File size: 4,681 Bytes
b5ece2c
 
3b69d63
b5ece2c
 
 
 
 
3b69d63
 
 
 
 
 
d5e4cb5
3b69d63
b5ece2c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
b110889
b5ece2c
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
import os
import sys
import subprocess
import argparse
import torch
import gradio as gr
from huggingface_hub import hf_hub_download

# Install local packages (face_detection, face_alignment) at runtime
# Cannot be in requirements.txt because app files are copied after pip install in Docker build
_app_dir = os.path.dirname(os.path.abspath(__file__))
for _pkg in ["face_detection", "face_alignment"]:
    _pkg_path = os.path.join(_app_dir, _pkg)
    if os.path.exists(_pkg_path):
        subprocess.run([sys.executable, "-m", "pip", "install", _pkg_path], check=True)

# Add auto_avsr to Python path
sys.path.insert(0, os.path.join(os.path.dirname(__file__), "auto_avsr"))

from lightning import ModelModule
from datamodule.transforms import VideoTransform
from preparation.detectors.retinaface.detector import LandmarksDetector
from preparation.detectors.retinaface.video_process import VideoProcess

# Download VSR model from HuggingFace Hub at startup
print("Downloading VSR model from HuggingFace Hub...")
model_path = hf_hub_download(
    repo_id="okregent/visnet-model",
    filename="vsr_trlrs2lrs3vox2avsp_base.pth",
)
print(f"Model ready at: {model_path}")

# Initialise args (model expects an argparse namespace)
parser = argparse.ArgumentParser()
args, _ = parser.parse_known_args(args=[])
setattr(args, "modality", "video")


class InferencePipeline(torch.nn.Module):
    SEGMENT_DURATION = 5   # seconds — matches the LRS3 training clip length
    MIN_FRAMES = 10        # skip segments shorter than this

    def __init__(self, args, ckpt_path, detector="retinaface"):
        super().__init__()
        self.modality = args.modality

        device = "cuda:0" if torch.cuda.is_available() else "cpu"
        self.landmarks_detector = LandmarksDetector(device=device)
        self.video_process = VideoProcess(convert_gray=False)
        self.video_transform = VideoTransform(subset="test")

        ckpt = torch.load(ckpt_path, map_location=lambda storage, loc: storage)
        self.modelmodule = ModelModule(args)
        self.modelmodule.model.load_state_dict(ckpt)
        self.modelmodule.eval()

    def load_video(self, data_filename):
        import torchvision
        frames, _, info = torchvision.io.read_video(data_filename, pts_unit="sec")
        fps = info.get("video_fps", 25.0)
        return frames.numpy(), fps

    def _process_segment(self, segment_frames):
        landmarks = self.landmarks_detector(segment_frames)
        processed = self.video_process(segment_frames, landmarks)
        if processed is None:
            return ""

        video_tensor = torch.tensor(processed)
        video_tensor = video_tensor.permute((0, 3, 1, 2))
        video_tensor = self.video_transform(video_tensor)

        with torch.no_grad():
            transcript = self.modelmodule(video_tensor)

        return transcript.strip()

    def forward(self, data_filename):
        data_filename = os.path.abspath(data_filename)
        assert os.path.isfile(data_filename), f"File not found: {data_filename}"

        video, fps = self.load_video(data_filename)
        segment_size = int(fps * self.SEGMENT_DURATION)
        total_frames = len(video)
        transcripts = []

        for start in range(0, total_frames, segment_size):
            end = min(start + segment_size, total_frames)
            segment = video[start:end]
            if len(segment) < self.MIN_FRAMES:
                continue
            result = self._process_segment(segment)
            if result:
                transcripts.append(result)

        return " ".join(transcripts)


# Load model once at startup
pipeline = InferencePipeline(args, model_path)


def transcribe(video_path):
    if video_path is None:
        return "Please upload a video file."
    try:
        result = pipeline(video_path)
        if not result:
            return (
                "No speech detected. Make sure the video clearly shows "
                "a speaker's face (front-facing, good lighting)."
            )
        return result
    except Exception as e:
        return f"Error: {str(e)}"


demo = gr.Interface(
    fn=transcribe,
    inputs=gr.Video(label="Upload Video (mp4 / avi / mov, max 100 MB)"),
    outputs=gr.Textbox(
        label="Transcription",
        lines=6,
    ),
    title="VisNet — Visual Speech Recognition",
    description=(
        "Upload a video to transcribe speech from lip movements — **no audio required**.\n\n"
        "**Tips for best results:** front-facing camera, clear face visibility, good lighting.\n\n"
        "⚠️ Running on CPU — inference may take several minutes for longer videos."
    ),

)

demo.queue()
demo.launch()