How to create model with custom input_size

[kondounagi]

Hello. Thank you @rwightman for this great repository !

I'm wondering whether timm.create_model() can specify custom input_size.
For example, is it possible for mobilenetv3_large_075 to set (3, 800, 800) as input_size, which is different from original (3, 224, 224) in default_cfg ?
I found no documentation about this custom input_size.
If you could tell me how to do that, I'm please to write documentation as a contribution to this repository !

[rwightman]

@kondounagi you don't create the model with an input_size, the models are fully convolutional and img size agnostic with the exception of vision_transformer which is fixed to a certain size. The default_cfg attached to the model is just metadata to indicate what resolution the datapipeline should be setup with to match the pretrained weights.

In the example train.py script, the input size fed to the model is determined by the data loader setup here: https://github.com/rwightman/pytorch-image-models/blob/master/train.py#L477

Earlier on, the value of that input_size is resolved between the default_args in the model and the command line input args here:
https://github.com/rwightman/pytorch-image-models/blob/master/train.py#L337

The arg --img-size gets converted from a single value to (3, img_size, img_size). The script command line args (--img-size 800) take priority over default_cfg, you can also override manually if you want and just pass a hard-coded tuple to the loader creation fn.