Skip to content

Remove dead ASR task and fix object-detection UI conditionals - #174

Open
Alvin-Nahabwe wants to merge 1 commit into
fix/dl-client-contractfrom
chore/remove-asr-fix-dl-ui
Open

Remove dead ASR task and fix object-detection UI conditionals#174
Alvin-Nahabwe wants to merge 1 commit into
fix/dl-client-contractfrom
chore/remove-asr-fix-dl-ui

Conversation

@Alvin-Nahabwe

Copy link
Copy Markdown
Contributor

Stacked on #173 (fix/dl-client-contract) — both branches touch server/deep_learning.R. Review/merge #173 first; this PR's base will retarget to main automatically.

Remove the ASR task

ASR was fully wired in the client but targets an API surface that no longer exists: no /train/asr endpoint, no asr key in the model registry, no ASR inference script. Selecting ASR populated nothing; submitting failed.

Removed the training panel, inference panel, checkpoint finder, job submission, result outputs, and the four dropdown entries. grep -i asr across the repo now returns nothing.

Fix two object-detection conditionals

Both are real bugs, independent of ASR:

  • RT-DETR was treated as a Transformers model. The API routes it to the Ultralytics trainer (is_ultralytics_model matches rtdetr), but the UI showed it the Transformers hyperparameters and hid the YOLO ones — so lr0, momentum, warmup_epochs etc. were unreachable for RT-DETR runs. Both panels now key off YOLO || RT-DETR.
  • Inference IoU / max-detections were hardcoded to yolo11. They were hidden for yolov12, yolo26, and rtdetr checkpoints, all of which the API accepts them for. The condition now matches any yolo* checkpoint except yolos (a Transformers model), plus rtdetr.

Cleanup

Removed 281 lines of commented-out legacy UI at the foot of deeplearning_ui.R. Every removed line was verified to be a comment or blank before deletion; it remains in git history.

Net: +17 / −538.

Verification

parse() passes on both changed files. The app was not run — no Shiny server available. Please click through the DL tab (task switching, RT-DETR arch selection, inference panel) before merging.

Open question

The Hub & Logging panels still expose push_to_hub, hub_user_id, log_to_wandb, wandb_project, wandb_entity for all three tasks. The API ignores every one of them, and all training scripts set WANDB_DISABLED=true. These controls silently do nothing. I've left them in place pending a decision: finish the W&B integration, or remove the panels?

🤖 Generated with Claude Code

The ASR task was fully wired in the client but targets an API surface
that no longer exists: there is no /train/asr endpoint, no `asr` key in
the model registry, and no ASR inference script. Selecting ASR populated
nothing and submitting failed. Removed the training panel, the inference
panel, the checkpoint finder, the job submission, the result outputs, and
the four dropdown entries. ASR now appears nowhere in the app.

Two object-detection conditionals were also wrong:

- RT-DETR is an Ultralytics model (the API routes it to the YOLO trainer
  via `is_ultralytics_model`), but the UI showed it the Transformers
  hyperparameters and hid the YOLO ones. Both panels now key off YOLO or
  RT-DETR.
- The inference IoU and max-detections controls were shown only when the
  checkpoint name contained 'yolo11', hiding them for yolov12, yolo26 and
  rtdetr checkpoints. The condition now matches any yolo* checkpoint
  except YOLOS (a Transformers model), plus rtdetr.

Finally, removed 281 lines of commented-out legacy UI at the foot of
deeplearning_ui.R, after verifying every removed line was a comment or
blank. It is preserved in git history.

Verified with parse() on both files. The app was not run.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant