Model properties

See what you can tune on your model—and why it matters. This page introduces the model_properties field and shows how to inspect, change, and group model settings by task.

Estimated read time: 5 minutes

This section focuses on the model_properties field in ModelSpec (and the matching writable properties on the model after loading). It shows what you can change on a DeGirum model and why—without drowning you in code.

First, quickly inspect your model to see which attributes are writable on your build. Then jump into focused subpages:

• Device selection: pick hardware targets, batching, queues, and timeouts

• Input preprocessing: control resize, crop, pad, and colorspace handling

• Output postprocessing: adjust confidence thresholds, NMS, class filters, and caps

• Image overlay: tune how detections render in demos or reviews

Two ways to specify model properties

• During load (stable defaults): pass durable choices via ModelSpec(model_properties={...}). Great for reproducible runs and team consistency.

• After load (live tuning): set model.<property> at runtime to explore and iterate.

Use ModelSpec for hardware and standard preprocessing choices.

Use the live model for tuning outputs, overlays, and experiments.

Inspect your model

After loading a model, run:

# What attributes & methods exist?
dir(model)

You'll see output like this:

[
    "__call__",
    "custom_postprocessor",
    "device_type",
    "devices_available",
    "devices_selected",
    "eager_batch_size",
    "extra_device_params",
    "frame_queue_depth",
    "image_backend",
    "inference_results_type",
    "inference_timeout_s",
    "input_crop_percentage",
    "input_image_format",
    "input_letterbox_fill_color",
    "input_numpy_colorspace",
    "input_pad_method",
    "input_resize_method",
    "input_shape",
    "label_dictionary",
    "measure_time",
    "model_info",
    "non_blocking_batch_predict",
    "output_class_set",
    "output_confidence_threshold",
    "output_max_classes_per_detection",
    "output_max_detections",
    "output_max_detections_per_class",
    "output_nms_threshold",
    "output_pose_threshold",
    "output_postprocess_type",
    "output_top_k",
    "output_use_regular_nms",
    "overlay_alpha",
    "overlay_blur",
    "overlay_color",
    "overlay_font_scale",
    "overlay_line_width",
    "overlay_show_labels",
    "overlay_show_probabilities",
    "predict",
    "predict_batch",
    "predict_dir",
    "reset_time_stats",
    "save_model_image",
    "supported_device_types",
    "time_stats",
]

To check which ones are writable, use this helper:

def list_properties_with_setters(obj):
    ro, rw = [], []
    cls = obj.__class__
    for name in dir(cls):
        attr = getattr(cls, name, None)
        if isinstance(attr, property):
            (ro if attr.fset is None else rw).append(name)
    return ro, rw


ro, rw = list_properties_with_setters(model)
print("READ‑ONLY:", sorted(ro))
print("WRITABLE:", sorted(rw))

Running this function on a loaded model will output:

• READ-ONLY: properties you can inspect but can't modify

• WRITABLE: properties you can update at runtime

READ‑ONLY: [
    "devices_available",
    "label_dictionary",
    "model_info",
    "supported_device_types",
]
WRITABLE: [
    "custom_postprocessor",
    "device_type",
    "devices_selected",
    "eager_batch_size",
    "extra_device_params",
    "frame_queue_depth",
    "image_backend",
    "inference_results_type",
    "inference_timeout_s",
    "input_crop_percentage",
    "input_image_format",
    "input_letterbox_fill_color",
    "input_numpy_colorspace",
    "input_pad_method",
    "input_resize_method",
    "input_shape",
    "measure_time",
    "non_blocking_batch_predict",
    "output_class_set",
    "output_confidence_threshold",
    "output_max_classes_per_detection",
    "output_max_detections",
    "output_max_detections_per_class",
    "output_nms_threshold",
    "output_pose_threshold",
    "output_postprocess_type",
    "output_top_k",
    "output_use_regular_nms",
    "overlay_alpha",
    "overlay_blur",
    "overlay_color",
    "overlay_font_scale",
    "overlay_line_width",
    "overlay_show_labels",
    "overlay_show_probabilities",
    "save_model_image",
]

Treat this output as ground truth for your build. Use it to verify assumptions before wiring up UI elements or CLI options based on these properties.

What’s changeable—by category

Use this section as a reference. The most common writable properties are grouped by use case. Confirm actual writability using your own read-only (ro) / read/write (rw) output.

Device selection & runtime

• Target runtime(s): device_type (str | list[str]) – set hardware preferences; if a list, the first available is selected

• Choose indices: devices_selected (list[int]) – which device IDs to run on

• Discoveries (inspect): supported_device_types (ro), devices_available (ro)

• Latency/throughput knobs: eager_batch_size, frame_queue_depth, inference_timeout_s, non_blocking_batch_predict

• Runtime-specific knobs: extra_device_params (dict) – pass vendor settings such as {"HAILO_BATCH_SIZE": 8}; overrides eager_batch_size on Hailo runtimes

Input preprocessing

• Resize strategy: input_pad_method – "stretch", "letterbox" (default), "crop-first", "crop-last"

• Interpolation: input_resize_method – "bilinear" (default) "nearest", "area", "bicubic", "lanczos"

• Cropping control: input_crop_percentage – for crop‑first/crop‑last modes

• Pad color: input_letterbox_fill_color – defaults to black; use (114,114,114) for neutral gray

• Colorspace/format: input_numpy_colorspace (RGB/BGR), input_image_format

• Image library: image_backend – "opencv" (default), "pil"

• Extras: save_model_image

• Model‑fixed (often ro): input_shape

Output postprocessing

• Filter thresholds: output_confidence_threshold, output_nms_threshold

• Caps: output_max_detections, output_max_detections_per_class, output_max_classes_per_detection, output_top_k

• Focus filters: output_class_set

• Advanced controls: output_use_regular_nms, output_postprocess_type, output_pose_threshold, inference_results_type, custom_postprocessor

Image overlay (rendering-only)

• Toggles: overlay_show_labels, overlay_show_probabilities

• Styling: overlay_line_width, overlay_font_scale, overlay_alpha, overlay_color, overlay_blur

Time statistics

• Enable/disable: measure_time

• Read stats: time_stats (ro)

• Reset: reset_time_stats

Predict APIs (reference-only)

• Sync: predict, __call__

• Batch: predict_batch

• Folders: predict_dir

You’ll also see helpful read-only properties like label_dictionary, and model_info.