bug in super-gradients/src/super_gradients/training/models/conversion.py

[JagdishKolhe]

Describe the bug

A clear and concise description of what the bug is.

To Reproduce

Steps to reproduce the behavior:

1. net = models.get(model_name, pretrained_weights="coco")
2. models.convert_to_onnx(model=net, out_path=f"{model_name}.onnx", torch_onnx_export_kwargs={'input_size': (3,640,640)})
3. See error

Expected behavior

Conversion should happen without giving validation errer

Screenshots

If applicable, add screenshots to help explain your problem. models.convert_to_onnx(model=net, out_path=f"{model_name}.onnx", torch_onnx_export_kwargs={'input_size': (3,640,640)}) File "/opt/conda/lib/python3.8/site-packages/super_gradients/common/decorators/factory_decorator.py", line 36, in wrapper return func(*args, **kwargs) File "/opt/conda/lib/python3.8/site-packages/super_gradients/common/decorators/factory_decorator.py", line 36, in wrapper return func(*args, **kwargs) File "/opt/conda/lib/python3.8/site-packages/super_gradients/training/models/conversion.py", line 89, in convert_to_onnx prep_model_for_conversion_kwargs["input_size"] = (1, *input_shape) TypeError: 'NoneType' object is not iterable

The change should be: if input_shape is not None: logger.warning( "input_shape is deprecated and will be removed in the next major release." " Use the input_size kwarg in prep_model_for_conversion_kwargs instead" ) prep_model_for_conversion_kwargs["input_size"] = (1, *input_shape)

[dagshub[bot]]

Join the discussion on DagsHub!

[shaydeci]

Thanks @JagdishKolhe , we will look into it and update.

[BloodAxe]

I don't think you need to use torch_onnx_export_kwargs directly. Passing input_shape argument should do the work just fine:

from super_gradients.common.object_names import Models
from super_gradients.training import models

net = models.get(Models.YOLO_NAS_L, pretrained_weights="coco")
models.convert_to_onnx(model=net, out_path=f"yolo_nas.onnx", input_shape=(3,640,640))

@JagdishKolhe may I ask you where did you get the code snippet? It may that we have outdated documentation somewhere and I'd like to address that to prevent such issue in future. Thanks!

[JagdishKolhe]

@BloodAxe Please check the code for the method convert_to_onnx in conversion.py (line no 84) It says that input_shape argument to the function will be deprecated in future.

if input_shape is not None: logger.warning( "input_shape is deprecated and will be removed in the next major release." " Use the input_size kwarg in prep_model_for_conversion_kwargs instead" )

[BloodAxe]

Ah I see what you mean. Yeah, it's definitely a bug