input_size vs test_input_size

[talhaanwarch]

What is difference between input_size and test_input_size in efficientnetv2 and if we have to pass data to the network, what input size should our data have?
For example in efficientnetv2_s, what should be my input 288 or 384
efficientnetv2_s': _cfg(url='',input_size=(3, 288, 288), test_input_size=(3, 384, 384), pool_size=(9, 9), crop_pct=1.0),

[rwightman]

@talhaanwarch If test_input_size is present, it is used for validation, ie it is the preferred 'test time' or inference resolution, input_size is the default train resolution (or both if test_input_size) is not defined. As per https://arxiv.org/abs/1906.06423, models trained with heavier augmentations such as RandAugment or AutoAugment often test better at a larger scale, this works very well with fully convolutional (up to the last global pool) models. It's becoming more common to see models evaluated that way, see EfficientNetV2 paper https://arxiv.org/abs/2104.00298 and NFNet papers as other examples https://arxiv.org/abs/2102.06171

[Doraemonzm]

hello, Ross.
I noticed that in "ResNet strikes back: An improved training procedure in timm", different resolutions are used for training and testing in training procedure 3 (160 for training resolution, 224 for test resolution). But I don't find a parse in timm to set different resolutions. I would like to know how can I train a model from scratch with different train-validation resolutions by your code. Will it help to improve the accuracy of the image classification model, thanks!

[rwightman]

@Doraemonzm easiest thing to do right now is set the command line image size for your train size and override the validation size manually in the code, I have a TODO to add an arg for validation so they can be different from cmd line.

That said, it hasn't been a super high priority. It's usually not that important to actually validate on the higher res during training, there is strong correlation between the val results at train res and higher validation res, so it's fine to just do the training at lower res and then run validation on best checkpoints after at a higher res (or test multiple res)