model-inference 0.0.5

package for inferencing models developed at phosphene

Torch CUDA install

If the identity_clustering package is installing the CPU version, you can install the right CUDA versions by manually installing it or running the following command:

python -m model_inference.torch_install

Sphinx Documentation

To view Sphinx documentation, clone this repository and run the following commands

cd pkg-faceswap-inference

#install sphinx and a theme.
pip install sphinx
pip install sphinx-rtd-theme

# generate docs
sphinx-build -b html docs/ build/

open the build folder there you will find index.html , open it with a browser

---

Inference Class Documentation

The Inference class provides functionalities for inferencing video files with models. It includes utilities for timing and tracking nested function calls to aid in performance analysis.

Table of Contents

• Class Attributes
• Methods
• Usage
• License

---

Class Attributes

• device (str): Specifies the device (cpu or cuda) used for computation.
• shape (tuple): Dimensions for resizing clustered faces, default is (224, 224).
• classes (list): The categories the model can predict, such as ["Real", "Fake"].
• timings (dict): Stores timing data for functions with nested call relationships.
• _clusters (None or dict): Stores clusters of faces post-clustering.

---

Methods

__init__(self, device: str, shape=(224, 224))

• Initializes the Inference class with specified device and shape attributes.

generate_video_data(self, video_path: str, print_timings=True)

• Processes a video to detect, crop, and cluster faces, and converts them to RGB format.

get_data(self, video_path: str, print_timings=True)

• Retrieves essential data for frames, bounding boxes, images, FPS, and clustered identities from a video.

get_predictions(self, model, images: torch.Tensor, device='cuda')

• Runs predictions on clustered face images using a model and returns logits and labels.

__cvt_to_rgb(self, faces: tuple) -> torch.Tensor

• Converts images from BGR to RGB format.

__plot_images_grid(self, tensor: torch.Tensor, images_per_row=4)

• Plots a grid of images from a 4D tensor.

__print_result(self, result: dict, image_data: List[torch.Tensor])

• Displays results, including images, based on the model’s predictions.

__print_timings(self, timings: dict)

• Outputs timing information for functions with nested call relationships.

__create_sequence_dict(self, identity_data)

• Organizes identity information into a sequence dictionary.

__draw_bounding_boxes(self, video_path: str, sequence_dict, result_video_path: str)

• Draws bounding boxes on faces in the video and saves the processed output.

Decorator timeit(func)

• Times functions and records their nested relationships in the timings attribute.

---

Usage

Here's a simple example of using the Inference class:

from inference import Inference
import torch
from models.models_list import ModelList

# Initialize Inference class
device = "cuda" if torch.cuda.is_available() else "cpu"
inference = Inference(device=device)

# Load your model from ModelList
model = ModelList().load_model("face_detection_model", device)

# Process video
video_path = "/path/to/video.mp4"
output_data, num_clusters = inference.generate_video_data(video_path)

# Get predictions for each cluster
for cluster_images in output_data:
    predictions = inference.get_predictions(model, cluster_images, device)
    inference.__print_result(predictions, cluster_images)

License

This project is licensed under the MIT License. See the LICENSE file for details.

---