Spaces:
Running
on
Zero
Running
on
Zero
Upload stf_utils.py
Browse files- stf_utils.py +125 -0
stf_utils.py
ADDED
@@ -0,0 +1,125 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import os
|
3 |
+
from concurrent.futures import ThreadPoolExecutor
|
4 |
+
from pydub import AudioSegment
|
5 |
+
import cv2
|
6 |
+
from pathlib import Path
|
7 |
+
import subprocess
|
8 |
+
from pathlib import Path
|
9 |
+
import av
|
10 |
+
import imageio
|
11 |
+
import numpy as np
|
12 |
+
from rich.progress import track
|
13 |
+
from tqdm import tqdm
|
14 |
+
|
15 |
+
import stf_alternative
|
16 |
+
|
17 |
+
|
18 |
+
|
19 |
+
def exec_cmd(cmd):
|
20 |
+
subprocess.run(
|
21 |
+
cmd, shell=True, check=True, stdout=subprocess.PIPE, stderr=subprocess.STDOUT
|
22 |
+
)
|
23 |
+
|
24 |
+
|
25 |
+
def images2video(images, wfp, **kwargs):
|
26 |
+
fps = kwargs.get("fps", 24)
|
27 |
+
video_format = kwargs.get("format", "mp4") # default is mp4 format
|
28 |
+
codec = kwargs.get("codec", "libx264") # default is libx264 encoding
|
29 |
+
quality = kwargs.get("quality") # video quality
|
30 |
+
pixelformat = kwargs.get("pixelformat", "yuv420p") # video pixel format
|
31 |
+
image_mode = kwargs.get("image_mode", "rgb")
|
32 |
+
macro_block_size = kwargs.get("macro_block_size", 2)
|
33 |
+
ffmpeg_params = ["-crf", str(kwargs.get("crf", 18))]
|
34 |
+
|
35 |
+
writer = imageio.get_writer(
|
36 |
+
wfp,
|
37 |
+
fps=fps,
|
38 |
+
format=video_format,
|
39 |
+
codec=codec,
|
40 |
+
quality=quality,
|
41 |
+
ffmpeg_params=ffmpeg_params,
|
42 |
+
pixelformat=pixelformat,
|
43 |
+
macro_block_size=macro_block_size,
|
44 |
+
)
|
45 |
+
|
46 |
+
n = len(images)
|
47 |
+
for i in track(range(n), description="writing", transient=True):
|
48 |
+
if image_mode.lower() == "bgr":
|
49 |
+
writer.append_data(images[i][..., ::-1])
|
50 |
+
else:
|
51 |
+
writer.append_data(images[i])
|
52 |
+
|
53 |
+
writer.close()
|
54 |
+
|
55 |
+
# print(f':smiley: Dump to {wfp}\n', style="bold green")
|
56 |
+
print(f"Dump to {wfp}\n")
|
57 |
+
|
58 |
+
|
59 |
+
def merge_audio_video(video_fp, audio_fp, wfp):
|
60 |
+
if osp.exists(video_fp) and osp.exists(audio_fp):
|
61 |
+
cmd = f"ffmpeg -i {video_fp} -i {audio_fp} -c:v copy -c:a aac {wfp} -y"
|
62 |
+
exec_cmd(cmd)
|
63 |
+
print(f"merge {video_fp} and {audio_fp} to {wfp}")
|
64 |
+
else:
|
65 |
+
print(f"video_fp: {video_fp} or audio_fp: {audio_fp} not exists!")
|
66 |
+
|
67 |
+
|
68 |
+
|
69 |
+
|
70 |
+
class STFPipeline:
|
71 |
+
def __init__(self,
|
72 |
+
stf_path: str = "../stf/",
|
73 |
+
device: str = "cuda:0",
|
74 |
+
template_video_path: str = "templates/front_one_piece_dress_nodded_cut.webm",
|
75 |
+
config_path: str = "front_config.json",
|
76 |
+
checkpoint_path: str = "089.pth",
|
77 |
+
root_path: str = "works"
|
78 |
+
|
79 |
+
):
|
80 |
+
|
81 |
+
config_path = os.path.join(stf_path, config_path)
|
82 |
+
checkpoint_path = os.path.join(stf_path, checkpoint_path)
|
83 |
+
work_root_path = os.path.join(stf_path, root_path)
|
84 |
+
|
85 |
+
model = stf_alternative.create_model(
|
86 |
+
config_path=config_path,
|
87 |
+
checkpoint_path=checkpoint_path,
|
88 |
+
work_root_path=work_root_path,
|
89 |
+
device=device,
|
90 |
+
wavlm_path="microsoft/wavlm-large",
|
91 |
+
)
|
92 |
+
self.template = stf_alternative.Template(
|
93 |
+
model=model,
|
94 |
+
config_path=config_path,
|
95 |
+
template_video_path=template_video_path,
|
96 |
+
)
|
97 |
+
|
98 |
+
|
99 |
+
def execute(self, audio: str):
|
100 |
+
Path("dubbing").mkdir(exist_ok=True)
|
101 |
+
save_path = os.path.join("dubbing", Path(audio).stem+"--lip.mp4")
|
102 |
+
reader = iter(self.template._get_reader(num_skip_frames=0))
|
103 |
+
audio_segment = AudioSegment.from_file(audio)
|
104 |
+
pivot = 0
|
105 |
+
results = []
|
106 |
+
with ThreadPoolExecutor(4) as p:
|
107 |
+
try:
|
108 |
+
|
109 |
+
gen_infer = self.template.gen_infer_concurrent(
|
110 |
+
p,
|
111 |
+
audio_segment,
|
112 |
+
pivot,
|
113 |
+
)
|
114 |
+
for idx, (it, chunk) in enumerate(gen_infer, pivot):
|
115 |
+
frame = next(reader)
|
116 |
+
composed = self.template.compose(idx, frame, it)
|
117 |
+
frame_name = f"{idx}".zfill(5)+".jpg"
|
118 |
+
results.append(it['pred'])
|
119 |
+
pivot = idx + 1
|
120 |
+
except StopIteration as e:
|
121 |
+
pass
|
122 |
+
|
123 |
+
images2video(results, save_path)
|
124 |
+
|
125 |
+
return save_path
|