Fix #999: Replace hardcoded paths in Australia notebook with get_data and tempdir

[proofbyhuman]

Description

This PR fixes #999 by removing hardcoded local user paths in the 'Working with deepforest data' documentation notebook (docs/user_guide/examples/Australia.ipynb).

Key changes:

• Replaced hardcoded paths (/Users/benweinstein/Downloads/Plot13Ortho.tif) with get_data("australia.tif").
• Replaced hardcoded output directories with cross-platform tempfile.gettempdir().
• Added required imports (os and tempfile) to code blocks.

Related Issue(s)

Closes #999

AI-Assisted Development

• 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

AI tools used (if applicable):
Gemini (Google DeepMind Agentic Assistant)

[vickysharma-prog]

Hi @proofbyhuman welcome to Deepforest, you have 10 PRs open (#1409–#1418). I read all ten diffs and their linked issues. Consolidating into one comment instead of ten threads.

The set isn't mergeable as a set

• Fix 889 separate batch sizes #1416 contains docs: improve checkpoint documentation and fix code snippet in user g… #1415's commit. Its commit list starts with docs: improve checkpoint documentation and fix code snippet in user g… and its diff duplicates the entire docs/user_guide/11_training.md change from docs: improve checkpoint documentation and fix code snippet in user g… #1415.
• feat: allow predict_file to accept pandas DataFrame instead of only C… #1409, feat: automatically log sample training images with annotations on tr… #1410, fix: automatically override num_workers=0 for TiledRaster prediction … #1412, Fix 889 separate batch sizes #1416 all edit overlapping regions of src/deepforest/main.py. feat: allow predict_file to accept pandas DataFrame instead of only C… #1409 rewrites predict_file's signature; Fix 889 separate batch sizes #1416 edits inside that same function. fix: automatically override num_workers=0 for TiledRaster prediction … #1412 and Fix 889 separate batch sizes #1416 both change predict_dataloader.
• test: assert release model eval scores on sample data #1413 adds a new predict_file(csv_file=...) call site — the keyword feat: allow predict_file to accept pandas DataFrame instead of only C… #1409 deprecates.

Whichever merges first conflicts with the rest.

CI hasn't run

The green checks are only pre-commit.ci and readthedocs. The test suite is gated on maintainer approval, which you asked for on #1412. No test in any of these ten has executed. #1409 also states the suite wasn't run locally.

Per-PR

#1410 — duplicates existing code. ImagesCallback.on_train_start at src/deepforest/callbacks.py:53 already logs sample training images at train start, same default of 5 (dataset_samples=5). The existing one guards on distributed.is_global_zero(trainer) (yours has no rank guard — every DDP rank would log), skips on fast_dev_run, and samples the actual dataset via dataset[idx], i.e. post-augmentation. Yours re-reads the raw CSV, so it shows pre-augmentation annotations and won't catch the transform/coordinate bugs the description cites. It's also always-on rather than opt-in. Suggest closing.

#1409 — call sites not updated. 11 sites still pass csv_file= as a keyword and would emit DeprecationWarning: src/deepforest/scripts/predict.py:55 (production code), tests/test_evaluate.py:24,50,315,379, tests/test_IoU.py:14, tests/test_main.py:433,1169, tests/profile_evaluate.py:15, tests/profile_predict_file.py:17, docs/user_guide/12_evaluation.md:201. Two more pass positionally (tests/test_main.py:441, docs/user_guide/16_prediction.md:133) — those won't warn, they silently rebind to input_file.

Also: root_dir now defaults to None, but the docstring's "if none, uses image_dir in config" fallback isn't implemented, so None fails.

#1416 — orphans config.batch_size. After this change nothing reads it, but it stays in src/deepforest/conf/config.yaml:8 and src/deepforest/conf/schema.py:148. Anyone setting it is silently ignored, no warning. Needs removal or explicit deprecation. #889's item 3 ("update the config doc") also isn't covered. Minor: bird.yaml keeps batch_size: 32 alongside the new keys, and there's an unrelated lambda reformat in evaluate().

#1415 — wrong issue. Closes #1380 is incorrect. #1380 asks for a HuggingFace model card for weecology/cropmodel-deadtrees; this edits docs/user_guide/11_training.md. Separately, the rewritten example calls m.create_trainer(fast_dev_run=True) then m.trainer.fit(m) without configuring config.train.csv_file, so it fails as written.

