NetfLips/inference_av2av.py at main · Prometheus-AI-3team/NetfLips · GitHub

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
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
import os
import argparse
import numpy as np
import torch
import torch.nn.functional as F

from fairseq import utils
from fairseq_cli.generate import get_symbols_to_strip_from_output

from av2unit.inference import load_model as load_av2unit_model
from unit2unit.inference import load_model as load_unit2unit_model
from unit2av.inference import load_model as load_unit2av_model, load_speaker_encoder_model

from util import process_units, extract_audio_from_video, save_video

class AVSpeechToAVSpeechPipeline:
    def __init__(self,
        av2unit_model, av2unit_task,
        unit2unit_task, unit2unit_generator,
        unit2av_model, speaker_encoder,
        use_cuda=False
    ):
        self.av2unit_model = av2unit_model
        self.av2unit_task = av2unit_task
        self.unit2unit_task = unit2unit_task
        self.unit2unit_generator = unit2unit_generator
        self.unit2av_model = unit2av_model
        self.speaker_encoder = speaker_encoder
        self.use_cuda = use_cuda

    def process_av2unit(self, lip_video_path, audio_path):
        task = self.av2unit_task
        video_feats, audio_feats = task.dataset.load_feature((lip_video_path, audio_path))
        audio_feats, video_feats = torch.from_numpy(audio_feats.astype(np.float32)) if audio_feats is not None else None, torch.from_numpy(video_feats.astype(np.float32)) if video_feats is not None else None
        if task.dataset.normalize and 'audio' in task.dataset.modalities:
            with torch.no_grad():
                audio_feats = F.layer_norm(audio_feats, audio_feats.shape[1:])

        collated_audios, _, _ = task.dataset.collater_audio([audio_feats], len(audio_feats))
        collated_videos, _, _ = task.dataset.collater_audio([video_feats], len(video_feats))

        sample = {"source": {
            "audio": collated_audios, "video": collated_videos,
        }}
        sample = utils.move_to_cuda(sample) if self.use_cuda else sample

        pred = task.inference(
            self.av2unit_model,
            sample,
        )
        pred_str = task.dictionaries[0].string(pred.int().cpu())

        return pred_str

    def process_unit2unit(self, unit):
        task = self.unit2unit_task
        unit = list(map(int, unit.strip().split()))
        unit = task.source_dictionary.encode_line(
            " ".join(map(lambda x: str(x), process_units(unit, reduce=True))),
            add_if_not_exist=False,
            append_eos=True,
        ).long()
        unit = torch.cat([
            unit.new([task.source_dictionary.bos()]),
            unit,
            unit.new([task.source_dictionary.index("[{}]".format(task.source_language))])
        ])

        sample = {"net_input": {
            "src_tokens": torch.LongTensor(unit).view(1,-1),
        }}
        sample = utils.move_to_cuda(sample) if self.use_cuda else sample

        pred = task.inference_step(
            self.unit2unit_generator,
            None,
            sample,
        )[0][0]

        pred_str = task.target_dictionary.string(
            pred["tokens"].int().cpu(),
            extra_symbols_to_ignore=get_symbols_to_strip_from_output(self.unit2unit_generator)
        )

        return pred_str

    def process_unit2av(self, unit, audio_path, video_path, bbox_path):
        unit = list(map(int, unit.strip().split()))

        sample = {
            "code": torch.LongTensor(unit).view(1,-1),
            "spkr": torch.from_numpy(self.speaker_encoder.get_embed(audio_path)).view(1,1,-1),
        }
        sample = utils.move_to_cuda(sample) if self.use_cuda else sample

        wav, video, full_video, bbox = self.unit2av_model(sample, video_path, bbox_path, dur_prediction=True)

        return wav, video, full_video, bbox

def main(args):
    use_cuda = torch.cuda.is_available() and not args.cpu

    av2unit_model, av2unit_task = load_av2unit_model(args.av2unit_path, args.modalities, use_cuda=use_cuda)
    unit2unit_task, unit2unit_generator = load_unit2unit_model(args.utut_path, args.src_lang, args.tgt_lang, use_cuda=use_cuda)
    cfg_path = os.path.join("unit2av", "config.json")
    unit2av_model = load_unit2av_model(args.unit2av_path, cfg_path, args.tgt_lang, use_cuda=use_cuda)
    speaker_encoder_model = load_speaker_encoder_model(os.path.join("unit2av", "encoder.pt"), use_cuda=use_cuda)

    pipeline = AVSpeechToAVSpeechPipeline(
        av2unit_model, av2unit_task,
        unit2unit_task, unit2unit_generator,
        unit2av_model, speaker_encoder_model,
        use_cuda=use_cuda
    )

    temp_audio_path = os.path.splitext(args.in_vid_path)[0]+".temp.wav"
    lip_video_path = os.path.splitext(args.in_vid_path)[0]+".lip.mp4"
    bbox_path = os.path.splitext(args.in_vid_path)[0]+".bbox.pkl"
    extract_audio_from_video(args.in_vid_path, temp_audio_path)

    src_unit = pipeline.process_av2unit(lip_video_path, temp_audio_path)
    tgt_unit = pipeline.process_unit2unit(src_unit)
    tgt_audio, tgt_video, full_video, bbox = pipeline.process_unit2av(tgt_unit, temp_audio_path, args.in_vid_path, bbox_path)

    save_video(tgt_audio, tgt_video, full_video, bbox, args.out_vid_path)

    os.remove(temp_audio_path)

def cli_main():
    parser = argparse.ArgumentParser()
    parser.add_argument(
        "--in-vid-path", type=str, required=True, help="File path of source video input"
    )
    parser.add_argument(
        "--out-vid-path", type=str, required=True, help="File path of translated video output"
    )
    parser.add_argument(
        "--src-lang", type=str, required=True,
        choices=["en","es","fr","it","pt"],
        help="source language"
    )
    parser.add_argument(
        "--tgt-lang", type=str, required=True,
        choices=["en","es","fr","it","pt"],
        help="target language"
    )
    parser.add_argument(
        "--modalities", type=str, default="audio,video", help="input modalities",
        choices=["audio,video","audio","video"],
    )
    parser.add_argument(
        "--av2unit-path", type=str, required=True, help="path to the mAV-HuBERT pre-trained model"
    )
    parser.add_argument(
        "--utut-path", type=str, required=True, help="path to the UTUT pre-trained model"
    )
    parser.add_argument(
        "--unit2av-path", type=str, required=True, help="path to the Unit AV Renderer"
    )
    parser.add_argument("--cpu", action="store_true", help="run on CPU")
    args = parser.parse_args()
    main(args)

if __name__ == "__main__":
    cli_main()