Yolov5 change silu to relu accuracy problem after exporting to onnx.

[maro-jeon]

Hi,
I saw your study about changing activation function from silu to relu.
The score shows that ReLU mAP is 55.7 (swish pretrained).
so I just changed swish activation to relu activation without any fine-tuning process.

from

to

Then, the result of accuracy was zero.

Can you describe more about how to achieve the score 55.7 (swish pretrained) for reproducing the score.

Thanks,
Maro JEON

[Tyler-D]

You have to retrain the model.

[maro-jeon]

Thanks, you just fine tune ? or train many epochs ?

[Tyler-D]

Train for 300 epochs.