Fine-Tuning

https://d2l.ai/chapter_computer-vision/fine-tuning.html

I find hotdog is 934 in ImageNet

in Exercises4:
in my opinion, we can copy the hotdog_w into the parameters of fc.
image

when i use xavier_uniform_ only, the training process is
image

when i copy hotdog_w into fc in the begining of training, the loss is smaller than before.
image

4 Likes

Hi @min_xu, I was wondering why we update finetune_net.fc.weight.data[0 : , ] to hotdog_w[0] and not finetune_net.fc.weight.data[1 : , ] to hotdog_w[0]. Shouldn’t the 0th index correspond to label 0 and index 1 correspond to label 1 (which indicates that the image contains a hotdog)

How would we go about implementing fine-tuning for a pretrained model that doesn’t support .fc?
I’ve been using alexnet and that doesn’t support .fc

@Danfriedz Try .classifier instead of .fc if you are fine-tuning torchvision.models.alexnet().

im trying to implement vgg16 to finetune, i have changed the classifer[0] for the linear equation. however i keep getting runtime errors saying CUDA is out of memory, is there any other values to change besides making the batch size smaller