Shapes dont match if resolution other than default 640

[elicz]

Change from:
compute_loss = torch_losses.Loss(device=device, nc=num_classes)

to
compute_loss = torch_losses.Loss(device=device, nc=num_classes, input_shape=imgsz)

keras_cv_attention_models/keras_cv_attention_models/yolov8/train.py

Line 79 in 9ee3d73

compute_loss = torch_losses.Loss(device=device, nc=num_classes)

(I was using EfficientNet backbone, resolution of inputs 320x320, and keras backend - shapes mismatched in loss function)

[leondgarse]

Ya, you are right. Fix is updated. It's actually updated in torch_coco_train_script.py, but missed here.