Comparing the Weights and Biases of several different Yolov8s models trained through transfer learning

I have 3 different yolov8s models which I would like to assess:

1. Yolov8s trained with the normal model.train() command

2. Yolo8vs model trained with a frozen backbone

3. Yolov8s model trained with all layers frozen

I am using a callback function to freeze the weights, see below:

def freeze_layer(trainer):
    model = trainer.model
    num_freeze = 10
    print(f"Freezing {num_freeze} layers")
    freeze = [f'model.{x}.' for x in range(num_freeze)]  # layers to freeze
    for k, v in model.named_parameters():
        v.requires_grad = True  # train all layers
        if any(x in k for x in freeze):
            print(f'freezing {k}')
            v.requires_grad = False
    print(f"{num_freeze} layers are freezed.")

if __name__ == "__main__":
    model = YOLO(yolov8s.pt)
    model.add_callback("on_train_start", freeze_layer)
    model.train(
        data="coco128.yaml",
        epochs=300,
        imgsz=640
        )

I want to be able to assess the differences in the weights and biases between these 3 neural networks. I want to see which neurons are dead after transfer learning and assess the major differences between all three neural networks. Are there any standard solutions to use to gain this kind of insight into a neural network?

I have compared the standard training metrics that yolo provides like MAP and Loss and everything is as expected. The more layers you freeze the worse performce gets. And when I freeze the whole network the performance is much worse than the other 2 instances.

I barely know where to begin when it comes to comparing the same network architetchure but just trained differently.

Any help is much appreciated.

YOLO architecture is quite complicated (depending on which size you're using), and it can be quite counter-intuitive to compare / match the weights of intermediate representations in a specific layer. One approach would be to Visualize the intermediate activations at each layer. You can also plot the distribution of the weights for each layer. Or calculating the norm between the weights of each model.