How to perform quantization

[escorciav]

❔Question

Dear @debapriyamaji, thanks a lot for providing the code of YOLOv5Pose. I was wondering,

• Did you release the scripts (functions, classes, utilities) for quantization?

• Could you please provide more details about the mixed-precision results?

Any pointer it's highly appreciated, as it's my first time using Pytorch quantization API.

Additional context

• I'm using the Pytorch Quantization API. I added the corresponding torch.quantization.QuantStub/DeQuantStub & nn.quantized.FloatFunctional boilerplate.

• ATM I achieve low performance with static PTQ & QINT8. I'm testing on my desktop CPU, i.e., I haven't done on-device testing yet.

Results

Dataset: COCO-val
Quantization type: Static PTQ
Model: yolov5s6_pose_640_ti_lite/last.pt (md5sum: f3a72894e9d0d4a4095174b0e795d7b8)
Data type: QINT8

 Average Precision  (AP) @[ IoU=0.50:0.95 | area=   all | maxDets= 20 ] =  0.079
 Average Precision  (AP) @[ IoU=0.50      | area=   all | maxDets= 20 ] =  0.289
 Average Precision  (AP) @[ IoU=0.75      | area=   all | maxDets= 20 ] =  0.012
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=medium | maxDets= 20 ] =  0.053
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= large | maxDets= 20 ] =  0.118
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets= 20 ] =  0.248
 Average Recall     (AR) @[ IoU=0.50      | area=   all | maxDets= 20 ] =  0.618
 Average Recall     (AR) @[ IoU=0.75      | area=   all | maxDets= 20 ] =  0.147
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=medium | maxDets= 20 ] =  0.171
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= large | maxDets= 20 ] =  0.353