Show HN: Continuous Machine Learning – CI/CD for Machine Learning Projects | Hacker News Reader

Show HN: Continuous Machine Learning – CI/CD for Machine Learning Projects

cml.dev

171 pointsrhythmvertigo38 comments

Hi, I'm one of the project creators. Continuous Machine Learning (CML) is an open source project to help ML projects use CI/CD with Github Actions and Gitlab CI (https://github.com/iterative/cml).

CML automatically generates human-readable reports with metrics and data viz in every pull/merge request, and helps you use storage and GPU/CPU resources from cloud services. CML addresses three hurdles for making ML compatible with CI:

1. In ML, pass/fail tests aren’t enough. Understanding model performance might require data visualizations and detailed metric reports. CML automatically generates custom reports after every CI run with visual elements like tables and graphs. You can even get a Tensorboard.dev link as part of your report.

2. Dataset changes need to trigger feedback just like source code. CML works with DVC so dataset changes trigger automatic training and testing.

3.Hardware for ML is an ecosystem in itself. We’ve developed use cases with CML and Docker Machine to automatically provision and deploy cloud compute instances (CPU & GPU) for model training.

Our philosophy is that ML projects- and MLOps practices- should be built on top of traditional software tools and CI systems, and not as a separate platform. Our goal is to extend DevOps’ wins from software development to ML. Check out our project site (https://cml.dev) and repo, and please let us know what you think!

This is really cool. We've been recommending DVC to our users for a long time, and this looks like a natural step forward for the Iterative ecosystem.

Yeah, we were seeing a lot of users organically asking about CI/CD. DVC had to come first to address some of the data management issues, which seemed like the biggest hurdle to CI. But we're excited to add this now.

This looks really amazing. Do you have any plans to support on-prem GPUs? If we had Gitlab runners with GPUs in them would this project pick it up and use those runners for training/analysis?

Yep, it will work with your GPUs! You just set up self-hosted runners as usual https://docs.gitlab.com/runner/

GitHub & GitLab have both made it quite easy to use your own resources as runners. I recently met someone who was doing Actions with a Jetson Nano on their dresser :)

That's really cool. I might have to play around with this. Do you have any docs on what you do to deploy a model? Something I've been doing at work is dealing with the output of some ML code we have. We end up with ~150GB of data that needs to be synced to a file share in prod. I'm assuming DVC can be used for this.

After the run, output files, upload as a data set in DVC or something?

Documenting this full workflow would save a lot of confused devopsy people (like myself) survive in the world of ML. Thanks for this hard work you've all put into this!

150Gb ML model file? That's cool!

Yes, DVC can help with that. Where the data lives? S3/GCS or just a server with SSH?

Disclaimer: I'm a creator of DVC.

They live in an Azure Storage Account (Azure's not-S3). They're mounted as a network storage into pods in our kube cluster. All we need to do to deploy is copy the data into `/<thing>/<timestamp>` and the code finds the newest version and loads it up. So really I'm just looking for a way to abstract azure blob storage from my ML people and to allow them to do the equivalent of `docker tag...` to choose what we roll out.

I think DVC (+CML) is a good solution for this. It "wraps" artifacts that you store into Git. And Git repo abstracts access to the cloud. In your case of the mounted storage it will look like `git pull` + `dvc checkout` after model is "merged" in the production/master branch.

CML can automate and make the process of preparing the model to be merged into that branch reliable, visible, robust, etc.

I'm happy to help with this flow, ping me on Twitter - @shcheklein in DM or ivan on DVC Discord.

Not to be too self-promotional here, but I'm a maintainer of Cortex, a model deployment platform that sounds like it might be useful: https://github.com/cortexlabs/cortex

With DVC/Cortex, you can set things up so that all you have to do is run `dvc push` to update your model and `cortex deploy` to deploy it.

You can also deploy runners with GPU on premise using CML docker image with GPU already supported having to install only the ndivia drivers and nvidia-docker in your machine.

docker run --name myrunner -d -e RUNNER_IDLE_TIMEOUT=1800 -e RUNNER_LABELS=cml -e RUNNER_REPO=$my_repo_url -e repo_token=$my_repo_token dvcorg/cml-gpu-py3-cloud-runner

It works for Gitlab and Github. Just only point your url and repo token

Deployment, inference and management can participate in this as well!

Here is the missing part for a total e2e solution: https://github.com/marketplace/actions/algorithmia-ci-cd

{disclaimer, we built this Github action}

coooool! going to try this out :)

Hi doppenhe, we have that part already implemented using cml-send-github-check and dvc metrics diff. You can compare the metric that you prefer with dvc and then just set the status of the github check uploading your full report. Of course, you can also fail the workflow as your Github action does, but I think is more useful to see it as a report in the check.

disclaimer: I'm work with CML

Hey! Disclaimer - I'm one of the DVC maintainers :) Super excited for the team on this release!

For the last two years we have seen over and over again how our users take DVC and use it inside Gitlab, Github, etc. This product was born partially as a result of these discussions, partially as an initial visions for the ML tools ecosystem - Hashicorp-like.

Having A software engineering background I really hope that integrating ML workflow into engineering tools will be the future of this space. And with CML and other tools (e.g. https://github.blog/2020-06-17-using-github-actions-for-mlop...) we see this happening.

Very cool to see this!

I've added integrating CML closer into GitLab to our direction page with https://gitlab.com/gitlab-com/www-gitlab-com/-/merge_request...

