lower object percentage using -quantized

[rimu123]

Hi, thanks for the great work! I developed a new model like mobilenet-yolov3. I get lower object percentage when using -quantized . It is about 10 percentage lower than without quantized in GPU. I am very sure I add input_calibration generated by myself dataset in cfg. I heard that the small model with quantification has very low precision. I don't know if it is correct, can you give me some advice? thank you！

[AlexeyAB]

@rimu123 Hi,

I get lower object percentage when using -quantized

I recommend to use lower threshold for INT8 -thresh 0.15 than for Float32 -thresh 0.25. It will solve your issue.

---

I heard that the small model with quantification has very low precision. I don't know if it is correct, can you give me some advice?

Just check your mAP with -quantized and without it.

• /darknet detector map data/obj.data yolo_obj.cfg data/yolo_obj_10000.weights -thresh 0.25

• /darknet detector map data/obj.data yolo_obj.cfg data/yolo_obj_10000.weights -thresh 0.15 -quantized

Usually there is much lower probability of class 15 instead of 25, i.e. -60%. But there is small drop in accuracy mAP, something like: 75% instead of 78%, i.e. only -1%.

More: https://github.com/AlexeyAB/yolo2_light/issues/24

[rimu123]

@AlexeyAB thank you for your reply. Yes ,I did it follow you. for this: /darknet detector map data/obj.data yolo_obj.cfg data/yolo_obj_10000.weights -thresh 0.25 mAP = 50.3% for this: /darknet detector map data/obj.data yolo_obj.cfg data/yolo_obj_10000.weights -thresh 0.15 -quantized mAP = 40.6% There is big drop in accuracy mAP i.e. -10%. I think that is not normal. I disable quantized for first few layers. but it does not work. Now I don't know how to do.

[AlexeyAB]

I disable quantized for first few layers. but it does not work. Now I don't know how to do.

• How did you do it?

• What model do you use, is it yolov3.cfg?

• How many iterations did you train?

• Try to use default input_calibration= parameters

[rimu123]

@AlexeyAB thank you for your reply.

How did you do it?

I modified the line of code in additionally.c. #3209

if (params.index == 0 || activation == LINEAR || (params.index > 1 && stride>1) || size==1)
        quantized = 0; // disable Quantized for 1st and last layers

for：

if (params.index < 4 || activation == LINEAR || (params.index > 4 && stride>1) || size==1)
      quantized = 0; // disable Quantized for 1st and last layers

I think that means disable quantized for first three layers.

What model do you use, is it yolov3.cfg?

I developed a new model that base bone is peleenet, a small classification network. the model has 27M and 192 layers.

How many iterations did you train?

I trained the new model 300k for iterations. I think that mean the model has converged.

Try to use default input_calibration= parameters

I use default input_calibration= parameters of yolov3.cfg for model, mAP = 45.7, about 5% drop. It is surprising. I do not know why. I am very sure I used validation set to generate input_calibration. i.e. input_calibration = 15.497, 12.2827, 15.2862, 15.7211, 15.497, 15.8132, 15.8132, 15.497, 15.8132, 15.497, 15.8132, 15.8132, 15.497, 15.8132, 15.497, 15.8132, 15.497, 15.8132, 15.497, 15.8132, 15.497, 15.8132, 15.497, 15.8132, 15.497, 15.8132, 15.497, 15.8132, 15.8132, 15.497, 15.8132, 15.497, 15.8132, 15.497, 15.8132, 15.497, 15.8132, 15.497, 15.8132, 15.497, 15.8132, 15.497, 15.8132, 15.497, 15.8132, 15.8132, 15.497, 15.8132, 15.497, 15.8132, 15.497, 15.8132, 15.497, 15.8132, 15.8132, 15.8132, 15.8132, 15.8132, 13.7261, 15.497, 15.497, 15.8132, 15.8132, 15.8132, 15.8132, 15.8132, 15.8132, 15.497, 15.497, 15.8132, 15.8132, 15.8132, 15.8132, 15.8132, 15.8132, 16 why it is worse effect than default input_calibration?

[AlexeyAB]

@rimu123

if (params.index < 4 || activation == LINEAR || (params.index > 4 && stride>1) || size==1) quantized = 0; // disable Quantized for 1st and last layers I think that means disable quantized for first three layers.

Yes.

---

why it is worse effect than default input_calibration?

May be something wrong in my code for calibration.

---

I use default input_calibration= parameters of yolov3.cfg for model, mAP = 45.7, about 5% drop. It is surprising. I do not know why. I am very sure I used validation set to generate input_calibration. i.e. input_calibration = 15.497, 12.2827, 15.2862, 15.7211, ....

Also you can try to tune input calibration manually. Just use 48 instead of 40, or 32 instead of 40 and check the mAP: https://github.com/AlexeyAB/yolo2_light/blob/117f196518597a902bc4c0552724f4fb99f09e3d/bin/yolov3.cfg#L25

[rimu123]

@AlexeyAB thank you for your reply. I tried to use 48, 32 instead of 40, and some of their combinations. In addition, I modified the first few input calibration learning from myself input_calibration. The best mAP = 45.8%. Maybe this is the best result. Of course, such modifications are blind. By the way, a lot of 40 in input_calibration, whether it mean that the dynamic range of the output of each layer is fixed? Similarly, whether it also reflects some of the characteristics of cnn?

[abhigoku10]

@rimu123 hi when applied quantization did ur weight size reduce ?/

[ReekiLee]

@AlexeyAB Hello,

1. Does -quantized work only in the yolo2_light repo?

2. What'more, I found whenever I ran with -quantized or without it, the prediction time always equals to 0.00000 as follows: Loading weights from yolov3-tiny_final_572.weights...Done! test1.jpg: Predicted in 0.000000 seconds. classes= 1: 100% (left_x: 69 top_y: 251 width: 83 height: 98) classes= 1: 100% (left_x: 248 top_y: 174 width: 74 height: 81) classes= 1: 100% (left_x: 454 top_y: 206 width: 62 height: 88) classes= 1: 92% (left_x: 461 top_y: 352 width: 80 height: 97) classes= 1: 97% (left_x: 704 top_y: 141 width: 68 height: 76) classes= 1: 99% (left_x: 779 top_y: 282 width: 84 height: 90) Not compiled with OpenCV, saving to predictions.png instead Could you tell me where maybe the bug is?

3. When I calculate mAP, it shows: Set -points flag: -points 101 for MS COCO -points 11 for PascalVOC 2007 (uncomment difficult in voc.data) -points 0 (AUC) for ImageNet, PascalVOC 2010-2012, your custom dataset in the end. Could you clarify what's the meaning of it?

4. At last, in the Yolo2_lights repo, what's the meaning of difficult=data/difficult_2007_test.txt in the voc.data file? If I don't run the voc_label_difficult.py, and set difficult the same as valid, is it ok?

Thank you in advance!!!