FaceNet PyTorch

2.6.0 · active · verified Thu Apr 16

facenet-pytorch provides pretrained PyTorch models for face detection (MTCNN) and facial recognition (InceptionResnetV1). It simplifies the process of integrating robust face analysis capabilities into Python applications, offering an easy-to-use API for tasks like detecting faces, extracting facial embeddings, and preparing faces for classification. The library is actively maintained, with regular updates to support newer PyTorch versions and address community feedback.

Common errors

```
ModuleNotFoundError: No module named 'facenet_pytorch'
```
cause The `facenet-pytorch` package is not installed in the active Python environment or there's a typo in the import statement.

fix Ensure the package is correctly installed: `pip install facenet-pytorch`. Verify your virtual environment is activated and the import path is `from facenet_pytorch import ...`.
```
ERROR: Cannot install facenet-pytorch because these package versions have conflicting dependencies.
```
cause Dependency conflicts, most commonly with `torch` or `numpy` versions required by other installed packages.

fix Use a fresh virtual environment. Install `torch` and `torchvision` first with specific compatible versions (e.g., `pip install torch==2.2.0 torchvision==0.17.0`), then install `facenet-pytorch`. Consult `facenet-pytorch`'s PyPI page or GitHub for exact dependency ranges for your version.
```
RuntimeError: No faces detected or image is empty
```
cause The `MTCNN` model failed to detect any faces in the provided image, or the image itself was not valid/empty.

fix Verify the input image contains detectable faces and is properly loaded. Check `MTCNN` parameters like `min_face_size`, `thresholds`, and `image_size` if faces are very small or unusual. Inspect the image processing pipeline to ensure the image is not empty or corrupted. Handle `None` return from `mtcnn()` gracefully.

Warnings

breaking Version 2.6.0 of `facenet-pytorch` requires specific PyTorch (`torch`) and torchvision versions. Installing with incompatible versions can lead to `pip` dependency resolution errors or runtime issues.
Fix: Ensure your `torch` version is between 2.2.0 and 2.3.0 (`<2.3.0,>=2.2.0`) and `torchvision` is between 0.17.0 and 0.18.0 (`<0.18.0,>=0.17.0`). Consider using a virtual environment and installing `torch` and `torchvision` first, then `facenet-pytorch`.
gotcha Older versions of `facenet-pytorch` (prior to v2.5.3) could throw an `MTCNN module error` when no face was found in an image, leading to unexpected crashes.
Fix: Upgrade to `facenet-pytorch` version 2.5.3 or newer, where this issue was addressed.
gotcha The library's `MTCNN` and `InceptionResnetV1` models expect `PIL.Image` objects as input. Passing other image formats (e.g., NumPy arrays without conversion) can lead to errors.
Fix: Always convert your image data to a `PIL.Image` object before passing it to `MTCNN` or `InceptionResnetV1`. E.g., `Image.fromarray(numpy_array)` or `Image.open(image_path)`.
deprecated `numpy` is throwing deprecation warnings for creating `ndarray` from nested sequences due to `facenet-pytorch`'s internal usage. This doesn't break functionality but indicates future incompatibility.
Fix: While this is often an upstream issue in `facenet-pytorch` itself or its dependencies, users should keep their `numpy` version updated and monitor for future releases that address this warning.

Install

pip install facenet-pytorch Install stable release

Imports

MTCNN
wrong:
```
import facenet_pytorch.MTCNN
```
correct:
```
from facenet_pytorch import MTCNN
```
MTCNN is a class within the facenet_pytorch package, not a top-level module.
InceptionResnetV1
wrong:
```
import facenet_pytorch.InceptionResnetV1
```
correct:
```
from facenet_pytorch import InceptionResnetV1
```
InceptionResnetV1 is a class within the facenet_pytorch package, not a top-level module.

Quickstart

This quickstart demonstrates how to use `facenet-pytorch` to detect a face in an image using `MTCNN` and then compute its 512-dimensional embedding using `InceptionResnetV1`. It sets up a PyTorch device (GPU if available, otherwise CPU) and initializes both models. It includes a placeholder for image loading and handles cases where no face is detected.

import torch
from facenet_pytorch import MTCNN, InceptionResnetV1
from PIL import Image
import os

# Set device for GPU if available, else CPU
device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
print(f'Running on device: {device}')

# Initialize MTCNN for face detection
mtcnn = MTCNN(
    image_size=160,
    margin=0,
    min_face_size=20,
    thresholds=[0.6, 0.7, 0.7],
    factor=0.709,
    post_process=True,
    device=device
)

# Initialize InceptionResnetV1 for face recognition
resnet = InceptionResnetV1(pretrained='vggface2').eval().to(device)

# Create a dummy image for demonstration (replace with your image path)
# In a real scenario, load an image from disk or URL
# Example: img = Image.open('path/to/your/image.jpg').convert('RGB')

# For a runnable example, we create a blank image
try:
    # Attempt to load a real image for better demo, if it exists
    dummy_image_path = os.path.join(os.path.dirname(__file__), 'dummy_face.jpg')
    if os.path.exists(dummy_image_path):
        img = Image.open(dummy_image_path).convert('RGB')
    else:
        # Create a blank image if no dummy_face.jpg is found
        img = Image.new('RGB', (250, 250), color = 'red')
        print("No 'dummy_face.jpg' found. Using a blank red image. Face detection will likely fail.")
except Exception as e:
    img = Image.new('RGB', (250, 250), color = 'red')
    print(f"Could not load image, creating a blank red image. Error: {e}")

# Detect faces
img_cropped = mtcnn(img)

if img_cropped is not None:
    # Calculate face embedding
    img_embedding = resnet(img_cropped.unsqueeze(0)).detach().cpu()
    print("Face detected and embedding calculated.")
    print(f"Embedding shape: {img_embedding.shape}")
else:
    print("No face detected.")

view raw JSON →