obrookes
commited on
Commit
β’
4ce4025
1
Parent(s):
eb24165
add str byte decoding
Browse files
frames.py
ADDED
@@ -0,0 +1,220 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import contextlib
|
2 |
+
import functools
|
3 |
+
import io
|
4 |
+
import os
|
5 |
+
import time
|
6 |
+
from typing import Union
|
7 |
+
|
8 |
+
import av
|
9 |
+
import numpy as np
|
10 |
+
import torch
|
11 |
+
|
12 |
+
|
13 |
+
class FrameSelectionMethod:
|
14 |
+
"""
|
15 |
+
Enum-like class for frame selection methods π
|
16 |
+
"""
|
17 |
+
|
18 |
+
RANDOM: str = "random" # π²
|
19 |
+
UNIFORM: str = "uniform" # π
|
20 |
+
SEQUENTIAL: str = "sequential" #
|
21 |
+
|
22 |
+
|
23 |
+
def seek_to_second(container, stream, second):
|
24 |
+
# Convert the second to the stream's time base
|
25 |
+
timestamp = int(
|
26 |
+
second * stream.time_base.denominator / stream.time_base.numerator
|
27 |
+
)
|
28 |
+
# Seek to the timestamp
|
29 |
+
container.seek(timestamp, stream=stream)
|
30 |
+
return container
|
31 |
+
|
32 |
+
|
33 |
+
def duration_in_seconds(stream):
|
34 |
+
return float(stream.duration * stream.time_base)
|
35 |
+
|
36 |
+
|
37 |
+
def frame_timestamp_in_seconds(frame, stream):
|
38 |
+
return float(frame.pts * stream.time_base)
|
39 |
+
|
40 |
+
|
41 |
+
def duration_in_seconds_from_path(video_path, modality):
|
42 |
+
with av.open(video_path) as container:
|
43 |
+
stream = next(s for s in container.streams if s.type == modality)
|
44 |
+
return duration_in_seconds(stream)
|
45 |
+
|
46 |
+
|
47 |
+
def suppress_stderr(func):
|
48 |
+
@functools.wraps(func)
|
49 |
+
def wrapper(*args, **kwargs):
|
50 |
+
with open(os.devnull, "w") as devnull:
|
51 |
+
with contextlib.redirect_stderr(devnull):
|
52 |
+
return func(*args, **kwargs)
|
53 |
+
|
54 |
+
return wrapper
|
55 |
+
|
56 |
+
|
57 |
+
@suppress_stderr
|
58 |
+
def extract_frames_pyav(
|
59 |
+
video_data: Union[str, bytes],
|
60 |
+
modality: str,
|
61 |
+
starting_second: float,
|
62 |
+
ending_second: float,
|
63 |
+
num_frames: int,
|
64 |
+
rng: np.random.Generator,
|
65 |
+
frame_selection_method: str = "RANDOM",
|
66 |
+
key_frames_only: bool = False,
|
67 |
+
stereo_audio_if_available: bool = False,
|
68 |
+
single_image_frame: bool = False,
|
69 |
+
) -> torch.Tensor:
|
70 |
+
frame_dict = {}
|
71 |
+
|
72 |
+
video_source = (
|
73 |
+
io.BytesIO(video_data) if isinstance(video_data, bytes) else video_data
|
74 |
+
)
|
75 |
+
|
76 |
+
with av.open(video_source) as container:
|
77 |
+
stream = next(s for s in container.streams if s.type == modality)
|
78 |
+
if key_frames_only:
|
79 |
+
stream.codec_context.skip_frame = "NONKEY"
|
80 |
+
|
81 |
+
container = seek_to_second(container, stream, starting_second)
|
82 |
+
|
83 |
+
# Get the duration of the video
|
84 |
+
video_duration = duration_in_seconds(stream)
|
85 |
+
# print(f"Video duration: {video_duration} seconds")
|
86 |
+
|
87 |
+
# Get the FPS of the video
|
88 |
+
video_fps = stream.average_rate
|
89 |
+
# print(f"Video FPS: {video_fps}")
|
90 |
+
|
91 |
+
for frame in container.decode(stream):
|
92 |
+
# logger.info(f"Frame timestamp: {frame}")
|
93 |
+
frame_timestamp = frame_timestamp_in_seconds(frame, stream)
|
94 |
+
# logger.info(f"Frame timestamp: {frame_timestamp}")
|
95 |
+
array_frame = torch.from_numpy(
|
96 |
+
frame.to_ndarray(
|
97 |
+
format="rgb24" if modality == "video" else None
|
98 |
+
)
|
99 |
+
)
|
100 |
+
|
101 |
+
if modality == "video" and len(array_frame.shape) == 2:
|
102 |
+
array_frame = array_frame.unsqueeze(0)
|
103 |
+
|
104 |
+
if modality == "audio" and not stereo_audio_if_available:
|
105 |
+
array_frame = array_frame[0].unsqueeze(0)
|
106 |
+
|
107 |
+
if frame_timestamp > ending_second:
|
108 |
+
break
|
109 |
+
frame_dict[frame_timestamp] = array_frame
|
110 |
+
# logger.info(f"Frame dict: {frame_dict}")
|
111 |
+
if single_image_frame:
|
112 |
+
break
|
113 |
+
|
114 |
+
frame_values = (
|
115 |
+
torch.stack(list(frame_dict.values()))
|
116 |
+
if modality == "video"
|
117 |
+
