Issue exporting VGT model to ONNX - RuntimeError: invalid unordered_map<K, T> key

[meichenberger78]

trafficstars

Hi all,

I'm attempting to export the VGT model to ONNX format, but I'm encountering an issue where I get the following error:

File ...\VGT\object_detection\ditod\VGTbeit.py:795
vis_x = checkpoint.checkpoint(blk, vis_x, rel_pos_bias, training_window_size)
...
RuntimeError: invalid unordered_map<K, T> key

I’ve set up the export for the model’s backbone using dummy data for both image and grid inputs as follows:

VGTTrainer.py

def export_to_onnx(self, export_path, opset_version=17, device="cpu"):

dummy_image = torch.randn(3, 1132, 800).to(device)
dummy_input_ids = torch.randint(0, 30522, (1, 1209)).to(device) 
dummy_bbox = torch.randn(1, 1209, 4).to(device)
dummy_height = torch.tensor([1132], dtype=torch.int64).to(device)
dummy_width = torch.tensor([800], dtype=torch.int64).to(device)

torch.onnx.export(
    self.model.backbone,
   (dummy_image, dummy_bbox, dummy_input_ids, dummy_height, dummy_width), 
    export_path,
    opset_version=opset_version,
   input_names=["image", "bbox", "input_ids", "height", "width"],
    output_names=["instances"],
    dynamic_axes= {
                        "image": { 1: "height", 2: "width"},
                        "bbox": {0: "num_boxes"},
                        "input_ids": {0: "num_input_ids"}
                    }
)

I encounter the same exception if I try exporting the entire model (self.model) with dummy data.

Any guidance on troubleshooting or resolving this export error would be greatly appreciated!

Thank you!