Glenn Jocher

@liu15509348793 you're correct. For deterministic algorithms to take effect, PyTorch 2.0.0 or above is required. If you're using an earlier version, deterministic behavior may not be fully supported.

You're welcome! If you have any more questions, feel free to ask.

@lpy5618 if the run is stable for a few epochs and then flips to `nan`, it’s usually triggered by a specific batch (bad/degenerate mask/label, extreme augmentation outcome) or an unstable...

Sounds good—after upgrading, restart your Python session/kernel to ensure the updated package is actually imported, then rerun and paste a fresh `yolo checks` plus the log lines right before/at the...

@majnas seeing `nan` from the very first epoch strongly suggests something is off before any meaningful learning happens, and the `optimizer: MuSGD(...) with parameter groups 0 weight..., 0 weight..., 27...

@majnas glad to hear it’s stable after updating—if everything continues to run cleanly, feel free to close this; to keep things reproducible please pin your working `ultralytics` version in your...

@majnas thanks for confirming—since `yolo checks` still shows `opencv-python ❌`, I’d install it (it’s a core dependency for image I/O/augmentations) and pin your working `ultralytics>=8.4.6` environment for reproducibility (see the...

@majnas perfect—`yolo checks` is now clean (including `opencv-python`), and if YOLO26x-seg is stable on `ultralytics 8.4.9` you should be all set; I’d pin this working environment for reproducibility (e.g. save...

Hi @AqsaM1, The raw output for the YOLOv8 segmentation model indeed includes a tensor and a tuple. The tensor shape (3, 48, 8400) indicates 3 images, 48 channels (12 classes...

You don’t need a custom `Trainer` just to run multispectral segmentation: set `channels: N` in your dataset YAML and store your images as multi-page `.tif/.tiff` in `(C, H, W)` order...