Step 0. Environment


Note: You need to run pip uninstall mmcv first if you have mmcv installed. If mmcv and mmcv-full are both installed, there will be ModuleNotFoundError.

Install kneron-mmpose

  1. We recommend you installing mmcv-full with pip:

pip install mmcv-full -f{cu_version}/{torch_version}/index.html
Please replace {cu_version} and {torch_version} in the url to your desired one. For example, to install the latest mmcv-full with CUDA 11.0 and PyTorch 1.7.0, use the following command:
pip install mmcv-full -f
See here for different versions of MMCV compatible to different PyTorch and CUDA versions.

  1. Clone the Kneron-version kneron-mmpose repository.
git clone
cd kneron-mmpose
  1. Install required python packages for building and installing kneron-mmpose.

    pip install -r requirements/build.txt
    pip install -v -e .  # or "python develop"

Step 1: Training models on standard datasets

MMPose provides hundreds of existing and existing pose models in Model Zoo, and supports several standard datasets like COCO, MPII, FREIHAND, etc. This note demonstrates how to perform common object detection tasks with these existing models and standard datasets, including:

Train models on standard datasets

MMPose also provides out-of-the-box tools for training pose models. This section will show how to train models under configs on standard datasets i.e. FREIHAND.

Important: You might need to modify the config file according your GPUs resource (such as "samples_per_gpu","workers_per_gpu" ...etc due to your GPUs RAM limitation). The default learning rate in config files is for 1 GPUs and 64 img/gpu (batch size = 64*1 = 64).

Step 1-1: Prepare datasets

Public datasets such as FREIHAND. We suggest that you download and extract the dataset to somewhere outside the project directory and symlink (ln) the dataset root to kneron-mmpose/data(ln -s realpath/to/freihand kneron-mmpose/data/freihand), as shown below:

├── mmpose
├── configs
├── tools
├── data
│   ├── freihand
│   │   ├── training
│   │   ├── annotations
│   │   ├── evaluation
│   ├── rhd
│   │   ├── training
│   │   ├── annotations
│   │   ├── evaluation

It's recommended to symlink the dataset folder to kneron-mmpose folder. However, if you place your dataset folder at different place and do not want to symlink, you have to change the corresponding paths in config files (absolute path is recommended).

Step 1-2: Training Example with RSN18:

RSN: Range Sparse Net for Efficient, Accurate LiDAR 3D Object Detection

We only need the configuration file (which is provided in configs/hand/2d_kpt_sview_rgb_img/topdown_heatmap/freihand2d/ to train pose model:

python tools/ configs/hand/2d_kpt_sview_rgb_img/topdown_heatmap/freihand2d/
* (Note) you might need to create a folder name 'work_dir' in kneron-mmpose root folder because we set 'work_dir' as default folder in '' * (Note 2) The whole training process might take several days, depending on your computational resource (number of GPUs, etc). If you just want to take a quick look at the deployment flow, we suggest that you download our trained model so you can skip the training process:
mkdir work_dirs
cd work_dirs
wget (&)download_link_for_pretrained_model
cd ..(&&)
* (Note 3) This is a "training from scratch" tutorial, which might need lots of time and gpu resource. If you want to train a model on your custom dataset, it is recommended that you read,

Step 2: Test trained model

'tools/' is a script that generates inference results from test set with our pytorch model and evaluates the results to see if our pytorch model is well trained (if --eval argument is given). Note that it's always good to evluate our pytorch model before deploying it.

python toos/ \
    configs/hand/2d_kpt_sview_rgb_img/topdown_heatmap/freihand2d/ \
    work_dirs/latest.pth \
    --eval AUC EPE PCK \
* configs/hand/2d_kpt_sview_rgb_img/topdown_heatmap/freihand2d/ is your pose training config * work_dirs/latest.pth is your trained pose model

The expected result of the command above will be something similar to the following text (the numbers may slightly differ):

AUC: 0.8579125542518
EPE: 3.5918244972577695
PCK: 0.9885881741008469

Step 3: Export onnx

'tools/deployment/' is a script provided by MMPose to help user to convert our trained pth model to onnx:

python tools/deployment/ \
    configs/hand/2d_kpt_sview_rgb_img/topdown_heatmap/freihand2d/ \
    work_dirs/rsn18_freihand2d_224x224/latest.pth \
    --output-file work_dirs/rsn18_freihand2d_224x224/latest.onnx \
* 'configs/hand/2d_kpt_sview_rgb_img/topdown_heatmap/freihand2d/' is your pose training config * 'work_dirs/latest.pth' is your trained pose model

The output onnx should be the same name as 'work_dirs/latest.pth' with '.onnx' post-fix in the same folder.

Step 4: Convert onnx to NEF model for Kneron platform

Step 4-1: Install Kneron toolchain docker:

Step 4-2: Mout Kneron toolchain docker

Step 4-3: Import KTC and required lib in python shell

Step 4-4: Optimize the onnx model

onnx_path = '/data1/latest.onnx'
m = onnx.load(onnx_path)
m = ktc.onnx_optimizer.onnx2onnx_flow(m),'latest.opt.onnx')

Step 4-5: Configure and load data necessary for ktc, and check if onnx is ok for toolchain

# npu (only) performance simulation
km = ktc.ModelConfig((&)model_id_on_public_field, "0001", "720", onnx_model=m)
eval_result = km.evaluate()
print("\nNpu performance evaluation result:\n" + str(eval_result))

Step 4-6: Quantize the onnx model

We random sampled 50 images from voc dataset (50 images) as quantization data , we have to 1. Download the data 2. Uncompression the data as folder named voc_data50" 3. Put the voc_data50 into docker mounted folder (the path in docker container should be /data1/voc_data50)

The following script will do some preprocess(should be the same as training code) on our quantization data, and put it in a list:

import os
from os import walk

img_list = []
for (dirpath, dirnames, filenames) in walk("/data1/voc_data50"):
    for f in filenames:
        fullpath = os.path.join(dirpath, f)

        image =
        image = image.convert("RGB")
        image = Image.fromarray(np.array(image)[...,::-1])
        img_data = np.array(image.resize((640, 640), Image.BILINEAR)) / 256 - 0.5

Then perform quantization. The BIE model will be generated at /data1/output.bie.

# fixed-point analysis
bie_model_path = km.analysis({"input": img_list})
print("\nFixed-point analysis done. Save bie model to '" + str(bie_model_path) + "'")

Step 4-7: Compile

The final step is compile the BIE model into an NEF model.

# compile
nef_model_path = ktc.compile([km])
print("\nCompile done. Save Nef file to '" + str(nef_model_path) + "'")

You can find the NEF file at /data1/batch_compile/models_720.nef. models_720.nef is the final compiled model.