I am using yolov5 with coco128 dataset and I am trying to add more classes and data to such.
The thing is, I am not sure if the operations is working correctly or as it should based on the results that I am seeing:
The issue: Every time I initialize my configurations, the output is not showing additional classes only one class in connection with the additions that were integrated.
The total number of classes before the additions of the classes is 80 and after the addition is it 106.
A display of the models alterations can be seen in the verbose showing model 80 swithced to 106. This is very interesting but I am lacking the understanding of what is happening in the operations from the begining and it's challenging to decipher the stages that I am seeing.
What I am trying to understand:
What does training from scratch mean? "is it that the model is trainined on my classes only or is it training of the 80 classes in addition to what I've combined
The scanning is picking up only one class as demonstrated in the image and producing a labels.chche file but this is not done for the rest, WHY?
'if all of the classes are therein the location why is it not beeing processes here?' The only proof I have of the classes actually being training but this is not detecting the labels correctly, is at the output validation images and training images.
- In the verbose epoch interations, the class states (ALL) but only the original 80 classes from coco128 is displayed and randomised but none of the additional classes can be seen or is displayed, Please can someone tell me WHY?
Even when training on 150 epoch the outcome is the same.
"This gives me the intuition that the model is not training on the additional data as the labels are not being called for evaluation" I am sure I am wrong here, hence me asking??!!
My Code:
Python3 train.py --img 640 --cfg yolov5s.yaml --hyp hyp.scratch-high.yaml --batch 32 --epochs 5 --data coco128.yaml --weights yolov5s.pt --workers 24
I provided an image further breaking down the challenge and what is missing from a visual perspective.
