Lock Down Sample Configs #85

JSabadin · 2024-10-01T11:20:56Z

Configuration Updates

Summary

This PR introduces the creation of light and heavy configuration files for various tasks, including segmentation, detection, keypoints, and classification.

Key Changes

Added light and heavy model configurations for:
- Segmentation
- Detection
- Keypoints
- Classification

klemen1999

Left some comments. Overall let's try to handle variants similarly as we do with e.g. EfficientRep or RepPANNeck.

klemen1999 · 2024-10-01T12:40:45Z

configs/classification_light_model.yaml

@@ -1,25 +1,24 @@
 # Example configuration for training a predefined segmentation model


Nitpick: Let's add this comment in other configs sample configs as well and describe what are they for (e.g. "Example configuration for training a light segmentation model" - or something similar)

klemen1999 · 2024-10-01T12:41:08Z

configs/coco_model.yaml

@@ -105,7 +96,7 @@ trainer:
  n_sanity_val_steps: 1
  profiler: null
  verbose: True
-  batch_size: 4
+  batch_size: 8
  accumulate_grad_batches: 1
  epochs: &epochs 200
  n_workers: 8


Can we double check if train_metrics_interval parameter is even used anywhere? Because I believe we don't actually call a function where this is used. If that is the case then we should remove this config parameter (and the function)

klemen1999 · 2024-10-01T12:44:38Z

configs/example_export.yaml

    normalize:
      active: True

-  batch_size: 4
+  batch_size: 8
  epochs: &epochs 200
  n_workers: 4
  validation_interval: 10


Let's include exporter.output_names to this config as well. For detection models to work with dai.YoloNeuralNetwork out of the box (with included postprocessing) the output names should be [output1_yolov6r2, output2_yolov6r2, output3_yolov6r2]

klemen1999 · 2024-10-01T12:45:39Z

configs/segmentation_light_model.yaml

@@ -0,0 +1,44 @@
+# DDRNet-23-slim model for segmentation


Let's remove this comment about specific of DDRnet

klemen1999 · 2024-10-01T12:46:26Z

configs/segmentation_light_model.yaml

+# Refer to here for optimal hyperparameters for this model: https://github.com/Deci-AI/super-gradients/blob/4797c974c7c445d12e2575c468848d9c3e04becd/src/super_gradients/recipes/cityscapes_ddrnet.yaml#L4
+
+model:
+  name: ddrnet_segmentation


Change model names to reflect the actual model - e.g. segmentation_light (same comment for other configs)

klemen1999 · 2024-10-01T12:47:11Z

luxonis_train/attached_modules/losses/efficient_keypoint_bbox_loss.py

@@ -33,10 +33,10 @@ def __init__(
        n_warmup_epochs: int = 4,
        iou_type: IoUType = "giou",
        reduction: Literal["sum", "mean"] = "mean",
-        class_loss_weight: float = 1.0,
-        iou_loss_weight: float = 2.5,
+        class_loss_weight: float = 0.5,


Are these changes based on the model source repo/paper?

Those are based on YOLOv8.

klemen1999 · 2024-10-01T12:54:45Z

luxonis_train/config/predefined_models/classification_model.py

    task_name: str | None = None
+    visualizer_params: Params = field(


Should we rather update the default classification visualizer params with these values?

klemen1999 · 2024-10-01T13:01:00Z

luxonis_train/config/predefined_models/classification_model.py

@@ -14,19 +14,32 @@

 @dataclass
 class ClassificationModel(BasePredefinedModel):
-    backbone: str = "MicroNet"
+    variant: str | None = None


Can we make the variants more similar to how we handle them in e.g. EfficientRep? So make predefined model an actual class (instead of dataclass) with all the parameters being None by default (except variant="lite" by default). Then have VariantLiteral type ["lite", "heavy"] and a function that gets specific parameters based on variant and uses them if others aren't specifid (e.g. uses "ResNet-18" as backbone if variant="lite" and no backbone is set, but if backbone is explicitly set then use that one.

The same can be done for all ther predefined models.

klemen1999 · 2024-10-01T13:01:57Z

luxonis_train/config/predefined_models/segmentation_model.py

@@ -14,36 +14,61 @@

 @dataclass
 class SegmentationModel(BasePredefinedModel):
-    backbone: str = "MicroNet"
+    variant: str | None = None


Keep backbone as parameter (same as with ClassificationModel) so it can be explicitly set.

