transforms question from lecture No. 214

[geoffm39]

In lecture 214, we created some manual transforms for use for both efficientnet_b0 and efficientnet_b2 and their best available weights.

With the updated torchvision version I noticed that the transforms() from the DEFAULT weights used for the different models has a slightly different resize values. I'm not sure if these updated DEFAULT weights since the time of recording the lecture.

My question is as we are using the same transforms for the testing of both pretrained models, will this effect the efficientnet_b2 results?
would creating separate dataloaders for the two pretrained models with different transforms by calling transforms() on the two different DEFAULT weights be the best approach?

Thanks!
Geoff

[37Lime]

Hi Geoff,

The transforms() method provides you with the transforms that were used when the model was trained and obtained its weights. Because you are using those weights (I assume) for the prediction/test, it would be a good practice to also use the same transforms.

You can get more information here.

Best wishes.

[geoffm39]

Ok great, that's what I assumed.

Thanks for clarifying for me.