"All the Deep-Learning DevOps your research needs, and then some... Because ain't nobody got time for that"
TRAINS Agent is an AI experiment cluster solution.
It is a zero configuration fire-and-forget execution agent, which combined with trains-server provides a full AI cluster solution.
Full AutoML in 5 steps
- Install the TRAINS server (or use our open server)
pip install trains_agent(install the TRAINS agent on any GPU machine: on-premises / cloud / ...)- Add TRAINS to your code with just 2 lines & run it once (on your machine / laptop)
- Change the parameters in the UI & schedule for execution (or automate with an AutoML pipeline)
- 📉 📈 👀 🍺
Using the TRAINS agent, you can now set up a dynamic cluster with *epsilon DevOps
*epsilon - Because we are scientists 📐 and nothing is really zero work
(Experience TRAINS live at https://demoapp.trains.allegro.ai)
The TRAINS Agent was built to address the DL/ML R&D DevOps needs:
- Easily add & remove machines from the cluster
- Reuse machines without the need for any dedicated containers or images
- Combine GPU resources across any cloud and on-prem
- No need for yaml/json/template configuration of any kind
- User friendly UI
- Manageable resource allocation that can be used by researchers and engineers
- Flexible and controllable scheduler with priority support
- Automatic instance spinning in the cloud (coming soon)
We think Kubernetes is awesome.
Combined with KubeFlow it is a robust solution for production-grade DevOps.
We've observed, however, that it can be a bit of an overkill as an R&D DL/ML solution.
If you are considering K8S for your research, also consider that you will soon be managing hundreds of containers...
In our experience, handling and building the environments, having to package every experiment in a docker, managing those hundreds (or more) containers and building pipelines on top of it all, is very complicated (also, it’s usually out of scope for the research team, and overwhelming even for the DevOps team).
We feel there has to be a better way, that can be just as powerful for R&D and at the same time allow integration with K8S when the need arises.
(If you already have a K8S cluster for AI, detailed instructions on how to integrate TRAINS into your K8S cluster are coming soon.)
Full scale HPC with a click of a button
TRAINS Agent is a job scheduler that listens on job queue(s), pulls jobs, sets the job environments, executes the job and monitors its progress.
Any 'Draft' experiment can be scheduled for execution by a TRAINS agent.
A previously run experiment can be put into 'Draft' state by either of two methods:
- Using the 'Reset' action from the experiment right-click context menu in the TRAINS UI - This will clear any results and artifacts the previous run had created.
- Using the 'Clone' action from the experiment right-click context menu in the TRAINS UI - This will create a new 'Draft' experiment with the same configuration as the original experiment.
An experiment is scheduled for execution using the 'Enqueue' action from the experiment right-click context menu in the TRAINS UI and selecting the execution queue.
See creating an experiment and enqueuing it for execution.
Once an experiment is enqueued, it will be picked up and executed by a TRAINS agent monitoring this queue.
The TRAINS UI Workers & Queues page provides ongoing execution information:
- Workers Tab: Monitor you cluster
- Review available resources
- Monitor machines statistics (CPU / GPU / Disk / Network)
- Queues Tab:
- Control the scheduling order of jobs
- Cancel or abort job execution
- Move jobs between execution queues
The TRAINS agent executes experiments using the following process:
- Create a new virtual environment (or launch the selected docker image)
- Clone the code into the virtual-environment (or inside the docker)
- Install python packages based on the package requirements listed for the experiment
- Special note for PyTorch: The TRAINS agent will automatically select the torch packages based on the CUDA_VERSION environment variable of the machine
- Execute the code, while monitoring the process
- Log all stdout/stderr in the TRAINS UI, including the cloning and installation process, for easy debugging
- Monitor the execution and allow you to manually abort the job using the TRAINS UI (or, in the unfortunate case of a code crash, catch the error and signal the experiment has failed)
+-----------------+
| GPU Machine |
Development Machine | |
+------------------------+ | +-------------+ |
| Data Scientist's | +--------------+ | |TRAINS Agent | |
| DL/ML Code | | WEB UI | | | | |
| | | | | | +---------+ | |
| | | | | | | DL/ML | | |
| | +--------------+ | | | Code | | |
| | User Clones Exp #1 / . . . . . . . / | | | | | |
| +-------------------+ | into Exp #2 / . . . . . . . / | | +---------+ | |
| | TRAINS | | +---------------/-_____________-/ | | | |
| +---------+---------+ | | | | ^ | |
+-----------|------------+ | | +------|------+ |
| | +--------|--------+
Auto-Magically | |
Creates Exp #1 | The TRAINS Agent
\ User Change Hyper-Parameters Pulls Exp #2, setup the
| | environment & clone code.
