centernet

1. Introduction

This is a demo to run CenterNet(backbone dlav0) object detection with BMNNSDK.

2. Usage

Put this demo dir into BMNNSDK docker container and init the environment of sdk firstly.

Download the prepared bmodels from BaiDu Netdisk: access url: https://pan.baidu.com/s/1d3f8CjzC3BF2-2I2OF0q1g access code: lt59

2.1 Generate bmodels

Change directory to data/scripts. There is a torchscript file in this directory. It is made by weights file ctdet_coco_dlav0_1x.pth which is downloaded from CenterNet model zoo To be attention, we use dlav0 as the backbone of CenterNet and we concatenate the heatmap, wh, offset output to be one. This means the output shape of the pt is 1x84x128x128

cd data/scripts

2.1.1 fp32 bmodel

./gen_fp32_bmodel.sh

After a few minutes, we can get ctdet_coco_dlav0_1x_fp32.bmodel in ./models directory

2.1.2 int8 bmodel

# usage: ./gen_int8_bmodel.sh <batch_size> <img_size> <validation_image_dir>
./gen_int_bmodel.sh 1 512 ../val2017

We choose about 200 picutures from val2017 to quantization and calibrate int8 bmodel. You can use any picture you like.

After a few minutes, we get ctdet_coco_dlav0_1x_int8_b1.bmodel or ctdet_coco_dlav0_1x_int8_b4.bmodel depends on the batch_size you use.

2.2 python demo

cd py_bmcv_sail
python3 det_centernet_bmcv_1b_4b.py \
    --input=../data/ctdet_test.jpg \
    --bmodel=../data/models/ctdet_coco_dlav0_1x_int8_b4.bmodel \
    --tpu_id=0

This demo use bmcv for preprocess, sail for inference and numpy for postproces. If success, we get result image like ctdet_result_2022_-x-x-x-x-x_b_x.jpg in current directory.

2.3 cpp demo

cd cpp_bmcv_sail
# compile in PCIE mode
make -f Makefile.pcie
./centernet_bmcv_sail.pcie \
    --bmodel=/workspace/examples/centernet_test/CenterNet_object/data/models/ctdet_coco_dlav0_1x_fp32.bmodel \
    --image=/workspace/examples/centernet_test/CenterNet_object/data/ctdet_test.jpg \
    --conf=0.35 \
    --tpu_id=0
# compile in Soc mode
# TODO

This demo use bmcv for preprocess, sail for inference and numpy for postproces. If success, we get result image like ctdet_result_2022_-x-x-x-x-x_b_x.jpg in current directory.

WARNING: The python demo only supports 1 batch or 4 batch bmodel The cpp demo only supports 1 batch bmodel