Hi,
I am doing experiments with EfficientNet-B1 as backbone. Currently I have trained the model with input size of 240 × 240.
Now I want to try training with 512 × 512, can you please help me understand what will be effect of it on the model ?
You might want to check out popular repositories of models e.g., huggingface/pytorch-image-models: PyTorch image models, scripts, pretrained weights – ResNet, ResNeXT, EfficientNet, EfficientNetV2, NFNet, Vision Transformer, MixNet, MobileNet-V3/V2, RegNet, DPN, CSPNet, and more (github.com) which would have example results on how models perform when scaled up in terms of resolution. It would ultimately depend on your specific task, but generally the model should have somewhat higher accuracy (provided that the test-time resolution is also scaled up appropriately [1906.06423] Fixing the train-test resolution discrepancy (arxiv.org)), at roughly (512/240)^2 times the original compute cost.