point model training support

[jveitchmichaelis]

Description

This PR adds training functionality for the default point model. Most of the code has been distilled from my experimental branch. It should replicate the existing checkpoint, and I'm running a quick test to make sure.

Changes are relatively small in scope - added the optimal transport loss code, fleshed out compute_losses for the model and fixed a small bug in point visualization where the VertexAnnotator wouldn't accept a palette (we instead convert points to "detections" with a radius and plot them as circles).

I've included a point_pretrain config which was used to train the current default point model on NEON data, replication should be possible with:

uv deepforest --config-name point_pretrain train

Related Issue(s)

Closes #809 as we now support training, visualization of predictions, simple unit tests + should also include multi-class support in theory but we've not tested it.

AI-Assisted Development

Claude code for assistance with merging the changes from the other branch.

• I used AI tools (e.g., GitHub Copilot, ChatGPT, etc.) in developing this PR
• I understand all the code I'm submitting
• I have reviewed and validated all AI-generated code

[codecov]

Codecov Report

❌ Patch coverage is 92.96875% with 18 lines in your changes missing coverage. Please review.
✅ Project coverage is 85.75%. Comparing base (d2851ec) to head (aaad58a).
⚠️ Report is 8 commits behind head on main.

Files with missing lines	Patch %	Lines
src/deepforest/losses/ot_loss.py	93.75%	7 Missing ⚠️
src/deepforest/models/treeformer.py	94.30%	7 Missing ⚠️
src/deepforest/main.py	42.85%	4 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #1396      +/-   ##
==========================================
- Coverage   86.61%   85.75%   -0.87%     
==========================================
  Files          26       28       +2     
  Lines        3736     4121     +385     
==========================================
+ Hits         3236     3534     +298     
- Misses        500      587      +87     

Flag	Coverage Δ
unittests	85.75% <92.96%> (-0.87%)	⬇️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[jveitchmichaelis]

This looks good enough for review. I’ll run some more extensive tests to see if there’s any other features that need to be pulled in, but the core losses are all here.

[bw4sz]

I had one tiny issue that isn't probably related to this PR, but just in case. If you let the trainer default to the csv logger it hits

ValueError: dict contains fields not in fieldnames

which basically means that the schema is set on trainer.start and it hits metrics it wasn't
Pre-log the metric with a value of 0 or float('nan') in your setup() hook or on the first batch, so the logger registers the column header from the very beginning.

def on_train_start(self):
    # Register the metric key so CSVLogger knows about it from step 1
    self.log("epoch level metric", 0.0)

I think the right thing to do is merge this PR and address this at the module level, its not specific to this PR.

[jveitchmichaelis]

Which metric is this triggered for? I’ll have a look if I missed something when I pulled these changes from the other branch.

[bw4sz]

All of the epoch level metrics, this is just the csv logger, the tensorboard and comet logger handle this gracefully.

train_sinkhorn_beta_abs_max_epoch, train_ot_loss_epoch, train_count_mae_epoch

I think this is a different PR that we have an object that captures the metric names and logs them to 0 or 'nan' on train start, since this would be true for any metric in any of the workflows.