Skip to content

All In One

yolo is a packaged entry point that includes training, validation, and inference tasks. The following sections break down each operation and show how to import and call these functions directly.

Train Model

python -m yolo task=train

Common arguments:

Argument Type Description
name str Experiment name
model str Model backbone: v9-c, v7, v9-e, etc.
task.data.dataloader_workers int Number of CPU workers (num_workers)
out_path Path Output path for models and logs
weight Path \| bool \| None Pre-trained weights path. False = scratch, None = default
use_wandb bool Enable Weights & Biases tracking
use_tensorboard bool Enable TensorBoard logging
image_size int \| [int, int] Input image size
+quiet bool Disable all output (optional)
task.epoch int Total training epochs
task.data.batch_size int Batch size

Example

python -m yolo task=train task.data.batch_size=12 image_size=1280

Multi-GPU Training with DDP

For multi-GPU training using Distributed Data Parallel (DDP), replace python with torchrun --nproc_per_node=[GPU_NUM]:

torchrun --nproc_per_node=2 -m yolo task=train device=[0,1]

torchrun --nproc_per_node=2 -m yolo task=train device=\[0,1\]

Training on a Custom Dataset

Example dataset config (yolo/config/dataset/dev.yaml):

path: data/dev
train: train
validation: val

class_num: 80
class_list: ['Person', 'Bicycle', 'Car', ...]

auto_download:

Config fields:

Field Type Description
path str Path to the dataset
train, validation str Directory names under /images (and /labels/ for txt labels)
class_num int Number of dataset classes
class_list List[str] Optional class names for visualizing bounding boxes
auto_download dict Optional: auto-download configuration

Expected dataset structure:

DataSetName/
├── annotations/
│   ├── train_json_name.json
│   └── val_json_name.json
├── labels/
│   ├── train/
│   │   ├── AnyLabelName.txt
│   │   └── ...
│   └── validation/
│       └── ...
└── images/
    ├── train/
    │   ├── AnyImageNameN.{png,jpg,jpeg}
    │   └── ...
    └── validation/
        └── ...

Validation Model

During training this runs automatically. Run manually to generate a JSON file of predictions for a validation dataset. If the set includes JSON annotations, pycocotools evaluation runs automatically.

Common arguments:

Argument Type Description
task.nms.min_confidence float Minimum prediction confidence
task.nms.min_iou float Minimum IoU threshold for NMS

Example

python -m yolo task=validation task.nms.min_iou=0.9

yolo task=validation task.nms.min_iou=0.9

Model Inference

Note

Do not override dataset — the model requires class_num from it. If classes have names, provide class_list.

Common arguments:

Argument Type Description
task.fast_inference str onnx, trt, deploy, or None. deploy detaches the auxiliary head
task.data.source str \| Path \| int Webcam ID, image folder, video/image path
task.nms.min_confidence float Minimum prediction confidence
task.nms.min_iou float Minimum IoU threshold for NMS

Example

python -m yolo model=v9-m task.nms.min_confidence=0.1 task.data.source=0 task.fast_inference=onnx

yolo model=v9-m task.nms.min_confidence=0.1 task.data.source=0 task.fast_inference=onnx