You can also train on Maixhub.com, just upload your datasets and you will get the result(kmodel and usage code)

• Object classification(Mobilenet V1): judge class of image
• Object detection(YOLO v2): find a recognizable object in the picture

• Prepare environment, use CPU or GPU to train At your fist time train, CPU is recommended, just

pip3 install -r requirements.txt

or use aliyun's source if you are in China

pip3 install -r requirements.txt -i https://mirrors.aliyun.com/pypi/simple/

• Download nncase and unzip it to tools/ncc/ncc_v0.1, and the executable path is tools/ncc/ncc_v0.1/ncc
• python3 train.py init
• Edit instance/config.py according to your hardware
• Prepare dataset, in the datasets directory has some example datasets, input size if 224x224 or you just fllow maixhub's conduct

python3 train.py -t classifier -z datasets/test_classifier_datasets.zip train

or assign datasets directory

python3 train.py -t classifier -d datasets/test_classifier_datasets train

more command seepython3 train.py -h

and you will see output in the out directory, packed as a zip file

python3 train.py -t detector -d data.json train

{
    "train": [
        "G:/skiing/dataset/0",
        "G:/skiing/dataset/1"
    ],
    "val": [
        "G:/skiing/dataset/val"
    ]
}

more command seepython3 train.py -h

and you will see output in the out directory, packed as a zip file

Use docker or install tensorflow with GPU in your local environment

Tensorflow's version should >= 2.0, tested on 2.1

If you encounter the following error message, please just ignore it as this project was only built for Linux before.

no GPU, will use [CPU]
2021-01-12 22:04:48.773378: I tensorflow/stream_executor/platform/default/dso_loader.cc:49] Successfully opened dynamic library cudart64_110.dll
Traceback (most recent call last):
 File "<string>", line 1, in <module>
 File "C:\Users\<user>\AppData\Local\Programs\Python\Python38\lib\multiprocessing\spawn.py", line 109, in spawn_main
   fd = msvcrt.open_osfhandle(new_handle, os.O_RDONLY)
OSError: [Errno 9] Bad file descriptor

This repo can actually runs on Windows with GPU support now. The above error message will not affect your GPU training capability.

• OS: Windows 10 Insider Preview Build 20279.1
• GPU: GeForce RTX 3080
• GPU Driver: Game Ready Driver 465.12
• CUDA Toolkit: 11.0 and 11.2 side-by-side installation
• cuDNN: 11.1
• tensorflow: 2.4

see tensorflow official website (或者可以参考这篇教程)

• Install docker
• Install nvidia-docker
• Pull docker image

docker pull neucrack/tensorflow-gpu-py3-jupyterlab

or

docker pull daocloud.io/neucrack/tensorflow-gpu-py3-jupyterlab

• Test environment

docker run --gpus all -it --rm neucrack/tensorflow-gpu-py3-jupyterlab python -c "import tensorflow as tf; print('-----version:{}, gpu:{}, 1+2={}'.format(tf.__version__, tf.test.is_gpu_available(), tf.add(1, 2).numpy()) );"

if output is-----version:2.1.0, gpu:True, 1+2=3, that's ok(maybe version can > 2.1.0)

• Create docker container

docker run --gpus all --name jupyterlab-gpu -it -p 8889:8889 -e USER_NAME=$USER -e USER_ID=`id -u $USER` -e GROUP_NAME=`id -gn $USER` -e GROUP_ID=`id -g $USER` -v /home/${USER}:/tf neucrack/tensorflow-gpu-py3-jupyterlab

If used daocloud, image name should be change to daocloud.io/neucrack/tensorflow-gpu-py3-jupyterlab

This will mount your/home/$USER directory to /tf directory of container, the /tf is the root dir of jupyterlab

Stop by docker stop jupyterlab-gpu, start again by docker start jupyterlab-gpu To use sudo command, edit user password by

docker exec -it jupyterlab_gpu /bin/bash
passwd $USER
passwd root

• use jupyterlab

Open http://127.0.0.1:8889/lab? in browser, input token(see docker start log) and set new password

Use docker stop jupyterlab-gpu to stop server Use docker start jupyterlab-gpu to start service again

refer to tensorflow official website

Apache 2.0, see LICENSE