conversion of models from ultralytics/yolov8 to yours

[mahaling]

Can you share insights/code to convert ultralytics models to your version of yolo. I am specifically looking for conversion of yolov8. These are different from the base models you have provided.

[jahongir7174]

def port_weights(model1, model2):
    model1.eval()
    model2.eval()
    with torch.no_grad():
        m1_std = model1.state_dict().values()
        m2_std = model2.state_dict().values()
        for m1, m2 in zip(m1_std, m2_std):
            m1.copy_(m2)
    state = {'model': model1.half()}
    torch.save(state, 'weights/bestpt')

model1 is the dst model
model2 is the src model

[mahaling]

Thanks for this piece of code. But it looks like the architecture of ultralytics' YOLO is different from what your code has. Can you please confirm?

[ItayMeiri]

Hello @jahongir7174
I really appreciate your work - could you please provide a minimal example of porting? For instance, if I download the v8 weights for COCO from ultralytics and try to do the following:

import torch
from ultralytics import YOLO
from nets.nn import yolo_v8_n


downloaded_weights = "yolov8n.pt"
model = YOLO(downloaded_weights)
model_nn = yolo_v8_n(num_classes=80)

def port_weights(model1, model2):
    model1.eval()
    model2.eval()
    with torch.no_grad():
        m1_std = model1.state_dict().values()
        m2_std = model2.state_dict().values()
        for m1, m2 in zip(m1_std, m2_std):
            m1.copy_(m2)
    state = {'model': model1.half()}
    torch.save(state, 'weights/bestpt')

port_weights(model_nn, model)

I would get an error RuntimeError: The size of tensor a (32) must match the size of tensor b (16) at non-singleton dimension 0
I'm wondering how you ported the weights?