#1414 — fragile detection. Catches bare Exception and substring-matches "401" / "Unauthorized". huggingface_hub raises typed GatedRepoError / HfHubHTTPError with .response.status_code — catch those and check == 401. Substring matching false-positives on any error text containing "401" (commit hash, path, port). The test patches with a plain Exception("401 Client Error..."), so it validates the substring match against its own mock, not real HF behavior. Also #1401 is a CI failure on a gated repo; a better message doesn't fix it, so this likely doesn't close the issue.

#1413 — network-dependent pin. Pins 0.80/0.72 against a live download of weecology/deepforest-tree. Fails for everyone the moment the model is republished, including on unrelated PRs. Deriving from integer counts (44/55, 44/61) is right, but consider pinning a revision or adding a marker so it can be deselected. Also uses csv_file=, deprecated by your #1409.

#1412 — removes a deliberate error. Deletes the workers > 0 ValueError in predict_tile and replaces it with a silent auto-override. That error looks intentional; turning a loud failure into a quiet correction is a maintainer decision. Test also calls m.load_model("weecology/deepforest-tree") — network-dependent and slow.

#1417 — matches #1274's spec. inverse_transform is identical to RandomPadTo's, worth sharing rather than duplicating. apply_transform returns a different shape than its input — fine at the default p=1.0, but you expose p, and if Kornia applies to a subset of a batch at p<1.0 the shapes may not concat. Worth testing with p=0.5.

#1418 — valid. australia.tif does ship in src/deepforest/data/. But the notebook's stored cell outputs weren't regenerated, so the rendered docs page still shows results from the old Plot13Ortho.tif. #999 is about the page rendering incorrectly, so re-run the notebook before committing.

#1411 — fine. Bare print() matches the surrounding code (datasets/training.py:71). callbacks.py uses pl_module.print() (rank-aware) but that isn't available in a dataset constructor.

On the AI-assisted workflow

Please review the Ai given code before submitting
You disclosed AI use on all ten via the template and named the tool each time — that's correct and not the issue. What's showing is that the review step is being skipped: #1411's description contains a stray proofbyhuman/antigravity string, #1410's ends with verification.ification, #1415 targets the wrong issue, and #1410 rebuilds something already in callbacks.py. Those are all catchable in a minute of reading. The "I have reviewed and validated all AI-generated code" checkbox is the load-bearing part, not the disclosure.

The specific gap is that these tools don't survey the existing codebase before adding to it — grepping for an existing implementation before starting would have caught #1410, and probably the #1409/#1413/#1416 overlap too.

Suggestion

Pick 2–3, draft the rest until those land. #1417, #1418 and #1409 look fine. That gives maintainers something reviewable and lets later PRs build on a merged base instead of conflicting.

[proofbyhuman]

Hola Viki Sharma, estoy muy agradecido con la devolución, tendré en cuenta todas las sugerencias y actuaré en consecuencia muchas gracias . atte. Ramiro guevara El jue, 30 jul 2026 a las 13:16, vicky sharma ***@***.***>) escribió:

…

*vickysharma-prog* left a comment (weecology/DeepForest#1418) <#1418 (comment)> Hi @proofbyhuman <https://github.com/proofbyhuman> welcome to Deepforest, you have 10 PRs open (#1409 <#1409>–#1418 <#1418>). I read all ten diffs and their linked issues. Consolidating into one comment instead of ten threads. The set isn't mergeable as a set - *#1416 <#1416> contains #1415 <#1415>'s commit.* Its commit list starts with docs: improve checkpoint documentation and fix code snippet in user g… and its diff duplicates the entire docs/user_guide/11_training.md change from #1415 <#1415>. - *#1409 <#1409>, #1410 <#1410>, #1412 <#1412>, #1416 <#1416> all edit overlapping regions of src/deepforest/main.py.* #1409 <#1409> rewrites predict_file's signature; #1416 <#1416> edits inside that same function. #1412 <#1412> and #1416 <#1416> both change predict_dataloader. - *#1413 <#1413> adds a new predict_file(csv_file=...) call site* — the keyword #1409 <#1409> deprecates. Whichever merges first conflicts with the rest. CI hasn't run The green checks are only pre-commit.ci and readthedocs. The test suite is gated on maintainer approval, which you asked for on #1412 <#1412>. No test in any of these ten has executed. #1409 <#1409> also states the suite wasn't run locally. Per-PR *#1410 <#1410> — duplicates existing code.* ImagesCallback.on_train_start at src/deepforest/callbacks.py:53 already logs sample training images at train start, same default of 5 (dataset_samples=5). The existing one guards on distributed.is_global_zero(trainer) (yours has no rank guard — every DDP rank would log), skips on fast_dev_run, and samples the actual dataset via dataset[idx], i.e. post-augmentation. Yours re-reads the raw CSV, so it shows pre-augmentation annotations and won't catch the transform/coordinate bugs the description cites. It's also always-on rather than opt-in. Suggest closing. *#1409 <#1409> — call sites not updated.* 11 sites still pass csv_file= as a keyword and would emit DeprecationWarning: src/deepforest/scripts/predict.py:55 (production code), tests/test_evaluate.py:24,50,315,379, tests/test_IoU.py:14, tests/test_main.py:433,1169, tests/profile_evaluate.py:15, tests/profile_predict_file.py:17, docs/user_guide/12_evaluation.md:201. Two more pass positionally (tests/test_main.py:441, docs/user_guide/16_prediction.md:133) — those won't warn, they silently rebind to input_file. Also: root_dir now defaults to None, but the docstring's "if none, uses image_dir in config" fallback isn't implemented, so None fails. *#1416 <#1416> — orphans config.batch_size.* After this change nothing reads it, but it stays in src/deepforest/conf/config.yaml:8 and src/deepforest/conf/schema.py:148. Anyone setting it is silently ignored, no warning. Needs removal or explicit deprecation. #889 <#889>'s item 3 ("update the config doc") also isn't covered. Minor: bird.yaml keeps batch_size: 32 alongside the new keys, and there's an unrelated lambda reformat in evaluate(). *#1415 <#1415> — wrong issue.* Closes #1380 is incorrect. #1380 <#1380> asks for a HuggingFace model card for weecology/cropmodel-deadtrees; this edits docs/user_guide/11_training.md. Separately, the rewritten example calls m.create_trainer(fast_dev_run=True) then m.trainer.fit(m) without configuring config.train.csv_file, so it fails as written. *#1414 <#1414> — fragile detection.* Catches bare Exception and substring-matches "401" / "Unauthorized". huggingface_hub raises typed GatedRepoError / HfHubHTTPError with .response.status_code — catch those and check == 401. Substring matching false-positives on any error text containing "401" (commit hash, path, port). The test patches with a plain Exception("401 Client Error..."), so it validates the substring match against its own mock, not real HF behavior. Also #1401 <#1401> is a CI failure on a gated repo; a better message doesn't fix it, so this likely doesn't close the issue. *#1413 <#1413> — network-dependent pin.* Pins 0.80/0.72 against a live download of weecology/deepforest-tree. Fails for everyone the moment the model is republished, including on unrelated PRs. Deriving from integer counts (44/55, 44/61) is right, but consider pinning a revision or adding a marker so it can be deselected. Also uses csv_file=, deprecated by your #1409 <#1409>. *#1412 <#1412> — removes a deliberate error.* Deletes the workers > 0 ValueError in predict_tile and replaces it with a silent auto-override. That error looks intentional; turning a loud failure into a quiet correction is a maintainer decision. Test also calls m.load_model("weecology/deepforest-tree") — network-dependent and slow. *#1417 <#1417> — matches #1274 <#1274>'s spec.* inverse_transform is identical to RandomPadTo's, worth sharing rather than duplicating. apply_transform returns a different shape than its input — fine at the default p=1.0, but you expose p, and if Kornia applies to a subset of a batch at p<1.0 the shapes may not concat. Worth testing with p=0.5. *#1418 <#1418> — valid.* australia.tif does ship in src/deepforest/data/. But the notebook's stored cell outputs weren't regenerated, so the rendered docs page still shows results from the old Plot13Ortho.tif. #999 <#999> is about the page rendering incorrectly, so re-run the notebook before committing. *#1411 <#1411> — fine.* Bare print() matches the surrounding code (datasets/training.py:71). callbacks.py uses pl_module.print() (rank-aware) but that isn't available in a dataset constructor. On the AI-assisted workflow Please review the Ai given code before submitting You disclosed AI use on all ten via the template and named the tool each time — that's correct and not the issue. What's showing is that the review step is being skipped: #1411 <#1411>'s description contains a stray proofbyhuman/antigravity string, #1410 <#1410>'s ends with verification.ification, #1415 <#1415> targets the wrong issue, and #1410 <#1410> rebuilds something already in callbacks.py. Those are all catchable in a minute of reading. The "I have reviewed and validated all AI-generated code" checkbox is the load-bearing part, not the disclosure. The specific gap is that these tools don't survey the existing codebase before adding to it — grepping for an existing implementation before starting would have caught #1410 <#1410>, and probably the #1409 <#1409>/#1413 <#1413>/#1416 <#1416> overlap too. Suggestion Pick 2–3, draft the rest until those land. #1417 <#1417>, #1418 <#1418> and #1409 <#1409> look fine. That gives maintainers something reviewable and lets later PRs build on a merged base instead of conflicting. — Reply to this email directly, view it on GitHub <#1418?email_source=notifications&email_token=CA3AEFPYC3F2JQTUREQQZUT5HNYG7A5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTKMJTGM2DEMBZGUZ2M4TFMFZW63VHNVSW45DJN5XKKZLWMVXHJLDGN5XXIZLSL5RWY2LDNM#issuecomment-5133420953>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/CA3AEFJSKM3LLSFOLN3MRST5HNYG7AVCNFSNUABFKJSXA33TNF2G64TZHMYTENBSHEYDAMRTHNEXG43VMU5TKMBRGMZDANRVHA42C5QC> . Triage notifications, keep track of coding agent tasks and review pull requests on the go with GitHub Mobile for iOS <https://github.com/notifications/mobile/ios/CA3AEFPLE5N46LR4YKJMWY35HNYG7A5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTKMJTGM2DEMBZGUZ2M4TFMFZW63VHNVSW45DJN5XKKZLWMVXHJKTGN5XXIZLSL5UW64Y> and Android <https://github.com/notifications/mobile/android/CA3AEFPE74ZBVHOZJ5V4T7D5HNYG7A5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTKMJTGM2DEMBZGUZ2M4TFMFZW63VHNVSW45DJN5XKKZLWMVXHJLTGN5XXIZLSL5QW4ZDSN5UWI>. Download it today! You are receiving this because you were mentioned.Message ID: ***@***.***>