If you want to show the results in the Merge Request itself like we do for things like test results and security scans please let us know. Open an issue and if there is no response my Twitter handle is @sytses

Hey Sytse! Thanks for the kind words. We're very interested in a deep integration with Gitlab :) Can you share some examples of these security scan reports, please? Right now, we return CML reports as comments in Merge Requests like this: https://gitlab.com/iterative.ai/cml-cloud-case/-/merge_reque.... We'd appreciate any tips or suggestions.

For some examples, you can see these designs for including feedback on security scans directly in an MR: https://gitlab.com/gitlab-org/gitlab/-/issues/12896/designs

Our public documentation on these reports and how you can integrate with them is at https://docs.gitlab.com/ee/development/integrations/secure/ in our documentation if you want to check it out! tl;dr - we have a JSON format for security results that if you format your results in, it will show up in the various interfaces in the product automatically.

Awesome! Other replies to your comment already gave suggestions. Consider asking them for a call to give in if you need more information.

Thanks! we'll check it out and be in touch.

Very interesting, I've been looking for something like this to add to our ML pipeline. a few questions:

1) can we see examples of generated reports?

2) what happens if training fails?

3) what kind of metrics can it graph? can we have it track our custom metrics?

4) can we connect with external services like with webhooks,slack, and other integrations.

5) is this a docker technology, or how does it deal with images and dependencies?

Great work!

Thanks, and good questions!

1. Yes! Let me link some reports and example repos:

- A basic classification problem with scikit learn: https://github.com/iterative/cml_base_case/pull/2

- CML with DVC & Vega-Lite graphs: https://github.com/iterative/cml_dvc_case/pull/4

- Neural style transfer with EC2 GPU: https://github.com/iterative/cml_cloud_case/pull/2

2. If training fails, you'll be notified that your run failed in the GitHub Action dashboard (or GitLab CI/CD dashboard). See here for some real life examples of failure ;) : https://github.com/iterative/cml_cloud_case/actions

3. CML reports are markdown documents, so you can write any kind of text to them. If your metrics are output in a file `metrics.txt`, you can have your runner execute `cat metrics.txt >> report.md` and then have CML pass on the report to GitHub/GitLab. Likewise, any graphing library is supported because you can add standard image files (.png, .jpg) to the report. So custom metrics and custom graphs. We like DVC for managing and plotting metrics, but we're biased because we also maintain it.

4. Yep, GitHub Actions is pretty powerful and flexible. Works with whatever external services you can connect to your Action!

5. It's not strictly a Docker technology. We use Docker images preinstalled with the CML library in our examples, but you can just install the library with npm in your own image. https://github.com/iterative/cml#using-your-own-docker-image

Let me know if there's anything else I can tell you about

I think taking the approach of "help[ing] your team make informed, data-driven decisions" through generating reports is valuable here. In my opinion, it goes too far if we start continuously deploying ML code like it's a SWE project. To take an example in the case of autonomous vehicles, pushing continuous updates to perception modules without thoroughly exploring the ramifications of an update could be potentially catastrophic.

Obviously we can't predict every error by thinking hard, but datasets will never serve as a full representation of what models might experience in the real world. Continuous deployment to an ML model could affect undefined behavior in unpredictable ways.

Definitely- automatic model deployment would only work for a subset of applications. We think a lot more about having models available as candidates for deployment, which can then be inspected by domain experts on a team.

One of our motivations for building visual reports that appear like comments in a pull request is giving teams metrics & info to discuss when deciding if merge is right. That way, the automated part is training and testing, but the decision making is human (i.e., data scientists whose skills are better used interpreting models & data than running repetitive training scripts).

Agreed! I think, it's similar to software engineering. It's not that often that large websites are being deployed to prod without someone approving/QA-ing it manually first, right?

CI/CD systems in this case help automating this as much as possible, but do not completely replace decision making process, I would say.

ML is a relatively young field, and decades behind Software Engineering in terms of best practices for running production systems. CI/CD massively improved the innovation cycle time and quality of production software, and I believe it is key for building robust production ML systems as well. CML looks like a really easy to use product for bringing CI/CD to ML projects.

Thanks for the kind words! DevOps has had profound results and we think culturally and technically, it's the right moment to push for those practices in ML.

Just something I wanted to share which might be of interest on this side of things.

The CD Foundation has a SIG around MLOps which is pretty active and has some awesome folk participating.

For anyone who's interested in this space, there's some more detail here: https://cd.foundation/blog/2020/02/11/announcing-the-cd-foun...

Awesome to see a github native workflow for CI/CD in the ML space! This team is closet I seen that's like Hashicorp for ML

Thanks! Hashicorp is our inspiration :D

Very exciting! I used this team's previous product (DVC) at a research lab at Caltech. This looks like a very useful tool.

Out of curiosity, what kind of research? I'm very interested in getting DVC and Git more broadly into academic research circles, but finding a lot of barriers in my home field.

Longtime DVC user here - this is going to be so helpful. We use DVC for all of our model and data versioning, but what's been missing is the ability to cleanly integrate that into our CI workflow. Looks like that's solved now! The cml.yaml syntax also looks quite nice, very easy to follow. Looking forward to trying this out.

Really glad to hear that, rkaplan. Let us know if we can be of any help!

CML looks really awesome. I've been on one of your online meetups. Are you planning to host more in the future? It would be great to learn real production use cases!

Yes, we're aiming to have another at the end of July! If you have a particular use case you're interested in, we'd love to know. We might be able to develop some materials around it.

Btw. Here's a recording of the first DVC meetup https://www.youtube.com/watch?v=19GMtrFykSU I like it a lot. Highly recommended!

Reply on news.ycombinator.com