Xilinx Contest 2023

TUTORIAL MNIST

En la carpeta MNIST_ALVEO se encuentra el tutorial MNIST. Sus scripts han sido MODIFICADOS para funcionar con la Alveo U280.

Si venimos de una instalación limpia, debemos configurar la máquina host siguiendo este tutorial: https://github.com/Xilinx/Vitis-AI/tree/1.4.1/setup/alveo Básicamente, solo debemos hacer source del fichero install.sh de dicho tutorial. Debemos instalar Vitis 1.4, ya que dicho Tutorial sólo funciona con dicha versión: docker pull xilinx/vitis-ai:1.4.1.978

Quantizing & Compiling TRT_Pose model with Vitis-AI

NOTE: all these steps should be performed inside a Vitis-AI Pytorch container or with Vitis-AI installed locally 

Vitis-AI image used for this work: https://hub.docker.com/r/xilinx/vitis-ai-pytorch-cpu

1. Download pretrained model from trt_pose github → https://drive.google.com/open?id=1XYDdCUdiF2xxx4rznmLb62SdOUZuoNbd

2. Download COCO validation dataset, 2017 with keypoints annotations

   • Validation dataset
   • Annotations

3. Preprocess COCO annotations with vitis/calib_dataset/preprocess_coco_person.py

   • python3 preprocess_coco_person.py <input_annotation_file> <output_annotation_file>
   • Use person_keypoints_val2017.json as input file

4. Apply quantization with vitis/quantize_trt_pose.py

   • python3 quantize_trt_pose.py --model_dir <pretrained_model.pth> --quant_mode calib

5. Compile model with vitis/vai_c_xir_cust_pytorch.py

   • python3 vai_c_xir_cust_pytorch.py -x <xmodel> -a <arch> -o <output_dir>
   • <xmodel> is the .xmodel file resulting from quantize_trt_pose.py execution
   • <arch> is a JSON file containing the DPU architecture. This is found in the Vitis-AI tool installation under the route /opt/vitis_ai/compiler/arch/<DPU>/<HW_PLATFORM>/arch.json

6. The result should be 3 files:

   • md5sum.txt
   • meta.json
   • <model_name>.xmodel