search

WongKinYiu / ScaledYOLOv4

Scaled-YOLOv4: Scaling Cross Stage Partial Network
GNU General Public License v3.0
2.02k stars 570 forks source link

Support PyTorch 1.9 Mish version #271

Open digantamisra98 opened 3 years ago

digantamisra98 commented 3 years ago

With the new PyTorch release, for consistency, we can switch to the Mish version supported in PyTorch by default. Link - https://pytorch.org/docs/stable/generated/torch.nn.Mish.html#torch.nn.Mish

WongKinYiu commented 3 years ago

cool, thanks for the information. but the server i rent not yet support pytorch 1.9, will try it on my pc first. by the way, some of my code support w&b now.

jaideep11061982 commented 3 years ago

@WongKinYiu how can we add custom Augmentation like Albumentations ones

WongKinYiu commented 3 years ago

add augmentation code to: https://github.com/WongKinYiu/ScaledYOLOv4/blob/9336ffe48abe0df874844118d6fa47fea25b0bf5/utils/datasets.py#L594-L596

digantamisra98 commented 3 years ago

@WongKinYiu Oh that's pretty cool, do you have a supporting W&B dashboard?

WongKinYiu commented 3 years ago

I am not sure, this branch and this repo now support W&B. And have plan to update this branch and this repo to support W&B too.

digantamisra98 commented 3 years ago

@WongKinYiu Additionally, I wanted to bring it to your notice that considering Mish is an integral component in YOLOv4, it would be appropriate if it can be cited in the CVPR Scaled YOLOv4 paper to adhere to academic ethos and morals. Feel free to let me know if you have differing thoughts on the same.

WongKinYiu commented 3 years ago

I update pytorch to the latest docker image, but do not know why it has no nn.Mish(). https://docs.nvidia.com/deeplearning/frameworks/pytorch-release-notes/rel_21-06.html#rel_21-06

digantamisra98 commented 3 years ago

@WongKinYiu Oh, I am not sure, I checked the dist from the NVIDIA doc but doesn't seem that it includes all the features released in PyTorch 1.9.