pretrained-models

I used this, it works.

[peter-wang321]

Sorry, I also encountered issues like:

unexpected key in source state_dict: norm.module.weight, module.norm.weight, norm.weight, norm.module.bias, module.norm.bias, norm.bias, head.module.weight, module.head.weight, head.weight, head.module.bias, module.head.bias, head.bias, patch_embed.module.proj.weight, patch_embed.module.proj.bias, patch_embed.module.norm.weight, patch_embed.module.norm.bias, patch_embed.proj.weight, patch_embed.proj.bias, patch_embed.norm.module.weight, patch_embed.norm.weight, patch_embed.norm.module.bias, patch_embed.norm.bias, layers.0.blocks.module.1.attn_mask, layers.0.blocks.0.module.norm1.weight, layers.0.blocks.0.module.norm1.bias, layers.0.blocks.0.module.norm2.weight, layers.0.blocks.0.module.norm2.bias, layers.0.blocks.0.module.attn.relative_position_index, layers.0.blocks.0.module.attn.relative_position_bias_table, layers.0.blocks.0.norm1.weight, layers.0.blocks.0.norm1.bias, layers.0.blocks.0.attn.qkv.module.weight, layers.0.blocks.0.attn.qkv.weight, layers.0.blocks.0.attn.qkv.module.bias, layers.0.blocks.0.attn.qkv.bias, layers.0.blocks.0.attn.proj.module.weight, layers.0.blocks.0.attn.proj.weight,

[peter-wang321]

请问您，swin的预训练权重链接是什么？我在官网上下载的不能匹配进去。、

Have you reproduced the results based on mit_b3 or mit_b5?

[JiaDingCN]

Sorry, I also encountered issues like:

unexpected key in source state_dict: norm.module.weight, module.norm.weight, norm.weight, norm.module.bias, module.norm.bias, norm.bias, head.module.weight, module.head.weight, head.weight, head.module.bias, module.head.bias, head.bias, patch_embed.module.proj.weight, patch_embed.module.proj.bias, patch_embed.module.norm.weight, patch_embed.module.norm.bias, patch_embed.proj.weight, patch_embed.proj.bias, patch_embed.norm.module.weight, patch_embed.norm.weight, patch_embed.norm.module.bias, patch_embed.norm.bias, layers.0.blocks.module.1.attn_mask, layers.0.blocks.0.module.norm1.weight, layers.0.blocks.0.module.norm1.bias, layers.0.blocks.0.module.norm2.weight, layers.0.blocks.0.module.norm2.bias, layers.0.blocks.0.module.attn.relative_position_index, layers.0.blocks.0.module.attn.relative_position_bias_table, layers.0.blocks.0.norm1.weight, layers.0.blocks.0.norm1.bias, layers.0.blocks.0.attn.qkv.module.weight, layers.0.blocks.0.attn.qkv.weight, layers.0.blocks.0.attn.qkv.module.bias, layers.0.blocks.0.attn.qkv.bias, layers.0.blocks.0.attn.proj.module.weight, layers.0.blocks.0.attn.proj.weight,

Have you tested the released swin transformer models? Do they have the correct accuracy?

[JiaDingCN]

It looks like you have modified the model setting code or used the wrong config. You may check whether the code is changed or the running command is right.

[peter-wang321]

I have tested all the pre-trained models on the nyuv2 dataset, the swin-large-384 +FineTune from SUN 300eps did not show the correct accuracy, which is just 56.76 instead of 60.9. However, other models have a similar accuracy as the one reported in the Repo.

The running command I used is: CUDA_VISIBLE_DEVICES=0,1,2 python -m torch.distributed.launch --nproc_per_node=3 --use_env main.py --backbone swin_large_window12 --dataset nyudv2 -c rerun_54.8_swin_large_window12_finetune_dpr0.15_100+200+100 \ --dpr 0.15 --num-epoch 100 200 100 --is_pretrain_finetune --resume ./swin-large-384.pth.tar --eval --resume ./pretrained/finetune-swin-large-384.pth.tar

Another quick question, I am not sure what the "54.8" means in the "-c" para.

Thanks.

[JiaDingCN]

I have tested all the pre-trained models on the nyuv2 dataset, the swin-large-384 +FineTune from SUN 300eps did not show the correct accuracy, which is just 56.76 instead of 60.9. However, other models have a similar accuracy as the one reported in the Repo.

The running command I used is: CUDA_VISIBLE_DEVICES=0,1,2 python -m torch.distributed.launch --nproc_per_node=3 --use_env main.py --backbone swin_large_window12 --dataset nyudv2 -c rerun_54.8_swin_large_window12_finetune_dpr0.15_100+200+100 --dpr 0.15 --num-epoch [100 200 100](tel:100 200 100) --is_pretrain_finetune --resume ./swin-large-384.pth.tar --eval --resume ./pretrained/finetune-swin-large-384.pth.tar

Another quick question, I am not sure what the "54.8" means in the "-c" para.

Thanks.

You may delete the first —resume. Also, the typo in -c is fixed.

[JiaDingCN]

Closed due to inactivity.

[iacopo97]

@peter-wang321 Have you managed to obtain the results?

[peter-wang321]

yes except the last model.

On Thu, Aug 29, 2024 at 3:49 AM iacopocurti @.***> wrote:

@peter-wang321 https://github.com/peter-wang321 Have you managed to obtain the results?

— Reply to this email directly, view it on GitHub https://github.com/JiaDingCN/GeminiFusion/issues/7#issuecomment-2316936806, or unsubscribe https://github.com/notifications/unsubscribe-auth/AN6CL5V45TXJGYTAHBCOU3TZT3HADAVCNFSM6AAAAABMT7F36OVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDGMJWHEZTMOBQGY . You are receiving this because you were mentioned.Message ID: @.***>