support RT-DETR outputs by yijint · Pull Request #221 · PUTvision/qgis-plugin-deepness

yijint · 2025-08-11T18:07:40Z

Support model predictions from https://docs.ultralytics.com/models/rtdetr/

…with attributes

przemyslaw-aszkowski

Hi,
Thank you for your contribution!
I left a few comments for you to look at.
Moreover, the integration tests do not pass, please have a look at it: https://github.com/PUTvision/qgis-plugin-deepness/actions/runs/16948111454/job/48079565026?pr=221

Let me know if you need any help!

przemyslaw-aszkowski · 2025-08-14T10:39:19Z

src/deepness/common/processing_parameters/detection_parameters.py

    Parameters for Inference of detection model (including pre/post-processing) obtained from UI.
    """
-
+    from deepness.processing.models.model_base import ModelBase


why import here instead of at the top of the file?

przemyslaw-aszkowski · 2025-08-14T10:41:42Z

src/deepness/metadata.txt

 qgisMinimumVersion=3.22
 description=Inference of deep neural network models (ONNX) for segmentation, detection and regression
-version=0.7.0
+version=0.6.5


Please increase the version number to 0.8.0

przemyslaw-aszkowski · 2025-08-14T10:43:08Z

src/deepness/processing/map_processor/map_processor_detection.py

            filtered_bounding_boxes = [det for det in bounding_boxes if det.clss == channel_id]
            print(f'Detections for class {channel_id}: {len(filtered_bounding_boxes)}')
-
-            vlayer = QgsVectorLayer("multipolygon", self.model.get_channel_name(0, channel_id), "memory")


Why was this code removed? Is it intended? I think it was added in a recent PR. Maybe there were some merge conflicts not resolved correctly?

bartoszptak · 2025-08-14T10:52:29Z

src/deepness/processing/classify_chip.py

+# import packages
+from qgis.core import QgsTask, QgsApplication, QgsProject, QgsVectorLayer, QgsField, QgsWkbTypes
+from qgis.PyQt.QtCore import QVariant
+import torch


You can't use PyTorch in Deepness

Copilot

Pull Request Overview

This PR adds support for RT-DETR model outputs in the deepness package, extending the existing detection capabilities to handle predictions from Ultralytics RT-DETR models. Additionally, it introduces a new chip classification feature for analyzing image patches within bounding boxes.

Adds RT-DETR detector type with custom preprocessing and postprocessing logic
Implements chip classification functionality with PyTorch model support
Updates model base class to handle different preprocessing paths based on detector type

Reviewed Changes

Copilot reviewed 10 out of 10 changed files in this pull request and generated 10 comments.

Show a summary per file

File	Description
src/deepness/processing/models/preprocessing_utils.py	Adds RT-DETR specific normalization functions and debug print statements
src/deepness/processing/models/model_base.py	Updates preprocessing logic to handle RT-DETR differently from other models
src/deepness/processing/models/detector.py	Implements RT-DETR postprocessing method with confidence filtering and NMS
src/deepness/processing/map_processor/map_processor_detection.py	Removes attribute setting code and adds debug parameters
src/deepness/processing/classify_chip.py	New file implementing chip classification task with PyTorch integration
src/deepness/metadata.txt	Version downgrade from 0.7.0 to 0.6.5
src/deepness/deepness_dockwidget.py	Adds comprehensive GUI for chip classification configuration
src/deepness/deepness.py	Integrates chip classification task and result handling
src/deepness/common/processing_parameters/detection_parameters.py	Adds RT_DETR enum value and parameters
src/deepness/common/processing_parameters/classify_chip_parameters.py	New dataclass for chip classification parameters

Comments suppressed due to low confidence (1)

src/deepness/deepness_dockwidget.py:272

Missing space before parameter 'i'. Should be 'tile_params_batched: List[TileParams], i)'.

        # insert the entire block just below "Training data export" block

_{Tip: Customize your code reviews with copilot-instructions.md. Create the file or learn how to get started.}

Copilot · 2025-08-14T14:37:55Z

src/deepness/processing/models/preprocessing_utils.py

    :return: Batch of tiles with standardized values
    """
+    print(f'params.mean: {params.mean}')
+    print(f'params.std: {params.std}')


Debug print statements should be removed from production code. Use proper logging instead.