else torch.cat(list(frame_dict.values()), dim=1).permute(1, 0)
|
118 |
+
)
|
119 |
+
|
120 |
+
if frame_selection_method == FrameSelectionMethod.RANDOM:
|
121 |
+
frame_indices = rng.choice(
|
122 |
+
len(frame_values),
|
123 |
+
min(num_frames, len(frame_values)),
|
124 |
+
replace=key_frames_only,
|
125 |
+
)
|
126 |
+
elif frame_selection_method == FrameSelectionMethod.UNIFORM:
|
127 |
+
frame_indices = np.linspace(
|
128 |
+
0,
|
129 |
+
len(frame_values),
|
130 |
+
min(num_frames, len(frame_values)),
|
131 |
+
endpoint=False,
|
132 |
+
dtype=int,
|
133 |
+
)
|
134 |
+
elif frame_selection_method == FrameSelectionMethod.SEQUENTIAL:
|
135 |
+
frame_indices = np.arange(0, min(num_frames, len(frame_values)))
|
136 |
+
|
137 |
+
frame_indices = sorted(set(frame_indices))
|
138 |
+
output = frame_values[frame_indices]
|
139 |
+
|
140 |
+
if modality == "video" and len(output.shape) == 3:
|
141 |
+
output = output.unsqueeze(0)
|
142 |
+
|
143 |
+
return output
|
144 |
+
|
145 |
+
|
146 |
+
def test_extract_frames_video_pyav():
|
147 |
+
video_path = "/data/datasets/tali-wit-2-1-buckets/video_data.parquet/550/550321/4chLRYT8ylY/360p_90.mp4"
|
148 |
+
video_path = "/data/datasets/tali-wit-2-1-buckets//video_data.parquet/10/10586/SA7bKo4HRTg/360p_0.mp4"
|
149 |
+
modality = "video"
|
150 |
+
start_time = 10
|
151 |
+
end_time = 20
|
152 |
+
num_frames = 30
|
153 |
+
rng = np.random.default_rng()
|
154 |
+
|
155 |
+
for selection_method in [
|
156 |
+
FrameSelectionMethod.RANDOM,
|
157 |
+
FrameSelectionMethod.UNIFORM,
|
158 |
+
FrameSelectionMethod.SEQUENTIAL,
|
159 |
+
]:
|
160 |
+
for i in range(5):
|
161 |
+
time_list = []
|
162 |
+
for key_frames_only in [False]:
|
163 |
+
start_fn_time = time.time()
|
164 |
+
frames = extract_frames_pyav(
|
165 |
+
video_path=video_path,
|
166 |
+
modality=modality,
|
167 |
+
starting_second=start_time,
|
168 |
+
ending_second=end_time,
|
169 |
+
num_frames=num_frames,
|
170 |
+
rng=rng,
|
171 |
+
frame_selection_method=selection_method,
|
172 |
+
key_frames_only=key_frames_only,
|
173 |
+
)
|
174 |
+
end_fn_time = time.time()
|
175 |
+
time_list.append(end_fn_time - start_fn_time)
|
176 |
+
print(
|
177 |
+
f"Using {selection_method} frame selection method π², with key_frames_only: {key_frames_only}, have extracted {frames.shape}, mean time {np.mean(time_list)} seconds, std time {np.std(time_list)} seconds"
|
178 |
+
)
|
179 |
+
|
180 |
+
|
181 |
+
def test_extract_frames_audio_pyav():
|
182 |
+
video_path = "/data/datasets/tali-wit-2-1-buckets/video_data.parquet/550/550321/4chLRYT8ylY/360p_90.mp4"
|
183 |
+
video_path = "/data/datasets/tali-wit-2-1-buckets//video_data.parquet/10/10586/SA7bKo4HRTg/360p_0.mp4"
|
184 |
+
modality = "audio"
|
185 |
+
start_time = 10
|
186 |
+
end_time = 20
|
187 |
+
num_frames = 88200
|
188 |
+
rng = np.random.default_rng()
|
189 |
+
|
190 |
+
for selection_method in [
|
191 |
+
FrameSelectionMethod.RANDOM,
|
192 |
+
FrameSelectionMethod.UNIFORM,
|
193 |
+
FrameSelectionMethod.SEQUENTIAL,
|
194 |
+
]:
|
195 |
+
for i in range(5):
|
196 |
+
time_list = []
|
197 |
+
for key_frames_only in [False]:
|
198 |
+
start_fn_time = time.time()
|
199 |
+
frames = extract_frames_pyav(
|
200 |
+
video_path=video_path,
|
201 |
+
modality=modality,
|
202 |
+
starting_second=start_time,
|
203 |
+
ending_second=end_time,
|
204 |
+
num_frames=num_frames,
|
205 |
+
rng=rng,
|
206 |
+
frame_selection_method=selection_method,
|
207 |
+
key_frames_only=key_frames_only,
|
208 |
+
stereo_audio_if_available=False,
|
209 |
+
)
|
210 |
+
end_fn_time = time.time()
|
211 |
+
time_list.append(end_fn_time - start_fn_time)
|
212 |
+
print(
|
213 |
+
f"Using {selection_method} frame selection method π², with key_frames_only: {key_frames_only}, have extracted {frames.shape}, mean time {np.mean(time_list)} seconds, std time {np.std(time_list)} seconds"
|
214 |
+
)
|
215 |
+
|
216 |
+
|
217 |
+
if __name__ == "__main__":
|
218 |
+
# test_extract_frames_torchvision()
|
219 |
+
# test_extract_frames_video_pyav()
|
220 |
+
test_extract_frames_audio_pyav()
|