Inference artifact saving by addobosz · Pull Request #453 · GHOST-Science-Club/tree-classification-irim

addobosz · 2026-04-08T16:31:54Z

This pull request introduces several improvements and new features across the codebase, focusing on data augmentation, configuration flexibility, training/inference workflow, and code robustness. The most significant changes include adding configurable data augmentation, improving model checkpointing and artifact logging, and refactoring accuracy tracking and model freezing logic.

Data pipeline and augmentation improvements:

Added a configurable data augmentation section to src/config.yaml, enabling options such as random flips, rotation, brightness, contrast, saturation, and hue adjustments. The augmentation pipeline is implemented in get_train_transform() in src/dataset.py and applied during training dataset setup. [1] [2] [3]
Added ONNX as a dependency in pyproject.toml to support ONNX model export during inference.

Training and inference workflow enhancements:

Introduced an inference section in src/config.yaml to control mask size, ONNX export, checkpoint paths, artifact tracking, and device selection. Device handling in dataloader parameters is now more robust to "cuda"/"gpu" string variants. [1] [2]
Added a new SegmentationWrapper module for wrapping classifiers to produce segmentation-style outputs, supporting normalization and optional input rescaling.

Model and metrics tracking improvements:

Refactored accuracy tracking in src/models/classifier_module.py to maintain separate accuracy metrics for training, validation, and test phases, and improved logging granularity. [1] [2]
Updated the fine-grained model to support freezing the feature extractor backbone based on config, and fixed the forward pass to respect training/eval mode. [1] [2] [3]

Experiment tracking and artifact logging:

Enhanced experiment reproducibility and tracking by saving the label map alongside the best checkpoint, logging model artifacts to Weights & Biases (wandb), and improving run metadata handling. [1] [2] [3]
Improved dataset split reporting with a new summary function to check class distribution and missing classes in each split. [1] [2]

Miscellaneous:

Adjusted default batch size in config to 1 for training, likely for debugging or memory constraints.
Improved ModelCheckpoint/EarlyStopping handling for more flexible monitoring and mode selection.

These changes collectively improve the flexibility, robustness, and traceability of the training and inference pipeline.

…pipeline

- Add onnx>=1.20.0 to project dependencies - Update uv.lock with onnx and ml-dtypes packages - Update filelock from 3.19.1 to 3.20.1

- Remove trailing whitespace - Fix inconsistent indentation - Adjust line spacing - Format quotes to double quotes - Add missing blank line at end of file

github-actions · 2026-04-08T16:32:13Z

Dependency Review

The following issues were found:

❌ 2 vulnerable package(s)
✅ 0 package(s) with incompatible licenses
✅ 0 package(s) with invalid SPDX license definitions
✅ 0 package(s) with unknown licenses.

See the Details below.

Snapshot Warnings

⚠️: No snapshots were found for the head SHA 3c494f0.

Ensure that dependencies are being submitted on PR branches and consider enabling retry-on-snapshot-warnings. See the documentation for more information and troubleshooting advice.

Vulnerabilities

uv.lock

Name	Version	Vulnerability	Severity
onnx	1.20.0	ONNX Untrusted Model Repository Warnings Suppressed by silent=True in onnx.hub.load() — Silent Supply-Chain Attack	high
		onnx Vulnerable to Path Traversal via Symlink	high
		ONNX: Malicious ONNX models can crash servers by exploiting unprotected object settings.	high
		ONNX: TOCTOU arbitrary file read/write in save_external_dat	high
		ONNX: Arbitrary File Read via ExternalData Hardlink Bypass in ONNX load	moderate
		ONNX: External Data Symlink Traversal	moderate
filelock	3.20.1	filelock Time-of-Check-Time-of-Use (TOCTOU) Symlink Vulnerability in SoftFileLock	moderate

OpenSSF Scorecard

Package	Version	Score	Details
pip/onnx	1.20.0	Unknown	Unknown
pip/filelock	3.20.1	Unknown	Unknown
pip/ml-dtypes	0.5.4	Unknown	Unknown

Scanned Files

uv.lock

Copilot

Pull request overview

Note

Copilot was unable to run its full agentic suite in this review.

Adds end-to-end inference artifact saving and ONNX export support by logging the best checkpoint (plus label map) to W&B, introducing a segmentation-style wrapper for classifiers, and adding a standalone inference script.

Changes:

Added SegmentationWrapper to convert classifier logits into a dense per-pixel mask for ONNX/segmentation-like outputs.
Updated training flow to select/log the best checkpoint as a W&B artifact and persist a .label_map.json next to it.
Added an inference entrypoint with optional W&B artifact download + ONNX export, plus configurable train-time augmentation.

Reviewed changes

Copilot reviewed 9 out of 588 changed files in this pull request and generated 7 comments.

Show a summary per file

File	Description
src/models/segmentation_wrapper.py	New wrapper module that normalizes inputs and expands class probabilities into a `(B, C, H, W)` mask.
src/models/model_factory.py	Passes `freeze` through to `FineGrainedModel`.
src/models/diversified_model.py	Stabilizes top-k selection and pooling behavior; adds optional freezing of backbone.
src/models/classifier_module.py	Splits accuracy metrics by stage and aligns fine-grained training behavior with `self.training`.
src/main.py	Logs best checkpoint + label map as explicit W&B artifact; makes checkpoint monitor configurable; prints split summary.
src/inference.py	New inference script: load checkpoint (local or W&B artifact), run inference, optionally export/upload ONNX.
src/dataset.py	Adds configurable train augmentation pipeline and improves `pin_memory` device detection.
src/config.yaml	Adds inference config and augmentation config; changes defaults (e.g., batch size, device).
pyproject.toml	Adds `onnx` dependency required for export/metadata writing.

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

Copilot · 2026-04-09T10:42:13Z

+def _wandb_safe_metadata_value(value):
+    if isinstance(value, float) and not math.isfinite(value):
+        return None
+    if isinstance(value, torch.Tensor):
+        if value.numel() == 1:
+            scalar = value.item()
+            return None if isinstance(scalar, float) and not math.isfinite(scalar) else scalar
+        return value.tolist()
+    return value


math.isfinite is used but math is not imported in the shown imports at the top of main.py, which will raise NameError at runtime. Add import math near the top of the file (or replace math.isfinite with an equivalent already-imported utility).

Copilot · 2026-04-09T10:42:14Z


 training:
-  batch_size: 32
+  batch_size: 1


The committed defaults make the repo configuration brittle/non-portable: (1) raw_images_path is an absolute, user-specific filesystem path; (2) wandb_artifact points to a specific run artifact; (3) export_onnx: true enables export by default; and (4) training.batch_size changed from 32 to 1. Recommend reverting these to safe, portable defaults (e.g., comment out run-specific values, use relative example paths, set export_onnx: false by default, and restore a reasonable default batch_size).

Copilot · 2026-04-09T10:42:14Z

+inference:
+  mask_size: 224
+  export_onnx: true
+  checkpoint_path: "checkpoints/model.ckpt"
+  # wandb_artifact: "ghost-irim/ghost-irim/model-kx2co2o3:v0"
+  wandb_artifact: "ghost-irim/ghost-irim/best-model-inference-artifact-saving-vft1nd:v0"
+  wandb_project: "ghost-irim"
+  # Set to true to run inference on raw images (ONNX-style testing)
+  # Set to false to run inference on test dataset split
+  use_raw_images: false
+  # Path to directory containing raw images (relative to project root)
+  # raw_images_path: "test_images_qgis/10122025_193353"
+  # raw_images_path: "/home/adam/Documents/programming/tree-classification-irim/src/data/Abies_alba/train"
+  raw_images_path: "/home/adam/Documents/programming/tree-classification-irim/src/data/Robinia_pseudoacacia/train"
+device: "cuda"


The committed defaults make the repo configuration brittle/non-portable: (1) raw_images_path is an absolute, user-specific filesystem path; (2) wandb_artifact points to a specific run artifact; (3) export_onnx: true enables export by default; and (4) training.batch_size changed from 32 to 1. Recommend reverting these to safe, portable defaults (e.g., comment out run-specific values, use relative example paths, set export_onnx: false by default, and restore a reasonable default batch_size).

Copilot · 2026-04-09T10:42:14Z

+    run = wandb.init(project=project_name, job_type="inference")
+
+    artifact = run.use_artifact(artifact_path, type="model")
+    artifact_dir = artifact.download()
+


download_checkpoint_from_wandb() creates and finishes a W&B run immediately. As a result, later sections guarded by if wandb.run is not None: (ONNX artifact upload and metric logging) will never execute in the common 'download-from-artifact' path. Consider either (a) keeping the run open and finishing at the end of main(), or (b) moving W&B initialization to main() (optionally behind a config flag like inference.log_to_wandb), and having download_checkpoint_from_wandb() reuse the active run.

Copilot · 2026-04-09T10:42:14Z

+    run.finish()
+
+    return checkpoint_path


download_checkpoint_from_wandb() creates and finishes a W&B run immediately. As a result, later sections guarded by if wandb.run is not None: (ONNX artifact upload and metric logging) will never execute in the common 'download-from-artifact' path. Consider either (a) keeping the run open and finishing at the end of main(), or (b) moving W&B initialization to main() (optionally behind a config flag like inference.log_to_wandb), and having download_checkpoint_from_wandb() reuse the active run.

Copilot · 2026-04-09T10:42:15Z

+
+    def __init__(self, classifier: nn.Module, mask_size: int = 224, mean=None, std=None, input_rescale=False):
+        super().__init__()
+        self.classifier = classifier.eval()


Calling .eval() on the passed-in classifier mutates external state as a side effect of constructing SegmentationWrapper. This can be surprising if the caller expects to control train/eval mode (or reuses the classifier elsewhere). Prefer storing self.classifier = classifier and letting the caller set .eval() (as is already done in inference.py), or documenting clearly that the wrapper forces eval mode.

Copilot · 2026-04-09T10:42:15Z

+import wandb
+from omegaconf import OmegaConf
+from tqdm import tqdm
+from torchvision.utils import save_image


save_image is imported but never used anywhere in this file. Removing it avoids dead imports and reduces lint noise.

addobosz and others added 18 commits January 7, 2026 19:27

feat: segmentation-wrapper

9f76d05

adjust dvc.ymal for inference

a5986f4

add: push onnx to wandb artifacts

9bb9c0d

fix: change "gpu" parameter to "cuda"

593c87d

fix: adjust output shapes and fix deprecation warning

eb11741

fix: adjust onnx export to old pytorch version

79a9591

fix: change output mask structure

e3198e8

fix: remove save image line from inference

a854d77

fix: add preprocessing and standardization as it was in the training …

9e7628f

…pipeline

add: inference summary statistics

c8e856c

debug: standardization

3f659e7

build(deps): add onnx dependency and update lockfile

c3e876b

- Add onnx>=1.20.0 to project dependencies - Update uv.lock with onnx and ml-dtypes packages - Update filelock from 3.19.1 to 3.20.1

style: format inference.py with ruff

3107256

- Remove trailing whitespace - Fix inconsistent indentation - Adjust line spacing - Format quotes to double quotes - Add missing blank line at end of file

add inference on real images

f09b689

add explicit artifact saving

2b03320

switch dvc file to training

72bdb54

fix: validation is nan

6ee610d

add: training augmentations for overfitting mitigation and more logging

1f60ebb

style: ruff formatting

3c494f0

rojberr requested a review from Copilot April 9, 2026 10:40

Copilot AI reviewed Apr 9, 2026

View reviewed changes

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Inference artifact saving#453

Inference artifact saving#453
addobosz wants to merge 19 commits into
mainfrom
inference-artifact-saving

addobosz commented Apr 8, 2026 •

edited by rojberr

Loading

Uh oh!

github-actions Bot commented Apr 8, 2026 •

edited

Loading

Uh oh!

Copilot AI left a comment

Uh oh!

Copilot AI Apr 9, 2026

Uh oh!

Copilot AI Apr 9, 2026

Uh oh!

Copilot AI Apr 9, 2026

Uh oh!

Copilot AI Apr 9, 2026

Uh oh!

Copilot AI Apr 9, 2026

Uh oh!

Copilot AI Apr 9, 2026

Uh oh!

Copilot AI Apr 9, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Conversation

addobosz commented Apr 8, 2026 • edited by rojberr Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

github-actions Bot commented Apr 8, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Dependency Review

Snapshot Warnings

Vulnerabilities

uv.lock

OpenSSF Scorecard

Scanned Files

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

Copilot AI Apr 9, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Apr 9, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Apr 9, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Apr 9, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Apr 9, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Apr 9, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Apr 9, 2026

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

addobosz commented Apr 8, 2026 •

edited by rojberr

Loading

github-actions Bot commented Apr 8, 2026 •

edited

Loading