search

AlexeyAB / yolo2_light

Light version of convolutional neural network Yolo v3 & v2 for objects detection with a minimum of dependencies (INT8-inference, BIT1-XNOR-inference)
MIT License
301 stars 116 forks source link

INT8 for maxpool layers #55

Open Thilanka97 opened 5 years ago

Thilanka97 commented 5 years ago

@AlexeyAB Hey, Have you tried using INT8 for Max pooling layers? What is the reason you do not use INT8 for maxpool ? Thanks in advance!

Thilanka97 commented 5 years ago

@AlexeyAB I tried maxpooling with 8INT and the mAP is as bad as 3% !!. Do u think its that bad. Have you tried this ?

Thanks in advance!

AlexeyAB commented 5 years ago

@Thilanka97 I didn't do this, but yes - it is possible.

At first, try to implement maxpool-int32 and divide inputs by l->output_multipler (or by R_MULT / (l.input_quant_multipler * l.weights_quant_multipler ) where is l is a prevous convolutional layer. If I remember it correctly :)

jasonwu1977 commented 5 years ago

@Thilanka97 is there any progress you have found with int8 maxpool layers? I am currently working on this one as well, maybe we can talk with this one. jason_wu_23@hotmail.com <-- my email

abhigoku10 commented 5 years ago

@Thilanka97 @jasonwu1977 were you guys able to obtain the quantization for the yolov2 , i am not able to get the detections when i train only