app.py

import gradio as gr
import os
import re
import torch
from utils import VideoReader, FaceExtractor, confident_strategy, predict_on_video
from training.zoo.classifiers import DeepFakeClassifier

def detect(video):
    # Load model
    model = DeepFakeClassifier(encoder="tf_efficientnet_b7")
    path = os.path.join('weights', 'final_111_DeepFakeClassifier_tf_efficientnet_b7_ns_0_36')
    checkpoint = torch.load(path, map_location="cpu")
    state_dict = checkpoint.get("state_dict", checkpoint)
    model.load_state_dict({re.sub("^module.", "", k): v for k, v in state_dict.items()}, strict=True)
    model.eval()
    del checkpoint
    models = [model.float()]

    # Setting Video
    frames_per_video = 32
    video_reader = VideoReader()
    video_read_fn = lambda x: video_reader.read_frames(x, num_frames=frames_per_video)
    face_extractor = FaceExtractor(video_read_fn)
    input_size = 380
    strategy = confident_strategy

    # Predict
    pred = predict_on_video(
        face_extractor=face_extractor,
        video=video,
        batch_size=frames_per_video,
        input_size=input_size,
        models=models,
        strategy=strategy
    )
    prob = {'Fake': float(pred), 'Real': float(1 - pred)}
    return prob

gr_inputs = gr.Video(format='mp4', source='upload')
gr_outputs = gr.Label()
gr_ex = [
    [os.path.join(os.path.dirname(__file__),"sample/sample1.mp4")],
    [os.path.join(os.path.dirname(__file__),"sample/sample2.mp4")],
]
iface = gr.Interface(
    fn=detect,
    inputs=gr_inputs,
    outputs=gr_outputs,
    examples=gr_ex,
)
 
iface.launch()