So it would really be nice if the latest opset version of onnx would work because they are necessary to go from experimental level to production. Not having the option of opset >17 somehow negates the possibility to use torch for training models, since the new nvidia inference server with tensor rt rely on it.
I also tried dynamo_export for opset=18 but when tryining to use dynamic_shapes it does not work, which makes the model not batchable. It just makes dynamic shapes for the last two dimensions but not for batching.
My code for the dynamo_export is:
input = torch.randn([1, 3, 112, 112])
opts = torch.onnx.ExportOptions()
opts.dynamic_shapes = True
ox_p = torch.onnx.dynamo_export(model, input, export_options=opts)
Maybe you have an idea how to make the first dimension dynamic in dynamo_export?
Bets regards