lightforever/mlcomp

The goal of MLComp is to provide tools for training, inferencing, creating complex pipelines (especially for computer vision) in a rapid, well manageable way.
MLComp is compatible with: Python 3.6+, Unix operation system.

Part of Catalyst Ecosystem. Project manifest.

Features

• Amazing UI
• Catalyst support
• Distributed training
• Supervisor that controls computational resources
• Synchronization of both code and data
• Resource monitoring
• Full functionality of the pause and continue on UI
• Auto control of the requirements
• Code dumping (with syntax highlight on UI)
• Kaggle integration
• Hierarchical logging
• Grid search
• Experiments comparison
• Customizing layout system

Contents

• Screenshots

• Installation

• UI

• Usage

• Docs and examples

• Environment variables

Screenshots

Dags

Computers

Reports

Code

Graph

More screenshots

Installation

1. Install MLComp package

   sudo apt-get install -y \
   libavformat-dev libavcodec-dev libavdevice-dev \
   libavutil-dev libswscale-dev libavresample-dev libavfilter-dev
   
   pip install mlcomp
   mlcomp init
   mlcomp migrate

2. Setup your environment. Please consider Environment variables section

3. Run db, redis, mlcomp-server, mlcomp-workers:

   Variant 1: minimal (if you have 1 computer)

   Run all necessary (mlcomp-server, mlcomp-workers, redis-server), it uses SQLITE:

   mlcomp-server start --daemon=True

   Variant 2: full

   a. Change your Environment variables to use PostgreSql

   b. Install rsync on each work computer

   sudo apt-get install rsync

   Ensure that every computer is available by SSH protocol with IP/PORT you specified in the Environment variables file.

   rsync will perform the following commands:

   to upload

   rsync -vhru -e "ssh -p {target.port} -o StrictHostKeyChecking=no" \
   {folder}/ {target.user}@{target.ip}:{folder}/ --perms  --chmod=777

   to download

   rsync -vhru -e "ssh -p {source.port} -o StrictHostKeyChecking=no" \
   {source.user}@{source.ip}:{folder}/ {folder}/ --perms  --chmod=777

   c. Install apex for distributed learning

   d. To Run postgresql, redis-server, mlcomp-server, execute on your server-computer:

   cd ~/mlcomp/configs/
   docker-compose -f server-compose.yml up -d

   e. Run on each worker-computer:

   mlcomp-worker start

UI

Web site is available at http://{WEB_HOST}:{WEB_PORT}

By default, it is http://localhost:4201

The front is built with AngularJS.

In case you desire to change it, please consider front's Readme page

Usage

Run

mlcomp dag PATH_TO_CONFIG.yml

This command copies files of the directory to the database.

Then, the server schedules the DAG considering free resources.

For more information, please consider Docs

Docs and examples

API documentation and an overview of the library can be found here

You can find advanced tutorials and MLComp best practices in the examples folder of the repository.

FileSync tutorial describes data synchronization mechanism

Environment variables

The single file to setup your computer environment is located at ~/mlcomp/configs/.env

• ROOT_FOLDER - folder to save MLComp files: configs, db, tasks, etc.
• TOKEN - site security token. Please change it to any string
• DB_TYPE. Either SQLITE or POSTGRESQL
• POSTGRES_DB. PostgreSql db name
• POSTGRES_USER. PostgreSql user
• POSTGRES_PASSWORD. PostgreSql password
• POSTGRES_HOST. PostgreSql host
• PGDATA. PostgreSql db files location
• REDIS_HOST. Redis host
• REDIS_PORT. Redis port
• REDIS_PASSWORD. Redis password
• WEB_HOST. MLComp site host. 0.0.0.0 means it is available from everywhere
• WEB_PORT. MLComp site port
• CONSOLE_LOG_LEVEL. log level for output to the console
• DB_LOG_LEVEL. log level for output to the database
• IP. Ip of a work computer. The work computer must be accessible from other work computers by these IP/PORT
• PORT. Port of a work computer. The work computer must be accessible from other work computers by these IP/PORT (SSH protocol)
• MASTER_PORT_RANGE. distributed port range for a work computer. 29500-29510 means that if this work computer is a master in a distributed learning, it will use the first free port from this range. Ranges of different work computers must not overlap.
• NCCL_SOCKET_IFNAME. NCCL network interface.
• FILE_SYNC_INTERVAL. File sync interval in seconds. 0 means file sync is off
• WORKER_USAGE_INTERVAL. Interval in seconds of writing worker usage to DB
• INSTALL_DEPENDENCIES. True/False. Either install dependent libraries or not
• SYNC_WITH_THIS_COMPUTER. True/False. If False, all computers except that will not sync with that one
• CAN_PROCESS_TASKS. True/False. If false, this computer does not process tasks

You can see your network interfaces with ifconfig command. Please consider nvidia doc