Jordan Fix

Hi @MaxS1996, the current default behavior is to try to set the batch size to the number of input images provided. That is, you're setting the batch size to 1...

Hard to tell what's going wrong without digging in a bit. If you're willing to change some source code to help debug on the ONNX side, it would be useful...

So IIUC the model is expecting `[28, 28]` but `image-classifier` is creating images with `[1, 28, 28, 1]`? It shouldn't be so hard to allow for this, but would require...

Hmm, if this is a know issue in folly then I think the main workaround here is to just build folly separately with some commit that works and then place...

Hi @leejaymin, I think this makes more sense to implement inside `transformForPrecisionMode()`. We provide the Backend we are optimizing for in that function anyway. Then you can query the `Backend::isOpSupported()`...

Not really sure. It may have been there conservatively originally by the author who only was dealing with specific usecases/dimensions for gathering. Probably would make sense to remove it and...

>I think this transformation is dangerous since it can move reshape between an input placeholder and quantize. Then, your fully quantized model becomes partially quantized (the reshape is now float)...

> For a fully quantized IR, actual model inputs are the quantize nodes (the input image contains integer data, not float data) and if the reshape is be moved to...

CC: @vuzelac-cadence

I'm not totally sure the effective ask here -- are you suggesting that we would have an optimization that, when sinking transposes, if it finds an NHWC Convolution with transposed...