Feat/Improve Yolo-v4 Model #149
Conversation
|
Also can fix the suggestions by CodeFactor check |
| layer_names = self.net.getLayerNames() | ||
| self.output_layers = [layer_names[i[0] - 1] for i in self.net.getUnconnectedOutLayers()] | ||
|
|
||
| def read_config(self): |
There was a problem hiding this comment.
Similarly for this too. Maybe can put assert to check net.height and net.width are present in .cfg
| Obtains predicted boxes for predict_microservice | ||
| """ | ||
|
|
||
| def run(self, img, width_dict): |
There was a problem hiding this comment.
Validation of numpy matrix as input too
There was a problem hiding this comment.
I believe the input for both handler.py and local_inference.py is cv2.Mat, from cv2.imdecode and cv2.imread respectively. cv2.dnn.blobFromImage does accept a generic InputArray as image.
What should the validation here be for exactly?
There was a problem hiding this comment.
InputArray should be an interface for Mat. Since python numpy.ndarray is a binding to openCv Mat type. The validation should be numpy.ndarray. I think you can check with print(type(img))
|
|
||
| return self.get_filtered_boxes(img, height, width) | ||
|
|
||
| def get_filtered_boxes(self, img, height, width): |
There was a problem hiding this comment.
Can add validation to check if img is of the correct data type, numpy.ndarray?
Summary
The model has been improved, trained on more images, totaling 211 images. The inference code has also been refactored to improve maintainability.
Comparison
Old model was trained on 155 images, new model trained on 211 images. This table shows results which uses each model's own test split for 155 images, and the new model's test split for 56 new images.
Possible Improvement
With more compute time/power, k-fold cross validation could be used to improve performance on such a tiny dataset.
Refactor
Inferenceclasses have been created for the detection process. Seebase_inference.py,local_inference.py,service_inference.py. The predict_microservice has not yet been tested.