More than a framework

Machine Learning challenges

• Deployment and automation
• Reproducibility of models and predictions
• Diagnostics
• Scalability
• Collaboration
• Business uses
• Monitoring and management

MLops ecosystem powered by open source

Convenient DL serving

Experiments monitoring

Accelerated DL & RL

Train

Analyze

Deploy

Catalyst.Ecosystem

• Deep learning experiments standardisation
• Autoscaling from single GPU to SLURM cluster
• Experiments tracking
• Teamwork organisation, collaboration
• Community support

Catalyst.Ecosystem

• Reduces development costs
• Gives data scientists a tool for collaboration
• Finds useful insights from experiments
• Connects with the world-wide community

Eugene Kachan

Sergey Kolesnikov

David Kuryakin

​

Contributors�Researchers, Engineers

Artem Zolkin

Team

Friends

• Collaborations with top USA and RU universities on DL projects
• Several success ML cases with startups and companies
• Connections with top russian industrial researchers and community

catalyst-team/catalyst

pip install catalyst

• Fully reproducible pipelines
• CV, NLP, GANs and RecSys support
• Distributed and half-precision training
• Extra subpackages for �CV and NLP

catalyst-team/alchemy

pip install alchemy

• Experiment monitoring
• Hyperparameters storage
• Search, compare, and visualize training runs

​

alchemy.host

catalyst-team/reaction

pip install reaction

• DL models serving
• Load balancing
• Sync/Async API
• Batch mode handling
• Monitoring
• Swagger autodocs
• Telegram integration

PyTorch.Ecosystem

Tutorials

• Custom train loop
• CV, NLP, GANs and RecSys demos
• Classification tutorial
• Segmentation tutorial
• jupyter TensorBoard support
• FP16 with Nvidia Apex
• TTA usage
• And more: FocalLoss, OneCycle, Lookahead, RAdam, Ranger

minimal examples

catalyst-team/catalyst

pip install catalyst

@catalyst-slack

@catalyst_team

Sergey Kolesnikov, scitator@gmail.com

Extra

Catalyst.Reproducibility = 🖤

Catalyst has full DL & RL convergence tests for

DL, GANs and RL.

Catalyst.DL

Catalyst.Reproducibility = 🖤

DL pipelines reproducibility over different machines

Catalyst vs for-loop approach

VS

200+ lines!

Catalyst.RL

Catalyst.RL

• major DRL �on-policy (REINFORCE, PPO) �and off-policy algorithms �(DQN, DDPG, SAC, TD3)
• distributional value functions, n-step returns, �hyperbolical gammas, etc
• Trainers and samplers ensembling
• Distributed training support
• Open Source, PyTorch

Catalyst.RL

Comparison of catalyst.RL with a number of RL frameworks. �For a more detailed up-to-date table please check this Google Sheet.

Catalyst.Pipelines

Catalyst.Classification

• Resnet finetuning
• Cadene integration
• Augmentations prediction

Catalyst.Autolabel

Pseudo-Labeling is all you need

Catalyst.Detection

Anchor-free �object detection

Catalyst.Segmentation (semantic)

• Unet, ResnetUnet
• Linknet, ResnetLinknet
• FPNUnet, ResnetFPNUnet
• DSBowl18 – Watershed

Catalyst.Segmentation (instance)

• Unet, ResnetUnet
• Linknet, ResnetLinknet
• FPNUnet, ResnetFPNUnet
• DSBowl18 – Watershed

Catalyst.Video

Temporal Segmentation Networks

Catalyst.GAN

• Vanilla GAN
• Wasserstein GAN
• Wasserstein GAN GP
• Conditional GAN

Catalyst.NLP

• DistilBert pipeline
• Named-entity recognition
• Token classification

ReAction

• PyTorch models serving
• Load balancing
• Sync/Async API
• Batch mode handling
• Monitoring
• Swagger autodocs
• Telegram integration

Competitions

Catalyst.Competitions

• 3rd place solution of Task 3: Organ-at-risk segmentation from chest CT scans �at MICCAI 2019
• 4th place solution of Task 4: Gross Target Volume segmentation of lung cancer �at MICCAI 2019
• 3rd rank at DIBCO 2019

Bac Nguyen Xuan�Kaggle Competitions Master

Catalyst.Competitions – NeurIPS’18, 3rd place

Catalyst.Competitions – NeurIPS’19, 2nd place