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 86.72%. Comparing base (d2851ec) to head (c07788e).
⚠️ Report is 14 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #1418      +/-   ##
==========================================
+ Coverage   86.61%   86.72%   +0.11%     
==========================================
  Files          26       28       +2     
  Lines        3736     4069     +333     
==========================================
+ Hits         3236     3529     +293     
- Misses        500      540      +40     

Flag	Coverage Δ
unittests	86.72% <ø> (+0.11%)	⬆️

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.

[proofbyhuman]

Thanks for taking the time to go through all ten — one thread was far more useful than ten.

The process points are done: #1410 and #1414 are closed, and #1411, #1412, #1413, #1415 and #1416 are drafts. That leaves #1409, #1417 and #1418.

On re-running this notebook: I can't execute it as-is, and since the blocker is pre-existing code I'd rather ask than change something you didn't ask me to touch.

Cell 7 ends with:

plot = visualize.plot_results(results=boxes, image=r)
plt.imshow(plot)
plt.show()

plot_results returns matplotlib.figure.Figure (or Axes with axes=True), so plt.imshow raises TypeError: Image data of dtype object cannot be converted to float. Those three lines are unchanged from main — this PR doesn't touch them.

You're right that the outputs are the crux. To be precise about the current state: this branch cleared the stored outputs, because they embedded /Users/benweinstein/Downloads/... paths from the original run. With nbsphinx_execute = "never" in docs/conf.py, the rendered page now has no outputs at all rather than stale ones — which doesn't fix #999 either. So re-running is a hard requirement for this PR, not a nice-to-have; I just can't get there while cell 7 raises.

The minimal fix would be dropping the plt.imshow(plot) / plt.show() lines, since plot_results already draws with show=True. Would you prefer I do that here so the notebook can be executed and its outputs regenerated, or keep this PR to the path change and open a separate one for the execution fix?

One more thing I noticed while checking: australia.tif is 751x708 with 3 bands (EPSG:32753), not 4. The markdown in cells 3 and 5 walks through the four-band error and dropping the alpha channel, which was written for Plot13Ortho.tif and no longer matches what the code does. Let me know if you want that in scope too.