import torch
model = torch.jit.load("your_model.pt")
x = torch.randn(1, 3, 640, 640, requires_grad=True)
torch.onnx.export(model, x, "model.onnx", verbose=True, input_names=["input"], output_names=["output"])