About DB-Swin-B model, I tried to reimplementate it and get 55.2 mAP

[JR-Wang]

I re-trained DB-Swin-B model on coco dataset. The config file I use is configs/cbnet/htc_cbv2_swin_base_patch4_window7_mstrain_400-1400_giou_4conv1f_adamw_20e_coco.py, and the pretrained model I use is swin-base model trained on 22k imagenet, 384 input size. But the mAP I finally get is 55.2, lower than the result you provided in your readme.md (58.4).
What's wrong with my reimplementation?

[chuxiaojie]

We use checkpoint named swin_base_patch4_window7_224_22k.pth from https://github.com/SwinTransformer/storage/releases/download/v1.0.0/swin_base_patch4_window7_224_22k.pth

[JR-Wang]

ok, I will try this pretrained model. Many thanks

[cailk]

ok, I will try this pretrained model. Many thanks

Hi, have you tried the new pretrained model to reproduce Swin-B results? Does it work?

[JR-Wang]

ok, I will try this pretrained model. Many thanks

Hi, have you tried the new pretrained model to reproduce Swin-B results? Does it work?

I don't have enough V100 now, so I haven't finish my training. I saw your question "Reproduce DB-Swin-Large models #47". So have you tried to train a swin-L model? I found that when I use V100 to train swin-B model, the memory used is almost 100%. Obviously it will OOM to train swin-L model.

[cailk]

ok, I will try this pretrained model. Many thanks

Hi, have you tried the new pretrained model to reproduce Swin-B results? Does it work?

I don't have enough V100 now, so I haven't finish my training. I saw your question "Reproduce DB-Swin-Large models #47". So have you tried to train a swin-L model? I found that when I use V100 to train swin-B model, the memory used is almost 100%. Obviously it will OOM to train swin-L model.

Haven't yet. Since I don't have enough GPUs either, I need to make sure which config will work to reproduce 59.6AP on COCO-val.

BTW, you are now training a DB-Swin-B model, right? Could you let me know the results as soon as you have finished the training? Big thanks!~

[JR-Wang]

We use checkpoint named swin_base_patch4_window7_224_22k.pth from https://github.com/SwinTransformer/storage/releases/download/v1.0.0/swin_base_patch4_window7_224_22k.pth

Hello, I changed the pre-trained model. This time I got mAP 55.8, only 0.6 higher than last time. May I ask how many GPUs you used when you train your model? And should learning rate keep proportion to GPU number?
Another question is, I found that when I use V100 to train swin-B model, the memory used is almost 100%. Obviously it will OOM to train swin-L model. So how do you train your swin-L model? I have already used fp16 mode to reduce computation.

[JR-Wang]

ok, I will try this pretrained model. Many thanks

Hi, have you tried the new pretrained model to reproduce Swin-B results? Does it work?

I don't have enough V100 now, so I haven't finish my training. I saw your question "Reproduce DB-Swin-Large models #47". So have you tried to train a swin-L model? I found that when I use V100 to train swin-B model, the memory used is almost 100%. Obviously it will OOM to train swin-L model.

Haven't yet. Since I don't have enough GPUs either, I need to make sure which config will work to reproduce 59.6AP on COCO-val.

BTW, you are now training a DB-Swin-B model, right? Could you let me know the results as soon as you have finished the training? Big thanks!~

I finished training process to get 55.8 mAP, only 0.6 higher than last time. I wonder if I should decay learning rate since I don't have enough GPUs.

[cailk]

ok, I will try this pretrained model. Many thanks

Hi, have you tried the new pretrained model to reproduce Swin-B results? Does it work?

I don't have enough V100 now, so I haven't finish my training. I saw your question "Reproduce DB-Swin-Large models #47". So have you tried to train a swin-L model? I found that when I use V100 to train swin-B model, the memory used is almost 100%. Obviously it will OOM to train swin-L model.

Haven't yet. Since I don't have enough GPUs either, I need to make sure which config will work to reproduce 59.6AP on COCO-val.
BTW, you are now training a DB-Swin-B model, right? Could you let me know the results as soon as you have finished the training? Big thanks!~

I finished training process to get 55.8 mAP, only 0.6 higher than last time. I wonder if I should decay learning rate since I don't have enough GPUs.

I have finished the training of Swin-Large, which could achieve 58.8 mAP. This is pretty close to the result reported in the paper (59.1 mAP). And I use 8 V100 GPUs with batchsize=1. I also have trained a CBNet-Swin-base. However, because of my stupidity, I loaded a wrong Swin-Base pre-train checkpoint. I guess the results of CBNet-Swin-Base are also reproducible if you keep every hyper-parameters unchanging.

[JR-Wang]

ok, I will try this pretrained model. Many thanks

Hi, have you tried the new pretrained model to reproduce Swin-B results? Does it work?

I don't have enough V100 now, so I haven't finish my training. I saw your question "Reproduce DB-Swin-Large models #47". So have you tried to train a swin-L model? I found that when I use V100 to train swin-B model, the memory used is almost 100%. Obviously it will OOM to train swin-L model.

