$ git clone https://github.com/bubbliiiing/facenet-pytorch.git

Follow Docker official documentation to install Docker: Install Docker Engine on Ubuntu.

Then fetch the prebuilt NPU Docker Container and run it.

$ docker pull yanwyb/npu:v1
$ docker run -it --name vim4-npu1 -v $(pwd):/home/khadas/npu \
				-v /etc/localtime:/etc/localtime:ro \
				-v /etc/timezone:/etc/timezone:ro \
				yanwyb/npu:v1

Download Tool from khadas/vim4_npu_sdk.git.

$ git clone https://gitlab.com/khadas/vim4_npu_sdk.git

After training model, modify facenet-pytorch/nets/facenet.py as follows.

diff --git a/nets/facenet.py b/nets/facenet.py
index e7a6fcd..93a81f1 100644
--- a/nets/facenet.py
+++ b/nets/facenet.py
@@ -75,7 +75,7 @@ class Facenet(nn.Module):
             x = self.Dropout(x)
             x = self.Bottleneck(x)
             x = self.last_bn(x)
-            x = F.normalize(x, p=2, dim=1)
             return x
         x = self.backbone(x)
         x = self.avg(x)

Create a python file written as follows and run to convert model to ONNX.

export.py

import torch
import numpy as np
from nets.facenet import Facenet as facenet
 
model_path = "logs/ep092-loss0.177-val_loss1.547.pth"
net = facenet(backbone="mobilenet", mode="predict").eval()
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
net.load_state_dict(torch.load(model_path, map_location=device), strict=False)
 
img = torch.zeros(1, 3, 160, 160)
torch.onnx.export(net, img, "./facenet.onnx", verbose=False, opset_version=12, input_names=['images'])

Enter vim4_npu_sdk/demo and modify convert_adla.sh as follows.

convert_adla.sh

#!/bin/bash
 
ACUITY_PATH=../bin/
#ACUITY_PATH=../python/tvm/
adla_convert=${ACUITY_PATH}adla_convert
 
 
if [ ! -e "$adla_convert" ]; then
    adla_convert=${ACUITY_PATH}adla_convert.py
fi
 
$adla_convert --model-type onnx \
        --model ./model_source/facenet/facenet.onnx \
        --inputs "images" \
        --input-shapes  "3,160,160"  \
        --dtypes "float32" \
        --inference-input-type float32 \
		--inference-output-type float32 \
        --quantize-dtype int8 --outdir onnx_output  \
        --channel-mean-value "0,0,0,255"  \
        --source-file facenet_dataset.txt  \
        --iterations 394 \
        --disable-per-channel False \
        --batch-size 1 --target-platform PRODUCT_PID0XA003

Run convert_adla.sh to generate VIM4 model. The converted model is xxx.adla in onnx_output.

$ bash convert_adla.sh

Clone the source code from our khadas/vim4_npu_applications.

$ git clone https://github.com/khadas/vim4_npu_applications.git

$ sudo apt update
$ sudo apt install libopencv-dev python3-opencv cmake

There are two modes of this demo. One is converting face images into feature vectors and saving vectors in face library. Another is comparing input face image with faces in library and outputting Euclidean distance and cosine similarity.

Put facenet_int8.adla in vim4_npu_applications/facenet/data/.

# Compile
$ cd vim4_npu_applications/retinaface
$ mkdir build
$ cd build
$ cmake ..
$ make
 
# Run mode 1
$ sudo ./facenet -m ../data/facenet_int8.adla -p 1

After running mode 1, a file named face_feature_lib will generate in vim4_npu_applications/facenet. Had this file, you can run mode 2.

# Run mode 2
$ sudo ./facenet -m ../data/facenet_int8.adla -p ../data/img/lin_2.jpg