Suggested change

print(f'params.std: {params.std}')

logging.debug(f'params.mean: {params.mean}')

logging.debug(f'params.std: {params.std}')

Copilot · 2025-08-14T14:37:56Z

src/deepness/processing/models/preprocessing_utils.py

    :return: Batch of tiles with standardized values
    """
+    print(f'params.mean: {params.mean}')
+    print(f'params.std: {params.std}')


Debug print statements should be removed from production code. Use proper logging instead.

Suggested change

print(f'params.std: {params.std}')

logging.debug(f'params.mean: {params.mean}')

logging.debug(f'params.std: {params.std}')

Copilot · 2025-08-14T14:37:56Z

src/deepness/processing/models/model_base.py

+        # np.save(f"unprocessed_qgis_input{i}", tiles_batched)
        input_batch = self.preprocessing(tiles_batched)
+        # print(f'preprocessed tiles_batched.shape: {input_batch.shape}')
+        # np.save(f"preprocessed_qgis_input{i}", input_batch)


Commented debug code should be removed from the codebase.

Suggested change

# np.save(f"preprocessed_qgis_input{i}", input_batch)

input_batch = self.preprocessing(tiles_batched)

Copilot · 2025-08-14T14:37:57Z

src/deepness/processing/models/model_base.py

        """
+        TILE_SIZE = tiles_batched.shape[1]
+        # print(f'unprocessed tiles_batched.shape: {tiles_batched.shape}')
+        # np.save(f"unprocessed_qgis_input{i}", tiles_batched)


Commented debug code should be removed from the codebase.

Suggested change

# np.save(f"unprocessed_qgis_input{i}", tiles_batched)

Copilot · 2025-08-14T14:37:57Z

src/deepness/processing/models/model_base.py

+        # np.save(f"unprocessed_qgis_input{i}", tiles_batched)
        input_batch = self.preprocessing(tiles_batched)
+        # print(f'preprocessed tiles_batched.shape: {input_batch.shape}')
+        # np.save(f"preprocessed_qgis_input{i}", input_batch)


Commented debug code should be removed from the codebase.

Suggested change

# np.save(f"preprocessed_qgis_input{i}", input_batch)

input_batch = self.preprocessing(tiles_batched)

Copilot · 2025-08-14T14:37:57Z

src/deepness/processing/models/model_base.py

+        # np.save(f"unprocessed_qgis_input{i}", tiles_batched)
        input_batch = self.preprocessing(tiles_batched)
+        # print(f'preprocessed tiles_batched.shape: {input_batch.shape}')
+        # np.save(f"preprocessed_qgis_input{i}", input_batch)


Commented debug code should be removed from the codebase.

Suggested change

# np.save(f"preprocessed_qgis_input{i}", input_batch)

input_batch = self.preprocessing(tiles_batched)

Copilot · 2025-08-14T14:37:58Z

src/deepness/processing/models/detector.py

+            return [], [], []
+        outputs_x1y1x2y2 = self.xywh2xyxy(outputs_filtered)
+        boxes = outputs_x1y1x2y2[:,:4]
+        probabilities = outputs_x1y1x2y2[:, 4:][:,0]


The probability extraction is incorrect. This only takes the first probability column, but RT-DETR outputs should contain class probabilities. Use np.max(outputs_x1y1x2y2[:, 4:], axis=1) to get the maximum probability per detection.

Suggested change

probabilities = outputs_x1y1x2y2[:, 4:][:,0]

probabilities = np.max(outputs_x1y1x2y2[:, 4:], axis=1)

classes = np.argmax(outputs_x1y1x2y2[:, 4:], axis=1)

Copilot · 2025-08-14T14:37:58Z

src/deepness/processing/models/detector.py

+
+        boxes = np.array(boxes, dtype=int)
+        conf = probabilities
+        classes = np.argmax(np.expand_dims(conf, axis=1), axis=1)


This logic is incorrect. The conf variable contains single probability values, not class probabilities. Classes should be extracted from the original model output before filtering: classes = np.argmax(outputs_x1y1x2y2[:, 4:], axis=1).

