Rex Cheng
commited on
Commit
•
c58ca4b
1
Parent(s):
b0ec3f5
test
Browse files- mmaudio/data/__init__.py +0 -0
- mmaudio/data/av_utils.py +136 -0
mmaudio/data/__init__.py
ADDED
File without changes
|
mmaudio/data/av_utils.py
ADDED
@@ -0,0 +1,136 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from dataclasses import dataclass
|
2 |
+
from fractions import Fraction
|
3 |
+
from pathlib import Path
|
4 |
+
from typing import Optional
|
5 |
+
|
6 |
+
import av
|
7 |
+
import numpy as np
|
8 |
+
import torch
|
9 |
+
from av import AudioFrame
|
10 |
+
|
11 |
+
|
12 |
+
@dataclass
|
13 |
+
class VideoInfo:
|
14 |
+
duration_sec: float
|
15 |
+
fps: Fraction
|
16 |
+
clip_frames: torch.Tensor
|
17 |
+
sync_frames: torch.Tensor
|
18 |
+
all_frames: Optional[list[np.ndarray]]
|
19 |
+
|
20 |
+
@property
|
21 |
+
def height(self):
|
22 |
+
return self.all_frames[0].shape[0]
|
23 |
+
|
24 |
+
@property
|
25 |
+
def width(self):
|
26 |
+
return self.all_frames[0].shape[1]
|
27 |
+
|
28 |
+
|
29 |
+
def read_frames(video_path: Path, list_of_fps: list[float], start_sec: float, end_sec: float,
|
30 |
+
need_all_frames: bool) -> tuple[list[np.ndarray], list[np.ndarray], Fraction]:
|
31 |
+
output_frames = [[] for _ in list_of_fps]
|
32 |
+
next_frame_time_for_each_fps = [0.0 for _ in list_of_fps]
|
33 |
+
time_delta_for_each_fps = [1 / fps for fps in list_of_fps]
|
34 |
+
all_frames = []
|
35 |
+
|
36 |
+
# container = av.open(video_path)
|
37 |
+
with av.open(video_path) as container:
|
38 |
+
stream = container.streams.video[0]
|
39 |
+
fps = stream.guessed_rate
|
40 |
+
stream.thread_type = 'AUTO'
|
41 |
+
for packet in container.demux(stream):
|
42 |
+
for frame in packet.decode():
|
43 |
+
frame_time = frame.time
|
44 |
+
if frame_time < start_sec:
|
45 |
+
continue
|
46 |
+
if frame_time > end_sec:
|
47 |
+
break
|
48 |
+
|
49 |
+
frame_np = None
|
50 |
+
if need_all_frames:
|
51 |
+
frame_np = frame.to_ndarray(format='rgb24')
|
52 |
+
all_frames.append(frame_np)
|
53 |
+
|
54 |
+
for i, _ in enumerate(list_of_fps):
|
55 |
+
this_time = frame_time
|
56 |
+
while this_time >= next_frame_time_for_each_fps[i]:
|
57 |
+
if frame_np is None:
|
58 |
+
frame_np = frame.to_ndarray(format='rgb24')
|
59 |
+
|
60 |
+
output_frames[i].append(frame_np)
|
61 |
+
next_frame_time_for_each_fps[i] += time_delta_for_each_fps[i]
|
62 |
+
|
63 |
+
output_frames = [np.stack(frames) for frames in output_frames]
|
64 |
+
return output_frames, all_frames, fps
|
65 |
+
|
66 |
+
|
67 |
+
def reencode_with_audio(video_info: VideoInfo, output_path: Path, audio: torch.Tensor,
|
68 |
+
sampling_rate: int):
|
69 |
+
container = av.open(output_path, 'w')
|
70 |
+
output_video_stream = container.add_stream('h264', video_info.fps)
|
71 |
+
output_video_stream.codec_context.bit_rate = 10 * 1e6 # 10 Mbps
|
72 |
+
output_video_stream.width = video_info.width
|
73 |
+
output_video_stream.height = video_info.height
|
74 |
+
output_video_stream.pix_fmt = 'yuv420p'
|
75 |
+
|
76 |
+
output_audio_stream = container.add_stream('aac', sampling_rate)
|
77 |
+
|
78 |
+
# encode video
|
79 |
+
for image in video_info.all_frames:
|
80 |
+
image = av.VideoFrame.from_ndarray(image)
|
81 |
+
packet = output_video_stream.encode(image)
|
82 |
+
container.mux(packet)
|
83 |
+
|
84 |
+
for packet in output_video_stream.encode():
|
85 |
+
container.mux(packet)
|
86 |
+
|
87 |
+
# convert float tensor audio to numpy array
|
88 |
+
audio_np = audio.numpy().astype(np.float32)
|
89 |
+
audio_frame = AudioFrame.from_ndarray(audio_np, format='flt', layout='mono')
|
90 |
+
audio_frame.sample_rate = sampling_rate
|
91 |
+
|
92 |
+
for packet in output_audio_stream.encode(audio_frame):
|
93 |
+
container.mux(packet)
|
94 |
+
|
95 |
+
for packet in output_audio_stream.encode():
|
96 |
+
container.mux(packet)
|
97 |
+
|
98 |
+
container.close()
|
99 |
+
|
100 |
+
|
101 |
+
def remux_with_audio(video_path: Path, audio: torch.Tensor, output_path: Path, sampling_rate: int):
|
102 |
+
"""
|
103 |
+
NOTE: I don't think we can get the exact video duration right without re-encoding
|
104 |
+
so we are not using this but keeping it here for reference
|
105 |
+
"""
|
106 |
+
video = av.open(video_path)
|
107 |
+
output = av.open(output_path, 'w')
|
108 |
+
input_video_stream = video.streams.video[0]
|
109 |
+
output_video_stream = output.add_stream(template=input_video_stream)
|
110 |
+
output_audio_stream = output.add_stream('aac', sampling_rate)
|
111 |
+
|
112 |
+
duration_sec = audio.shape[-1] / sampling_rate
|
113 |
+
|
114 |
+
for packet in video.demux(input_video_stream):
|
115 |
+
# We need to skip the "flushing" packets that `demux` generates.
|
116 |
+
if packet.dts is None:
|
117 |
+
continue
|
118 |
+
# We need to assign the packet to the new stream.
|
119 |
+
packet.stream = output_video_stream
|
120 |
+
output.mux(packet)
|
121 |
+
|
122 |
+
# convert float tensor audio to numpy array
|
123 |
+
audio_np = audio.numpy().astype(np.float32)
|
124 |
+
audio_frame = av.AudioFrame.from_ndarray(audio_np, format='flt', layout='mono')
|
125 |
+
audio_frame.sample_rate = sampling_rate
|
126 |
+
|
127 |
+
for packet in output_audio_stream.encode(audio_frame):
|
128 |
+
output.mux(packet)
|
129 |
+
|
130 |
+
for packet in output_audio_stream.encode():
|
131 |
+
output.mux(packet)
|
132 |
+
|
133 |
+
video.close()
|
134 |
+
output.close()
|
135 |
+
|
136 |
+
output.close()
|