| | Start execution with the
+------------|------------+ | +--------------------+ new set of Hyper-Parameters.
| +---------v---------+ | | | TRAINS-SERVER | |
| | Experiment #1 | | | | | |
| +-------------------+ | | | Execution Queue | |
| || | | | | |
| +-------------------+<----------+ | | |
| | | | | | |
| | Experiment #2 | | | | |
| +-------------------<------------\ | | |
| | ------------->---------------+ | |
| | User Send Exp #2 | |Execute Exp #2 +--------------------+
| | For Execution | +---------------+ |
| TRAINS-SERVER | | |
+-------------------------+ +--------------------+
pip install trains_agentFull Interface and capabilities are available with
trains-agent --help
trains-agent daemon --helptrains-agent initNote: The TRAINS agent uses a cache folder to cache pip packages, apt packages and cloned repositories. The default TRAINS Agent cache folder is ~/.trains
See full details in your configuration file at ~/trains.conf
Note: The TRAINS agent extends the TRAINS configuration file ~/trains.conf
They are designed to share the same configuration file, see example here
For debug and experimentation, start the TRAINS agent in foreground mode, where all the output is printed to screen
trains-agent daemon --queue default --foregroundFor actual service mode, all the stdout will be stored automatically into a temporary file (no need to pipe)
trains-agent daemon --queue defaultGPU allocation is controlled via the standard OS environment NVIDIA_VISIBLE_DEVICES.
If NVIDIA_VISIBLE_DEVICES variable doesn't exist, all GPU's will be allocated for the trains-agent
If NVIDIA_VISIBLE_DEVICES is an empty string ("") No gpu will be allocated for the trains-agent
Example: spin two agents, one per gpu on the same machine:
NVIDIA_VISIBLE_DEVICES=0 trains-agent daemon --queue default &
NVIDIA_VISIBLE_DEVICES=1 trains-agent daemon --queue default &Example: spin two agents, with two gpu's per agent:
NVIDIA_VISIBLE_DEVICES=0,1 trains-agent daemon --queue default &
NVIDIA_VISIBLE_DEVICES=2,3 trains-agent daemon --queue default &For debug and experimentation, start the TRAINS agent in foreground mode, where all the output is printed to screen
trains-agent daemon --queue default --docker --foregroundFor actual service mode, all the stdout will be stored automatically into a file (no need to pipe)
trains-agent daemon --queue default --dockerExample: spin two agents, one per gpu on the same machine:
NVIDIA_VISIBLE_DEVICES=0 trains-agent daemon --queue default --docker &
NVIDIA_VISIBLE_DEVICES=1 trains-agent daemon --queue default --docker &Example: spin two agents, with two gpu's per agent:
NVIDIA_VISIBLE_DEVICES=0,1 trains-agent daemon --queue default --docker &
NVIDIA_VISIBLE_DEVICES=2,3 trains-agent daemon --queue default --docker &Priority Queues are also supported, example use case:
High priority queue: important_jobs Low priority queue: default
trains-agent daemon --queue important_jobs defaultThe TRAINS agent will first try to pull jobs from the important_jobs queue, only then it will fetch a job from the default queue.
-
Integrate TRAINS with your code
-
Execute the code on your machine (Manually / PyCharm / Jupyter Notebook)
-
As your code is running, TRAINS creates an experiment logging all the necessary execution information:
- Git repository link and commit ID (or an entire jupyter notebook)
- Git diff (we’re not saying you never commit and push, but still...)
- Python packages used by your code (including specific versions used)
- Hyper-Parameters
- Input Artifacts
You now have a 'template' of your experiment with everything required for automated execution
-
In the TRAINS UI, Right click on the experiment and select 'clone'. A copy of your experiment will be created.
-
You now have a new draft experiment cloned from your original experiment, feel free to edit it
- Change the Hyper-Parameters
- Switch to the latest code base of the repository
- Update package versions
- Select a specific docker image to run in (see docker execution mode section)
- Or simply change nothing to run the same experiment again...
-
Schedule the newly created experiment for execution: Right-click the experiment and select 'enqueue'
The TRAINS Agent can also be used to implement AutoML orchestration and Experiment Pipelines in conjunction with the TRAINS package.
Sample AutoML & Orchestration examples can be found in the TRAINS example/automl folder.
AutoML examples
- Toy Keras training experiment
- In order to create an experiment-template in the system, this code must be executed once manually
- Random Search over the above Keras experiment-template
- This example will create multiple copies of the Keras experiment-template, with different hyper-parameter combinations
Experiment Pipeline examples
- First step experiment
- This example will "process data", and once done, will launch a copy of the 'second step' experiment-template
- Second step experiment
- In order to create an experiment-template in the system, this code must be executed once manually