TorchOK
PIPELINE
FOR TESTING
NEURAL NETWORKS
In 2019,
EORA began
developing a pipeline for the development and training of neural network models with computer vision (CV) technology.
View projects
In three years, about
for large companies have been done with its help.
30 projects
It was named
the latest version of the pipeline is written on the Pytorch Lighting engine.
TorchOK:
All projects use image and video processing:
face recognition
search for similar images
segmentation
classification
For whom
Now the pipeline is available to our customers:
Outsource-developers
in the field of Data Science and CV
that don't have their own pipeline
Product teams
in the field of Deep Learning
Research groups
TorchOK pipeline features
Extensive library of computer models: from classic ResNet to the latest Swin transformer;
The wide selection of ready-made datasets: it is only necessary to prepare your data in the required format (CSV file with annotations and paths to images);
Modern infrastructure. TorchOK runs on machines with CPU, GPU, and also on multiple computers with multiple GPUs. There is support for TPU;
Metrics for evaluating computer vision models in TensorBoard and MLflow: classification, segmentation, metrics for finding similar images, for face recognition;
The single interface for loading and unloading computer models;
Convenient "packaging": TorchOK can be run through the Conda environment and in the cloud — using Docker containerizer or in Safe Maker on Amazon Web Services.
Our task is to bring “raw” computer models to production. Therefore, TorchOK will always have only the best technical solutions. Not the most accurate in terms of public benchmarks, but the most effective in terms of quality/speed.
Technical Director of EORA Data Lab
VLAD VINOGRADOV
Advantages of TorchOK
FOR DEVELOPERS
TorchOK can be cloned to work in several projects at the same time, for separate development teams. The errors and improvements found are eliminated and added to a single repository and are visible to everyone.
It is suitable for collaboration
You choose the training parameters of the neural network model: which network to use, which data to load, which metrics to count, how many GPUs you need.
It is easily customized
Usually teams write their own pipeline for each project. We use the TorchOK codebase as the basis, and then modify it within the team.
It is universal
Minimal knowledge in Deep Learning is enough to work in pipeline: prepare the dataset for training, train the neural network model and then transfer it for integration.
It's simple
FOR BUSINESS PROJECTS
Saving money: you have the finished result of many years of EORA work in your hands
Saving time: you can immediately start training your neural network
The most modern solutions in a simple "package"
Constantly updated database of neural network models
Only the best: we save that really works
Quality assurance
Selection of the best technical solutions and programs
Respected by the research community
The current version is the third
Yearly improvements since 2019
Team
Development of critical library blocks, integration of neural network architectures
Library core development, packaging, management
Correction of critical errors
Vyacheslav Shultz
Vladislav Vinogradov
Roman Bogachev
Konstantin Kubrak
Aelita Shaikhutdinova
Adding of the new task types support
Rashid Bayazitov
Upcoming product updates
Try
Transition to permanent support of key neural networks while maintaining the same type of structure: backbone, neck, head, hat
The ability to upload models to different frameworks: TensorRT, ONNX, OpenVINO
Ensuring compatibility of different frameworks and layers of neural network models
Selection of models in terms of quality and speed
Contact us