auto-ml

Just another AutoML implementation...

MLP Classifier (MultiLayer Perceptron)

Based on scikit-learn MLPClassifier

Installation

Install python 3

System dependent (see https://www.python.org/downloads/)

Install python packages with built-in package manager pip

pip install -r requirements.txt

Run

Train

python train.py -i train_input.json

"X" - samples to train with shape (n_samples, n_features)

"y" - samples classes with shape (n_samples)

Script generates "output_path" from train_input.json

Script dumps model to "model_path" from train_input.json

Predict

python predict.py -i predict_input.json

Script loads model from "model_path" from predict_input.json

Script generates "output_path" from predict_input.json

Field "y_pred" in "output_path" is predicted classes

Docker

Build

docker build -t auto-ml-app .

Train

docker run -it \
           -v "$(pwd)"/output:/usr/src/output \
           -w /usr/src/output \
           -v "$(pwd)"/train_input.json:/usr/src/input/train_input.json \
           auto-ml-app python /usr/src/app/train.py -i /usr/src/input/train_input.json

Predict

docker run -it \
           -v "$(pwd)"/output:/usr/src/output \
           -w /usr/src/output \
           -v "$(pwd)"/predict_input.json:/usr/src/input/predict_input.json \
           -v "$(pwd)"/output/model.joblib:/usr/src/input/model.joblib \
           auto-ml-app python /usr/src/app/predict.py -i /usr/src/input/predict_input.json

Save and load

docker save auto-ml-app | gzip > auto-ml-app.tar.gz

cat auto-ml-app.tar.gz | docker import - auto-ml-app:latest

docker load < auto-ml-app.tar.gz

Computer Vision

Based on

• MMCV
• MMDetection

Installation

Prerequisites

• Linux or macOS (Windows is in experimental support)
• Python 3.6+
• PyTorch 1.3+
• CUDA 9.2+ (If you build PyTorch from source, CUDA 9.0 is also compatible)
• GCC 5+
• MMCV

See https://github.com/open-mmlab/mmdetection/blob/master/docs/get_started.md

Install python 3

System dependent (see https://www.python.org/downloads/)

Install python packages with built-in package manager pip

pip install -r requirements_mmdet.txt

Train

Prepare data in data directory (Example)

python get_mmdet_configs.py 
python make_coco.py -i make_coco_input.json

Run

python train_mmdet.py -i train_mmdet_input.json

Results will be in data/works directory

Inference

Prepare data in data directory (Example, also requires config and checkpoint file from trained model)

cp test.jpg data/test.jpg

Run

python inference_mmdet.py -i inference_mmdet_input.json

Results will be in data directory (result.jpg by default)

Docker

Based on MMDetection docker

Prerequisites

• CUDA Driver or CUDA Toolkit (CUDA Driver included)
• NVIDIA Container Toolkit

What is the difference between CUDA, CUDNN, CUDA Driver, CUDA Toolkit, and NCVV?

Build

Change Dockerfile_mmdet ARGs according to host lib versions (see Dockerfile_mmdet)

ARG PYTORCH="1.8.0"
ARG CUDA="11.1"
ARG CUDNN="8"
ARG CUDA_MMCV="111"
ARG MMCV="1.3.5"
ARG MMDET="2.13.0"

Build image

docker build -t auto-ml-mmdet -f Dockerfile_mmdet_2 . 

Train

Prepare data in data directory (Example)

python get_mmdet_configs.py 
python make_coco.py -i make_coco_input.json

Data structure

• dataset - set of images with annotations (labels, bboxes, polygons, etc)
• model - torch or caffe neural net model (tensorflow or keras in the future?)
• config - text file with neural net structure, train params, dataset description
• checkpoint - binary or text file with neural net weights
• work - directory with train logs, metrics and intermediate checkpoints

data
|__ datasets
    |__ dataset1
        |__ images
        |__ annotation
        |__ ... (depends on)
    |__ dataset2
    |__ dataset3
    ...
|__ configs
    |__ model1
        |__ config1
        |__ config2
        |__ config3
        ...
    |__ model2
    |__ model3
    ...   
|__ checkpoints
    |__ model1
        |__ submodel1
            |__ checkpoint1
            |__ checkpoint2
            |__ checkpoint3
            ...
        |__ submodel2
        |__ submodel3
        ...
    |__ model2
    |__ model3
    ...
|__ works
    |__ work1
        |__ log
        |__ checkpoint1
        |__ checkpoint2
        |__ checkpoint3
        ...
        |__ checkpoint_latest
    |__ work2
    |__ work3
    ...

Run

docker run --gpus all --shm-size 8G \
           -v "$(pwd)"/data:/auto-ml/data \
           -v "$(pwd)"/train_mmdet_input.json:/auto-ml/train_mmdet_input.json \
           auto-ml-mmdet python3 train_mmdet.py

• --gpus all - GPU devices to add to the container ('all' to pass all GPUs)
• --shm-size 8G - Size of /dev/shm. The format is . number must be greater than 0. Unit is optional and can be b (bytes), k (kilobytes), m (megabytes), or g (gigabytes). If you omit the unit, the system uses bytes. If you omit the size entirely, the system uses 64m.
• -v "$(pwd)"/data:/auto-ml/data - mount host data dir "$(pwd)"/data to container data dir /auto-ml/data
• -v "$(pwd)"/train_mmdet_input.json:/auto-ml/train_mmdet_input.json - mount host train config "$(pwd)"/train_mmdet_input.json to container train config /auto-ml/train_mmdet_input.json
• python train_mmdet.py - run train script

Results will be in data/works directory

Inference

Prepare data in data directory (Example, also requires config and checkpoint file from trained model)

cp test.jpg data/test.jpg

Run

docker run --gpus all --shm-size 8G \
           -v "$(pwd)"/data:/auto-ml/data \
           -v "$(pwd)"/inference_mmdet_input.json:/auto-ml/inference_mmdet_input.json \
           auto-ml-mmdet python3 inference_mmdet.py

• --gpus all - GPU devices to add to the container ('all' to pass all GPUs)
• --shm-size 8G - Size of /dev/shm. The format is . number must be greater than 0. Unit is optional and can be b (bytes), k (kilobytes), m (megabytes), or g (gigabytes). If you omit the unit, the system uses bytes. If you omit the size entirely, the system uses 64m.
• -v "$(pwd)"/data:/auto-ml/data - mount host data dir "$(pwd)"/data to container data dir /auto-ml/data
• -v "$(pwd)"/inference_mmdet_input.json:/auto-ml/inference_mmdet_input.json - mount host inference config "$(pwd)"/inference_mmdet_input.json to container inference config /auto-ml/inference_mmdet_input.json auto-ml-mmdet
• python inference_mmdet.py - run inference script

Results will be in data directory (result.jpg and result.jsonby default)

JSON result format:

array
|__ bboxes
    |__ class1
        |__ bbox1
           |__ x
           |__ y
           |__ width
           |__ height
           |__ score
        |__ bbox2
        |__ bbox3
        |__ ...
    |__ class2
    |__ class3
    ...
|__ segms
    |__ class1
        |__ segm1 (boolean 2D array with size height x width)
        |__ segm2
        |__ segm3
        ... equals number of bboxes for corresponding class
    |__ class2
    |__ class3
    |__ ...

Save and load

docker save auto-ml-mmdet | gzip > ../auto-ml-mmdet.tar.gz

cat auto-ml-mmdet.tar.gz | docker import - auto-ml-mmdet:latest

docker load < auto-ml-mmdet.tar.gz