Help and configuring YOLO correctly ?

[newinai]

Hi,

I have some questions about YOLO configuration for my AI problem.

To resume :

• 2 classes to detect in my dataset, A with 4500 features and B with 1900 features.
• A shapes are generaly big and B shapes smaller (depends of the picture).
• Original pictures are 6000x4000px.

Configuration:

• 2 GPU.
• batch_size=16.
• train_times=3
• nb_epochs=100
• min_input_size=416 ; max_input_size=416

After multiple training trials :

• Yolo loss ~28 ; Yolo layer1 loss ~ < 1 ; Yolo layer2 loss ~ <1 ; Yolo layer3 loss: ~13

Evaluation :

• mAP is really bad ~ < 1
• Only A is detected

Some questions :

• Do i need to calculate my own dataset anchors ? If yes, is it compatible with my original size 6000px*4000px ?
• Is there a relation between anchors and min_input_size and max_input_size ?
• Do i need to increase min_input_size and max_input_size ?
• Why my loss not decreasing more ?
• Is it possible to print val_loss ? And is it a good thing to print it ? If no, when to stop training with only loss ?
• What do you suggest more ?

Thank you in advance !

[kabrau]

Do I need to calculate my own dataset anchors? If yes, is it compatible with my original size 6000px*4000px ?

• I think you have a problem because the YOLO reshape your images to go through the network
• "min_input_size": 288,
• "max_input_size": 448,

Is there a relation between anchors and min_input_size and max_input_size ?

• If you use gen_anchors.py to create anchors you do not have a problem, but, in my Fork, I change gen_acnhors.py to verify validations dataset (https://github.com/kabrau/keras-yolo3).

Do I need to increase min_input_size and max_input_size ?

• Yes

Why my loss not decreasing more ?

• Depends, there are many aspects, bad annotations, meaningless dataset, I do not understand what is your features (2 classes to detect in my dataset, A with 4500 features and B with 1900 features), for example, the classes are CAT or DOG, but you do not need features, the YOLO will create the features.

Is it possible to print val_loss ? And is it a good thing to print it ? If no, when to stop training with only loss ?

• Yes, I prefer to use validation dataset too, you need to change the fit_generator and add validations parameters. If you will have problems, send me a message.

What do you suggest more ?

• Show me one sample, maybe a can help you.

[sabrinatuli]

How can I compute min_input_size and max_input_size for my dataset? I have generated my own anchors, but don't know what to do with min and max_input_size.