Finetuning using input image whose size is smaller than images used to pretrain ResNet50

[luigi (Luigi Selmi)]

Hi all,
I am using the pretrained ResNet50 architecture available in fastai to finetune a set of satellite images whose size is 64x64 pixels. Since ResNet50 expects images whose size is 224x224 pixels I would like to know how fastai uses the smaller images to update the model. My data block is

blocks = DataBlock(blocks = (ImageBlock, CategoryBlock),
get_items=get_image_files,
splitter=RandomSplitter(seed=42),
get_y=using_attr(RegexLabeller(r’(.+)_\d+.jpg$’), ‘name’))

I don’t use any resizing to match the ResNet 224x224 pixel size. My questions are:

1. Can the fine-tuning work properly without any explicit resize ?
2. How fastai matches the size of my images (64x64) to that of the images used to pretrain ResNet50 (224x224) ?

Thanks in advance for any suggestion.
Luigi

[rossheaton (Ross)]

Hi Luigi, Jonathan’s post answers your question:

Feeding different sized images to fine tune ResNet34

It took me awhile to understand how a pre-trained model could use a different image size than the one it was trained on. I wrote up my notes on using resnet18 for MNIST classification. Maybe they will be helpful to someone. How does the pre-trained model use our image size? The PyTorch resnet18 model was trained on ImageNet images. These images are 3-channel 224x224 pixels. Our images are 3-channel 28x28. How can we use pre-trained weights for a model trained on different images sizes? Let’s lo…

[luigi (Luigi Selmi)]

Hi Ross,
thank you very much for your quick answer. Jonathan’s post makes perfect sense. I compared the same architecture (ResNet50) with a different set of satellite images for finetuning with a shape of 256x256 pixels and the number of parameters doesn’t change at each layer and of course the total is always the same.