train_rf_detr_seg

Train RF-DETR instance segmentation models.

🚀 Use with Ikomia API

1. Install Ikomia API

We strongly recommend using a virtual environment. If you're not sure where to start, we offer a tutorial here.

pip install ikomia

2. Create your workflow

from ikomia.dataprocess.workflow import Workflow

# Init your workflow
wf = Workflow()

# Add COCO instance segmentation dataset loader
coco = wf.add_task(name="dataset_coco")

coco.set_parameters({
    "json_file": "path/to/instances_train.json",
    "image_folder": "path/to/images",
    "task": "instance_segmentation",
})

# Add training algorithm
train = wf.add_task(name="train_rf_detr_seg", auto_connect=True)

# Launch training
wf.run()

☀️ Use with Ikomia Studio

Ikomia Studio offers a friendly UI with the same features as the API.

• If you haven't started using Ikomia Studio yet, download and install it from this page.
• For additional guidance on getting started with Ikomia Studio, check out this blog post.

📝 Set algorithm parameters

• model_name (str) - default rf-detr-seg-medium: Name of the RF-DETR segmentation model. Available models:
  • rf-detr-seg-nano
  • rf-detr-seg-small
  • rf-detr-seg-medium
  • rf-detr-seg-base
  • rf-detr-seg-large
• model_weight_file (str, optional): Path to custom starting weights.
• dataset_folder (str): Folder where the Ikomia dataset is converted to COCO format.
• batch_size (int) - default 2: Number of samples processed before the model is updated.
• grad_accum_steps (int) - default 4: Number of gradient accumulation steps.
• epochs (int) - default 100: Number of complete passes through the training dataset.
• dataset_split_ratio (float) - default 0.9: Ratio used to split the dataset into train and validation sets.
• input_size (int) - default 576: Training image size. It is adjusted to the selected RF-DETR segmentation model block size when needed.
• workers (int) - default 0: Number of worker threads for data loading. Use 0 on Windows if multiprocessing is unstable.
• weight_decay (float) - default 1e-4: Weight decay regularization.
• lr (float) - default 1e-4: Learning rate.
• lr_encoder (float) - default 1.5e-4: Encoder learning rate.
• output_folder (str): Folder where training runs, weights, and class_names.yaml are saved.
• early_stopping (bool) - default False: Stored for compatibility with the UI and future RF-DETR training options.
• early_stopping_patience (int) - default 10: Stored with the training configuration when early stopping is enabled.

The training dataset must contain polygon instance masks. For Ikomia datasets, those masks are read from segmentation_poly and exported to COCO segmentation annotations.

Pretrained RF-DETR segmentation weights are cached in train_rf_detr_seg/weights, following the same local cache pattern as infer_rf_detr.

Parameters should be in strings format when added to the dictionary.

from ikomia.dataprocess.workflow import Workflow

# Init your workflow
wf = Workflow()

# Add COCO instance segmentation dataset loader
coco = wf.add_task(name="dataset_coco")

coco.set_parameters({
    "json_file": "path/to/instances_train.json",
    "image_folder": "path/to/images",
    "task": "instance_segmentation",
})

# Add training algorithm
train = wf.add_task(name="train_rf_detr_seg", auto_connect=True)

train.set_parameters({
    "model_name": "rf-detr-seg-medium",
    "epochs": "100",
    "batch_size": "2",
    "grad_accum_steps": "4",
    "input_size": "576",
    "dataset_split_ratio": "0.9",
    "workers": "0",
    "weight_decay": "1e-4",
    "lr": "1e-4",
    "lr_encoder": "1.5e-4",
    "output_folder": "path/to/output/folder",
})

# Launch training
wf.run()

After training, use the generated weights and class_names.yaml with infer_rf_detr_seg.