Error trying to export models with timm encoder to onnx format

[rgsousa88]

Hi,

Is there any issue related to onnx exporting when using timm models as encoders? I'm trying to export a model with timm-MobileNetV3 as encoder and FPN as decoder.

I'm running the script in a conda environment with python=3.7, pytorch=1.8, segmenation_model_pytorch=0.2.1

Building network:

import segmentation_models_pytorch as smp

def build_fpn_mobv3(input_shape): 
    model = smp.FPN(encoder_name="timm-mobilenetv3_large_100",
                    encoder_weights=None, 
                    in_channels=3,
                    classes=1,
                    activation="sigmoid")

    shape = (1,3,) + input_shape
    x = torch.zeros(shape,dtype=torch.float32,device=torch.device('cpu'))
    model.eval()
    model(x)

    return model 

Exporting trained model

onnxpath = os.path.join(ckpt_path, f"{model_name}.onnx")
dynamic_axes = {0: 'batch_size'}
shape = (1,3,) + args.shape
dummy_input = torch.zeros(shape,requires_grad=True).float().to(device)
net.eval()

torch.onnx._export(net, dummy_input, onnxpath,
                   export_params=True, verbose=True,
                   input_names=['image'], output_names=['maps'],
                   keep_initializers_as_inputs=False,
                   dynamic_axes={'image': {0: 'batch'}, 'maps': {0: 'batch'}},
                   opset_version=10)

The error is: RuntimeError: Unsupported: ONNX export of Pad in opset 9. The sizes of the padding must be constant. Please try opset version 11.

Is there any workaround to export using opset version 10? I'm able to export using other encoders but none of them are timm ones.

Thanks for attention and time.

[EricPHassey]

Also wondering this. Any progress or updates we can do?

[github-actions[bot]]

This issue is stale because it has been open 60 days with no activity. Remove stale label or comment or this will be closed in 7 days.

[rgsousa88]

Hi, @EricPHassey

Sorry for the late reply... I have "solved" it by building a custom encoder using MobileNetV3 code available at torchvision.