Quantization and Deployment Tips for YOLO

Quantization and Deployment Tips for YOLO - 1.4 English

Vitis AI Optimizer User Guide (UG1333)

Document ID

UG1333

Release Date

2021-07-22

Version

1.4 English

After converting to the Caffe model, YOLOv3 can be deployed to the FPGA after quantization. Follow Vitis AI User Guide to quantize and deploy.

Set pre-processing for YOLO in prototxt.

YOLO has its own pre-processing scheme. This must be specified for better performance. Set yolo_height and yolo_width to the input height and width of network in transform_param as shown in the following code snippet.

layer {
  name: "data"
  type: "ImageData"
  top: "data"
  top: "label"
  include {
    phase: TEST
  }
  transform_param {
    mirror: false
    yolo_height: 768
    yolo_width: 1280
  }
  image_data_param {
    source: "file_list.txt"
    root_folder: "/data/dataset/"
    batch_size: 10
    shuffle: false
  }
}

Provide labels in file list.
Though labels are not required by quantization and will not affect the results, the ImageData layer in Caffe needs to check labels in the source file. Ensure the file list assigned in source has a number after each image file name. Example of source file list is shown as follows:
```
ILSVRC2012_val_00000001.JPEG 65
ILSVRC2012_val_00000002.JPEG 970
ILSVRC2012_val_00000003.JPEG 230
ILSVRC2012_val_00000004.JPEG 809
ILSVRC2012_val_00000005.JPEG 516
...
```