Haven't yet. Since I don't have enough GPUs either, I need to make sure which config will work to reproduce 59.6AP on COCO-val.
BTW, you are now training a DB-Swin-B model, right? Could you let me know the results as soon as you have finished the training? Big thanks!~

I finished training process to get 55.8 mAP, only 0.6 higher than last time. I wonder if I should decay learning rate since I don't have enough GPUs.

I have finished the training of Swin-Large, which could achieve 58.8 mAP. This is pretty close to the result reported in the paper (59.1 mAP). And I use 8 V100 GPUs with batchsize=1. I also have trained a CBNet-Swin-base. However, because of my stupidity, I loaded a wrong Swin-Base pre-train checkpoint. I guess the results of CBNet-Swin-Base are also reproducible if you keep every hyper-parameters unchanging.

Did you meet OOM when you train Swin-L? I alse tried to train Swin-L with each gpu one sample and got OOM on V100. Could you share your config file of Swin-L? Many thanks.

[cailk]

ok, I will try this pretrained model. Many thanks

Hi, have you tried the new pretrained model to reproduce Swin-B results? Does it work?

I don't have enough V100 now, so I haven't finish my training. I saw your question "Reproduce DB-Swin-Large models #47". So have you tried to train a swin-L model? I found that when I use V100 to train swin-B model, the memory used is almost 100%. Obviously it will OOM to train swin-L model.

Haven't yet. Since I don't have enough GPUs either, I need to make sure which config will work to reproduce 59.6AP on COCO-val.
BTW, you are now training a DB-Swin-B model, right? Could you let me know the results as soon as you have finished the training? Big thanks!~

I finished training process to get 55.8 mAP, only 0.6 higher than last time. I wonder if I should decay learning rate since I don't have enough GPUs.

I have finished the training of Swin-Large, which could achieve 58.8 mAP. This is pretty close to the result reported in the paper (59.1 mAP). And I use 8 V100 GPUs with batchsize=1. I also have trained a CBNet-Swin-base. However, because of my stupidity, I loaded a wrong Swin-Base pre-train checkpoint. I guess the results of CBNet-Swin-Base are also reproducible if you keep every hyper-parameters unchanging.

Did you meet OOM when you train Swin-L? I alse tried to train Swin-L with each gpu one sample and got OOM on V100. Could you share your config file of Swin-L? Many thanks.

I just use the original Swin-Large config file in this repo swin-large, and only set use_checkpoint=True. I use an 8*32g V100 machine with batchsize=1.

[JR-Wang]

ok, I will try this pretrained model. Many thanks

Hi, have you tried the new pretrained model to reproduce Swin-B results? Does it work?

I don't have enough V100 now, so I haven't finish my training. I saw your question "Reproduce DB-Swin-Large models #47". So have you tried to train a swin-L model? I found that when I use V100 to train swin-B model, the memory used is almost 100%. Obviously it will OOM to train swin-L model.

Haven't yet. Since I don't have enough GPUs either, I need to make sure which config will work to reproduce 59.6AP on COCO-val.
BTW, you are now training a DB-Swin-B model, right? Could you let me know the results as soon as you have finished the training? Big thanks!~

I finished training process to get 55.8 mAP, only 0.6 higher than last time. I wonder if I should decay learning rate since I don't have enough GPUs.

I have finished the training of Swin-Large, which could achieve 58.8 mAP. This is pretty close to the result reported in the paper (59.1 mAP). And I use 8 V100 GPUs with batchsize=1. I also have trained a CBNet-Swin-base. However, because of my stupidity, I loaded a wrong Swin-Base pre-train checkpoint. I guess the results of CBNet-Swin-Base are also reproducible if you keep every hyper-parameters unchanging.

Did you meet OOM when you train Swin-L? I alse tried to train Swin-L with each gpu one sample and got OOM on V100. Could you share your config file of Swin-L? Many thanks.

I just use the original Swin-Large config file in this repo swin-large, and only set use_checkpoint=True. I use an 8*32g V100 machine with batchsize=1.

I also use 32g V100. Maybe I should check whether my fp16 setting is right. Thank you for your reply.

[foolhard]

I re-trained DB-Swin-B model on coco dataset. The config file I use is configs/cbnet/htc_cbv2_swin_base_patch4_window7_mstrain_400-1400_giou_4conv1f_adamw_20e_coco.py, and the pretrained model I use is swin-base model trained on 22k imagenet, 384 input size. But the mAP I finally get is 55.2, lower than the result you provided in your readme.md (58.4). What's wrong with my reimplementation?

How to use Swin-Base model pretrained on 22k Imagenet in CBNetV2? How do you modify the config file?

[cailk]

I re-trained DB-Swin-B model on coco dataset. The config file I use is configs/cbnet/htc_cbv2_swin_base_patch4_window7_mstrain_400-1400_giou_4conv1f_adamw_20e_coco.py, and the pretrained model I use is swin-base model trained on 22k imagenet, 384 input size. But the mAP I finally get is 55.2, lower than the result you provided in your readme.md (58.4). What's wrong with my reimplementation?

How to use Swin-Base model pretrained on 22k Imagenet in CBNetV2? How do you modify the config file?

Hi, sorry for the late reply. I have checked my implementation, and it looks like the pre-training checkpoint should be swin_base_patch4_window7_224_22k. BTW, you can contact me by email for the config file~