Vision: ImageNet pre-trained model code and hyper-parameters
Hi,
Is the code used to trained the torchvision models (especially Resnet) on ImageNet available ? What are the hyper-parameters used ? Did you use specific methods (dropout, weight decay, specific augmentation such as cutout...etc) during training ?
Thank you very much
Jobanan
>All comments
Hi @Jobanan, you can find the training scripts for all pre-trained models here:
https://github.com/pytorch/vision/tree/master/references
The hyper parameters are displaying in the README, but for everything else you have to dig into the actual training script. For example, we used the following image transformation during training:
Let me know if you have further questions.
pmeier
on 14 Jul 2020
👍2
Was this page helpful?
0 / 5 - 0 ratings
Related issues
fmassa
路
30Comments
ppwwyyxx
路
33Comments
sumanthratna
路
28Comments
Jonas1312
路
23Comments
zsef123
路
23Comments
Most helpful comment
Hi @Jobanan, you can find the training scripts for all pre-trained models here:
https://github.com/pytorch/vision/tree/master/references
The hyper parameters are displaying in the README, but for everything else you have to dig into the actual training script. For example, we used the following image transformation during training:
https://github.com/pytorch/vision/blob/0344603ece20ba5e4086aa8c49b2cce0829aefe6/references/classification/train.py#L96-L103
Let me know if you have further questions.