Open ablyzniuk opened 6 months ago
How to keep different picture sizes(I meant dynamic input shape size, not only alongside batch axis) as the original ultralytics approach describes?
If you want dynamic shape input size, you'd better export an onnx with dynamic shape. But there will be some difference in inference code. Such as in cudaMalloc and set trt bindings.
How to keep different picture sizes(I meant dynamic input shape size, not only alongside batch axis) as the original ultralytics approach describes?