…ations

kozlov721

Left some comments. Additionally, the test_predefined_models in tests/integration/test_simple.py will need to be changed to use the new configs.
Otherwise LGTM

configs/classification_heavy_model.yaml

configs/classification_light_model.yaml

configs/example_export.yaml

configs/classification_heavy_model.yaml

luxonis_train/config/predefined_models/classification_model.py

klemen1999 · 2024-10-01T18:28:31Z

We should also update the README under predefined_models accordingly

codecov · 2024-10-02T18:35:15Z

Codecov Report

Attention: Patch coverage is 94.89796% with 5 lines in your changes missing coverage. Please review.

Please upload report for BASE (dev@c97f921). Learn more about missing BASE report.

✅ All tests successful. No failed tests found.

Files with missing lines	Patch %	Lines
...ain/config/predefined_models/segmentation_model.py	93.54%	2 Missing ⚠️
...n/config/predefined_models/classification_model.py	95.00%	1 Missing ⚠️
..._train/config/predefined_models/detection_model.py	95.23%	1 Missing ⚠️
...nfig/predefined_models/keypoint_detection_model.py	95.65%	1 Missing ⚠️

Additional details and impacted files

@@          Coverage Diff           @@
##             dev      #85   +/-   ##
======================================
  Coverage       ?   96.66%           
======================================
  Files          ?      142           
  Lines          ?     6324           
  Branches       ?        0           
======================================
  Hits           ?     6113           
  Misses         ?      211           
  Partials       ?        0

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

Co-authored-by: klemen1999 <[email protected]> Co-authored-by: Martin Kozlovsky <[email protected]>

JSabadin added 6 commits October 1, 2024 10:16

Refactor and clean up training config files

636af39

Making configurations more consistent

bb559c6

Hyperparameter adjustments

6f837e4

Correct image size

f213b1c

Correct loader

ca147ca

Correct hyperparameters

83c2b75

JSabadin requested a review from a team as a code owner October 1, 2024 11:20

JSabadin requested review from kozlov721, klemen1999, tersekmatija and conorsim and removed request for a team October 1, 2024 11:20

github-actions bot assigned JSabadin Oct 1, 2024

github-actions bot added the fix Fixing a bug label Oct 1, 2024

Correct detection heavy variant

a967a1d

klemen1999 reviewed Oct 1, 2024

View reviewed changes

Refining model variant logic and making minor adjustments to configur…

d10907b

…ations

JSabadin requested a review from klemen1999 October 1, 2024 15:14

JSabadin added 2 commits October 1, 2024 17:16

Removing unnecessary comments

dd393f0

Changed model names to reflect the actual models

a94c8d7

kozlov721 reviewed Oct 1, 2024

View reviewed changes

klemen1999 added 5 commits October 1, 2024 19:56

fixed some mistakes in configs

263fe43

removed unnecessary variable capturing in yamls

c987592

fix in classification predefined model

1078457

small reordering or parameters

7143e91

updated predefined model tests

ea3d3c8

JSabadin and others added 3 commits October 2, 2024 08:40

Update README and apply minor fix to pass tests.

80b0c04

small fix in README, removed mention of ImplicitKeypointBBoxHead

6801173

lowered number of epochs in tests

200a37d

fixed lowered epochs

1472d02

kozlov721 changed the title ~~Fix/lock down sample configs~~ Lock Down Sample Configs Oct 2, 2024

removed trailing whitespaces

3532623

klemen1999 merged commit ca9ab66 into dev Oct 3, 2024
9 checks passed

klemen1999 deleted the fix/lock-down-sample-configs branch October 3, 2024 08:21

kozlov721 added a commit that referenced this pull request Oct 9, 2024

Lock Down Sample Configs (#85)

809eeeb

Co-authored-by: klemen1999 <[email protected]> Co-authored-by: Martin Kozlovsky <[email protected]>

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Lock Down Sample Configs #85

Lock Down Sample Configs #85

JSabadin commented Oct 1, 2024

klemen1999 left a comment

klemen1999 Oct 1, 2024

klemen1999 Oct 1, 2024

klemen1999 Oct 1, 2024

klemen1999 Oct 1, 2024

klemen1999 Oct 1, 2024

klemen1999 Oct 1, 2024

JSabadin Oct 1, 2024

klemen1999 Oct 1, 2024

klemen1999 Oct 1, 2024

klemen1999 Oct 1, 2024

kozlov721 left a comment

klemen1999 commented Oct 1, 2024

codecov bot commented Oct 2, 2024

		@@ -1,25 +1,24 @@
		# Example configuration for training a predefined segmentation model

		task_name: str \| None = None
		visualizer_params: Params = field(

Lock Down Sample Configs #85

Lock Down Sample Configs #85

Conversation

JSabadin commented Oct 1, 2024

Configuration Updates

Summary

Key Changes

klemen1999 left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

kozlov721 left a comment

Choose a reason for hiding this comment

klemen1999 commented Oct 1, 2024

codecov bot commented Oct 2, 2024

Codecov Report