Suggested change

classes = np.argmax(np.expand_dims(conf, axis=1), axis=1)

classes = np.argmax(outputs_x1y1x2y2[pick_indxs, 4:], axis=1)

Copilot · 2025-08-14T14:37:58Z

src/deepness/processing/map_processor/map_processor_detection.py


-            bounding_boxes_in_tile_batched = self._process_tile(tile_img_batched, tile_params_batched)
+            bounding_boxes_in_tile_batched = self._process_tile(tile_img_batched, tile_params_batched, i)
+            i+=1


Missing space around operator. Should be 'i += 1'.

Suggested change

i+=1

i += 1

Copilot · 2025-08-14T14:37:59Z

src/deepness/metadata.txt

 qgisMinimumVersion=3.22
 description=Inference of deep neural network models (ONNX) for segmentation, detection and regression
-version=0.7.0
+version=0.6.5


Version appears to be downgraded from 0.7.0 to 0.6.5. This could indicate an error in version management or merge conflict.

Suggested change

version=0.6.5

version=0.7.0

przemyslaw-aszkowski · 2025-11-20T11:35:12Z

Closing due to no activity

yijint added 2 commits August 11, 2025 10:41

support RT-DETR outputs

91500e4

code to integrate+run classifier model and save outputs as new layer …

7e424cc

…with attributes

przemyslaw-aszkowski requested changes Aug 14, 2025

View reviewed changes

przemyslaw-aszkowski requested a review from Copilot August 14, 2025 10:49

bartoszptak reviewed Aug 14, 2025

View reviewed changes

This comment was marked as outdated.

Sign in to view

bartoszptak requested a review from Copilot August 14, 2025 14:31

Copilot AI reviewed Aug 14, 2025

View reviewed changes

przemyslaw-aszkowski closed this Nov 20, 2025

	print(f'params.std: {params.std}')
	logging.debug(f'params.mean: {params.mean}')
	logging.debug(f'params.std: {params.std}')

	# np.save(f"preprocessed_qgis_input{i}", input_batch)
	input_batch = self.preprocessing(tiles_batched)

	probabilities = outputs_x1y1x2y2[:, 4:][:,0]
	probabilities = np.max(outputs_x1y1x2y2[:, 4:], axis=1)
	classes = np.argmax(outputs_x1y1x2y2[:, 4:], axis=1)

	classes = np.argmax(np.expand_dims(conf, axis=1), axis=1)
	classes = np.argmax(outputs_x1y1x2y2[pick_indxs, 4:], axis=1)

Conversation

yijint commented Aug 11, 2025

Uh oh!

przemyslaw-aszkowski left a comment

Choose a reason for hiding this comment

Uh oh!

przemyslaw-aszkowski Aug 14, 2025

Choose a reason for hiding this comment

Uh oh!

przemyslaw-aszkowski Aug 14, 2025

Choose a reason for hiding this comment

Uh oh!

przemyslaw-aszkowski Aug 14, 2025

Choose a reason for hiding this comment

Uh oh!

bartoszptak Aug 14, 2025

Choose a reason for hiding this comment

Uh oh!

This comment was marked as outdated.

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull Request Overview

Reviewed Changes

Uh oh!

Copilot AI Aug 14, 2025

Choose a reason for hiding this comment

Uh oh!

Copilot AI Aug 14, 2025

Choose a reason for hiding this comment

Uh oh!

Copilot AI Aug 14, 2025

Choose a reason for hiding this comment

Uh oh!

Copilot AI Aug 14, 2025

Choose a reason for hiding this comment

Uh oh!

Copilot AI Aug 14, 2025

Choose a reason for hiding this comment

Uh oh!

Copilot AI Aug 14, 2025

Choose a reason for hiding this comment

Uh oh!

Copilot AI Aug 14, 2025

Choose a reason for hiding this comment

Uh oh!

Copilot AI Aug 14, 2025

Choose a reason for hiding this comment

Uh oh!

Copilot AI Aug 14, 2025

Choose a reason for hiding this comment

Uh oh!

Copilot AI Aug 14, 2025

Choose a reason for hiding this comment

Uh oh!

przemyslaw-aszkowski commented Nov 20, 2025

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants