# Folders

*Estimated read time: 2 minutes*

You can run inference over a whole directory of images using `model.predict_dir(...)`.

## Signature

{% code overflow="wrap" %}

```python
degirum.model.Model.predict_dir(
    path, *, recursive=False,
    extensions=['.jpg', '.jpeg', '.png', '.bmp']
)
```

{% endcode %}

* **What it does**: runs the full inference lifecycle for all files in a folder that match the given extensions, yielding one `InferenceResults` per image
* **Returns**: an iterator (`Iterator[InferenceResults]`), so you can use it directly in a `for` loop
* **Limitations**: supports only single-input models

## Common setup (used in all cases)

{% code overflow="wrap" %}

```python
from pathlib import Path
from degirum_tools import ModelSpec

# Configure & load once
model_spec = ModelSpec(
    model_name="yolov8n_coco--640x640_quant_axelera_metis_1",
    zoo_url="degirum/axelera",
    inference_host_address="@local",  # or "@cloud" / server URL
    model_properties={"device_type": ["AXELERA/METIS"]},
)
model = model_spec.load_model()
```

{% endcode %}

{% hint style="info" %}
Most results include `result.image_overlay` (NumPy BGR), which you can save for quick visualization.
{% endhint %}

## Basic folder (non-recursive, default extensions)

Use when you have a flat folder of standard image formats.

{% code overflow="wrap" %}

```python
in_dir = Path("images")  # folder with .jpg/.jpeg/.png/.bmp

for result in model.predict_dir(str(in_dir)):
    # result contains detections & convenience fields like image_overlay
    print(result)  # optional: inspect structured output
```

{% endcode %}

## Nested folders (recursive walk)

Use when working with datasets split across subdirectories (e.g., class-based folders).

{% code overflow="wrap" %}

```python
in_dir = "datasets/val"  # root directory with subfolders

for result in model.predict_dir(in_dir, recursive=True):
    # handle/save each result as it arrives
    pass
```

{% endcode %}

{% hint style="info" %}

* **Iterator-friendly**: `predict_dir` streams results; you don’t load everything into memory.
* **Filtering**: Use `extensions=` (a string or list) to limit which files are processed.
* **Reproducibility**: Keep input folders immutable and write outputs to a separate directory.
* **Colorspace**: Overlay images are in BGR (OpenCV-ready). Convert to RGB on save/view if needed.
  {% endhint %}

## **Minimal example**

{% code overflow="wrap" %}

```python
for result in model.predict_dir("./some_path"):
    print(result)
```

{% endcode %}