Question about training for model `ofa_D4_E6_K7`

[gunjupark]

Hello I want to train OFA model(ofa_mbv3) on 'Cifar100' or custom datasets.

so I want to get some training details about first supernet.

When I checked model in progressive-shrinking phase, I saw F.Linear(Kernel transformer) layer's weights were also trained.

When I want to train First Supernet (ofa_D4_E6_K7), should I train there Kernel transform matrix?

And I was wondering If you had some information about OFA net training on other dataset(like cifar 10, 100), I want to know them.

Thank you all the time.

[gunjupark]

on Cifar 10 or Cifar 100 I could train OFA's Init Supernet using paper's setting (Init_LR = 2.6) / 32 = 0.08125 per GPUs Supernet was trained quite well.

[darrenzhang1007]

on Cifar 10 or Cifar 100 I could train OFA's Init Supernet using paper's setting (Init_LR = 2.6) / 32 = 0.08125 per GPUs Supernet was trained quite well.

Hello，I want to ask you two questions:

1. How did you modify the code to be used for the classification of the cifar100 dataset；
2. After you modify the code, the cifar10 dataset has 100 categories, but the official pre-training model is 1000 classifications for ImageNet Dataset, and the loading pre-trained model will fail to load. Data dimensions do not match. How did you solve it?

Thanks a lot , Looking forward to your reply

[gunjupark]

on Cifar 10 or Cifar 100 I could train OFA's Init Supernet using paper's setting (Init_LR = 2.6) / 32 = 0.08125 per GPUs Supernet was trained quite well.

Hello，I want to ask you two questions:

1. How did you modify the code to be used for the classification of the cifar100 dataset；

I use same class (OFAMobileNetV3) in ofa's codes to train supernet. (e.g. progressive_shrinking.py)

Instead, I preprocessed Cifar's Images to fit Imagenet's Images size (e.g 40 -> 224 random resized crop)

2. After you modify the code, the cifar10 dataset has 100 categories, but the official pre-training model is 1000 classifications for ImageNet Dataset, and the loading pre-trained model will fail to load. Data dimensions do not match. How did you solve it?

Thanks a lot , Looking forward to your reply

I trained Supernet using cifar10 from scratch (I didnt use pretrained model for Imagenet) so I modify some codes in ofa/imagenet_codebase/data_provider & run_manager to set n_classes = 10 or 100

My step is 1. Training Supernet -> 2. Finetuing for all subnets (using Progressive Shirnking) -> 3. run EA codes in tutorial

[Jon-drugstore]

on Cifar 10 or Cifar 100 I could train OFA's Init Supernet using paper's setting (Init_LR = 2.6) / 32 = 0.08125 per GPUs Supernet was trained quite well.

how to change the code for training the supernet? using the train_ofa_net.py？Thanks