How to use dynamic_shape in ai_edge_torch.convert

[pradanaadn]

Hi, I want to understand the use of dynamic_shape in ai_edge_torch.convert for a CNN model. I'm trying the following:

batch = torch.export.Dim('batch', min=1, max=64)
tfl_converter_flags = {'target_spec': {'supported_ops': [tf.lite.OpsSet.TFLITE_BUILTINS]}}

dynamic_shape = {
    'x':{0:batch}
}
edge_model = ai_edge_torch.convert(resnet18.eval(), sample_inputs, dynamic_shapes=dynamic_shape, _ai_edge_converter_flags=tfl_converter_flags)

However, this returns the following TypeError: Only non-negative indices are allowed when broadcasting static shapes, but got shape (-9223372036854775808, 112, 112, 64).

My goal is to make the batch size dynamic. Below is the colab for reproduce my error:
https://colab.research.google.com/drive/1r0qCp1w3GHNBHeGHQtGeHI9eESiNoaHE?usp=sharing

[gaikwadrahul8]

Hi, @pradanaadn
Thank you for bringing this issue to our attention, I am able to reproduce the same behavior from my end also TypeError: Only non-negative indices are allowed when broadcasting static shapes, but got shape (-9223372036854775808, 112, 112, 64) so for investigation purpose I'm keeping gist-file here for reference so we will have to dig more into this issue but it seems like packages version dependancies issue between torch and ai_edge_torch, We will update you on this issue. Thank you for your cooperation and understanding.