Specific information to recreate leaderboard results?

[megvii-basedetection]

Specific information to recreate leaderboard results?

[fcjian]

@yinchimaoliang Do you have any plan to release the config file and model shown on the detection leaderboard?

[yinchimaoliang]

Hi! The input resolution is 640 x 1600, the backbone we use is ConvNext, the bev resolution is 256 x 256, we also use train and val split for training.

[bluffish]

Hi, thanks for the info! I'll make sure to give it a try!

[bluffish]

Hello, can you tell me which ConvNeXt architecture you guys use? Tiny, small base or large? Thanks!

[zehuichen123]

Hi, @yinchimaoliang, I wonder if the data augmentation in your submission is the same one used in the r50 config? So the only differences are (I) image resolution (2) bev resolution (3) image backbone (4) trainval for test, right?

[yinchimaoliang]

Hi! When the resolution of the input image is changed, the resize_lim and final_dim of ida should be changed accordingly.

[swtju14]

Hi, @bluffish. BEVDepth-pure extends the BEVDepth by using ConvNeXt-base as backbone.

[zehuichen123]

Hi, @yinchimaoliang thanks for your information. I am trying to use ConvNeXt but encountered OOM problem. I wonder if you adopt torch.utils.checkpoint to save your memory in the ConvNeXt backbone? If so, do you use it in the stage conv or also in the downsample layers ?Thanks!

[yinchimaoliang]

Hi, @zehuichen123 , we didn't adopt torch.utils.checkpoint, we used V100 with a memory of 32GB, you can try it if you are using machines with smaller memory size.

[zehuichen123]

Thanks! I am using V100 machine too. BTW, how did you set your batch size for each GPU as well as the learning rate? (since a batch size of 8 is impossible in this case)

[yinchimaoliang]

Thanks! I am using V100 machine too. BTW, how did you set your batch size for each GPU as well as the learning rate? (since a batch size of 8 is impossible in this case)

The batch size for each card is set to 2, and we use 4 machines. Learning rate is 2e-4.

[zehuichen123]

Got it! Thanks! 😊

[fcjian]

Hi, @zehuichen123, we also encountered OOM problem. Did you solve this problem？

[zw615]

Thanks! I am using V100 machine too. BTW, how did you set your batch size for each GPU as well as the learning rate? (since a batch size of 8 is impossible in this case)

The batch size for each card is set to 2, and we use 4 machines. Learning rate is 2e-4.

@yinchimaoliang Did you use base lr = 2e-4 or final lr = 2e-4? I see in the code the basic learning rate for 64 samples per batch is set to 2e-4. I wonder if it is the same for 2 (batch size) * 4 (machine) * 8(gpu per machine)?

[zw615]

Also, I see from here that you uses dcn head in the submitted version to the test set leaderboard, which is not mentioned in this issue. I wonder if there is anything else we should pay attention to to reproduce the leaderboard results (other than (I) image resolution (2) bev resolution (3) image backbone (4) use both train/val set for training)?

Hi! When the resolution of the input image is changed, the resize_lim and final_dim of ida should be changed accordingly.

For instance, what is the right value of resize_lim and final_dim of ida here?

[zw615]

Also, I have noticed ConvNeXT uses different drop_path_rate for different architectures. What value did you use for drop_path_rate?

[pummi823]

Why does the author not answer these questions （by zeyuwang615）？？ They are quite important T T