Dataset Adapters

Dataset adapters provide calibration and validation data to the compiler and accuracy measurement tools. Each adapter class corresponds to a task category and a dataset format.

See Deploy Custom Weights for how adapters are used in pipeline YAML files.

---

ObjDataAdapter​

For object detection models. Supports COCO 2014/2017 and custom datasets in YOLO/COCO JSON formats.

Definition: $AXELERA_FRAMEWORK/ax_datasets/objdataadapter.py

Field	Type	Description
data_dir_name	string	Dataset directory name, relative to the data root (default: data/)
label_type	string	Label format: YOLOv8, COCO JSON, COCO2017, COCO2014
ultralytics_data_yaml	string	Path to an Ultralytics data.yaml, relative to data_dir_name. Auto-generates cal/val/labels. Cannot be used with cal_data, val_data, or labels.
cal_data	string	Calibration data: directory with images or text file listing image paths
val_data	string	Validation data: directory with images or text file listing image paths
labels	string	Labels file (YAML or .names), relative to data_dir_name
repr_imgs_dir_path	string	Absolute path to a directory of representative calibration images. Alternative to cal_data.
download_year	string	COCO dataset year: "2014" or "2017" (for built-in COCO support)
format	string	COCO class format: default COCO-80, or "coco91" / "coco91-with-bg"
output_format	string	Bounding box format: "xyxy" (default), "xywh", "ltwh"
is_label_image_same_dir	bool	True if images and labels are in the same directory (default: False)
[val|cal]_img_dir_name	string	Override image directory for val or cal, relative to data_dir_name

KptDataAdapter​

Subclass of ObjDataAdapter for YOLO keypoint detection models. Uses COCO 2017 pose dataset.

Inherits all ObjDataAdapter fields. No additional fields.

SegDataAdapter​

Subclass of ObjDataAdapter for YOLO instance segmentation models. Uses COCO 2017 segmentation dataset.

Inherits all ObjDataAdapter fields, plus:

Field	Type	Default	Description
is_mask_overlap	bool	True	Whether masks are overlapped during evaluation
eval_with_letterbox	bool	True	Whether to use letterbox resize during mask evaluation
mask_size	tuple	(160, 160)	Mask dimensions (height, width) during evaluation

---

TorchvisionDataAdapter​

For classification models based on torchvision. Supports ImageNet-style datasets and standard torchvision datasets.

Definition: $AXELERA_FRAMEWORK/ax_datasets/torchvision.py

Field	Type	Default	Description
dataset_name	string	"ImageFolder"	torchvision dataset class to use

Beyond dataset_name, fields map to the corresponding torchvision dataset class arguments. Supported dataset classes:

ImageFolder​

For custom datasets organized as root/class_name/image.jpg.

Field	Type	Default
split	string	"train"
[val|cal]_data	string (required)	—
[val|cal]_index_pkl	string	None
is_one_indexed	bool	False

ImageNet​

Field	Type	Default
split	string	"train"

MNIST / CIFAR10​

Field	Type	Default
train	bool	True
download	bool	True

VOCDetection​

Field	Type	Default
year	string	"2011"
image_set	string	"train"
download	bool	False

LFWPairs / LFWPeople​

Field	Type	Default
image_set	string	"funneled"
download	bool	True
split	string	"test"

Caltech101​

Field	Type	Default
download	bool	False

---

See also​

• Deploy Custom Weights — how to configure adapters in a YAML file
• Measure Accuracy — running validation with a dataset source
• Model Zoo — pre-built models and their dataset types