Thomas Lefebvre

My friday night prompt to codex: "i refactored and now it's broken. help"

Making progress..slowly but surely. The model needs more training but here's a friday evening end to end pipeline.

I may have an acute overfitting case here...

Decided to fine-tune YOLO. I'm annotating data...how's your saturday?

I find them great to reason with to find solutions to problems you don't already know how to solve, or for extensive boilerplate/grunt work you'd want to get done. If you already know what to do and it's relatively fast to do so, I find it takes more time to write it all up to an LLM.

I'm trying Molmo and this is impressive

Overwriting the overridable overridden instructions.

Some progress...

I haven't gotten good result so far trying to detect pieces on squares using more traditional computer vision so I'm trying out DINO.

Stanford CS231N Deep Learning for Computer Vision I 2025 - YouTube Computer Vision has become ubiquitous in our society, with applications in search, image understanding, apps, mapping, medicine, drones, and self-driving car...

Just wrapped up CS231N. Definitely worth the time investment if you're willing to do it. Took me about 4 months to complete at 1 lecture/week.

www.youtube.com/playlist?lis...

We just spend many hours uploading vectors to S3 vectors bucket..only to realize the keys weren't unique and each batch did overwrite the previous one. Ended up with a bucket with 500 vectors at the end of the day 😬

Embed custom operations in ONNX graphs | Thomas Lefebvre Learn how to update an ONNX graph and add custom operations to create a self-contained model.

Winter is still around and I get cozy with ONNX.
Recently, I had to change a pre-trained model output and wrote a bit about it. My future self will thank me.

tlbvr.com/blog/onnx-gr...

We just named one of our service after this magical object. For all the Zelda fans out there.

is Macron...hungover?

Finally got it working on live camera. It's still supper laggy when I draw onto the frame but the most important part is the homography works in various angles. Still need to test how it behaves in various brightness settings.

I applied the same heuristics to another board where I printed the actual aruco markers. Works well! Defining the playable squares by hand is a bit tedious, but I'm sure we can streamline this.

And it's live! tlbvr.com/blog/categor...

Screenshot with a sneak peek of the second part of the blog series.

I'm finally writing the second part. Never too late.

Your phone listens.

Functional python snippet

Woaw. I'm finding my TS roots again (this is python)

On to detecting pieces — I'm not sure yet how I'm going to avoid false positive because of light/shadows.

Original board with squares mapped out.

Skewed board 1 retaining original mapping

Skewed board 2 retaining original mapping

Building a way to map game boards and make it resilient to jitter/skewing. Goal is to feed video frames to this algorithm and keep the mapping stable.

Detect label issues in your vision dataset | Thomas Lefebvre Leverage cross-validation and identify potential issues in your labeled data for your computer vision model

Using k-fold cross-validation has proven useful to find label issues in our dataset. I write a little bit about it here tlbvr.com/blog/k-fold-...

I wish python `match` would be an expression. There I said it.

Just started watching Stanford's computer vision lectures from this year www.youtube.com/watch?v=2fq9...

see you on the order side.

Does "reporting junk" do anything on iMessage? My phone number probably hung out in dark places and now i'm flooded with spam. Keep reporting but I don't feel like it does anything.

Lowering the loss of your model feels like a game. You initially take bets and get rewarded (maybe).

It's that time of year again when I can stick up my feet close to my gaming computer and it keeps them warm.

Similar to: why do I always see my queen blunders after I made the move.

Latency for served model REST APIs · Issue #7948 · mlflow/mlflow Willingness to contribute No. I cannot contribute this feature at this time. Proposal Summary For certain requirements, the latency of the model REST APIs have to be very low. We have observed that...

Going from 30s latency to ~400ms in model invocations by dropping the mlflow provided scoring server by a custom one. I don't understand

github.com/mlflow/mlflo...