Input Shape for Training and Inference with Pretrained Models

[hamac03]

Hi,

I'm currently training models using a backbone like ResNet50, which is pretrained on ImageNet with an input shape of 3x224x224. I was wondering if it's possible to use a larger input size, such as 3x320x320 or 3x416x416, and still benefit from the pretrained weights, or does the input shape need to be strictly 224x224?

Are there any constraints or considerations when using larger input sizes with pretrained models?

Best regards,
Ha

[qubvel]

Hey @hamac03, yes, you can use any size for convolutional backbones, e.g. 512x512. Usually, the only requirement is for the size to be divisible by 32. You can even train on 512x512 input and then inference on 1024x1024 and have similar or even better performance. But it's always better to check the metrics on your particular model and dataset.

[ogencoglu]

@hamac03 just to clarify: the requirement of "divisibly by 32" depends on the number of encoder-decoder stages. 5 stages of downsampling gives us 2^5=32 but it depends on the number of stages in your architecture.