Training scheme of the pretrained imagenet models?

[tzm1003306213]

Hi,

Are the pretrained models reported by torchvision using the same hyper-parameters as https://github.com/pytorch/examples/blob/master/imagenet/main.py? I used the default hyper-parameters to train mobilenet_v2, but the results were much worse than reported.

Thanks

[pmeier]

You can find the training scripts for all torchvision models here. For classification this is the script you want

https://github.com/pytorch/vision/blob/master/references/classification/train.py

[fmassa]

As @pmeier mentioned, we also provide the default hyperparameters for the pre-trained models in torchvision under the references folder.
For mobilenet_v2 we used https://github.com/pytorch/vision/tree/master/references/classification#mobilenetv2

python -m torch.distributed.launch --nproc_per_node=8 --use_env train.py\
     --model mobilenet_v2 --epochs 300 --lr 0.045 --wd 0.00004\
     --lr-step-size 1 --lr-gamma 0.98

Let us know if you have further questions

[magamba]

Hello,

I have a similar issue. I am using pretrained AlexNet and VGG models from torchvision for a scientific paper and, in order to interpret my results, I would like to know how the models were trained. I have checked here as suggested, but I am unable to find any reference to VGG19, its shallower variants, and AlexNet. Are they published anywhere else?

Thank you

[pmeier]

I expect these models were trained with the default parameters given in train.py, but I can't be sure. @fmassa ?

[fmassa]

AlexNet and VGG have been trained a long time ago by @colesbury , I think they might follow the same procedure as ResNet (and thus default parameters), but I'm not 100% sure. Original PR adding those is inn #23

[colesbury]

Models with batch normalization were trained with the default parameters. Models without batch normalization were trained with an initial learning rate of 0.01 (i.e. 1/10th the default learning rate).

See https://github.com/pytorch/examples/tree/master/imagenet#training

[pmeier]

Should we add this in the classification reference README? If yes, I could send a PR tomorrow.

[fmassa]

@pmeier yes please, if you could send a PR improving the README it would be great