Alexey

@Thilanka97 I quantize (FP32 -> INT8) weights before the inference. I just don't save it as 8-bit.

@Thilanka97 It loads yolov3.weights with FP32 weights. Then it quantizes the weights FP32 -> INT8 once during initialization, except 1st and one conv-layer before each [yolo]-layer. Then during inference it...

> Although you said that you use 32FP for the first layer, at the end of the first layer the output is quantized into 8 INT right? I mean the...

Try to run ./darknet detector test ... without image.jpg at the end of command And then enter image.jpg and press Enter, repeat several times. Do it on both frameworks. And...

The 1st and the last layers greatly affect accuracy. So it is done to loss not more than ~1% of mAP by using INT8.

@fidodido2503 What is wrong in this line?

I fixed it. Try to update your code from GitHub and recompile. 

@ahmed-nady Use the latest version of cuDNN

The last layer must be fp32, since it is used for object coordinates. May be it can be int8 only if we use much more anchors (initial w,h defined with...

@USMANHEART I fixed it. Update your code from GitHub.