-
Notifications
You must be signed in to change notification settings - Fork 47
/
readers.py
278 lines (242 loc) · 8.65 KB
/
readers.py
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
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
import math
from typing import Optional
import av
import torch
from pytorchvideo.transforms import Normalize, ShortSideScale
from torchvision.transforms._transforms_video import CenterCropVideo
try:
from torchaudio.io import StreamReader
except ImportError:
print("WARN: torchaudio not installed", flush=True)
def get_video_meta(path):
with av.open(path) as cont:
n_frames = cont.streams[0].frames
codec = cont.streams[0].codec.name
codec_long_name = cont.streams[0].codec.long_name
tb = cont.streams[0].time_base
all_pts = []
for x in cont.demux(video=0):
if x.pts is None:
continue
all_pts.append(x.pts)
assert len(all_pts) == n_frames
return {
"all_pts": sorted(all_pts),
"codec": codec,
"codec_long_name": codec_long_name,
"tb": tb,
"width": cont.streams.video[0].width,
"height": cont.streams.video[0].height,
}
def _yuv_to_rgb(img: torch.Tensor) -> torch.Tensor:
img = img.to(torch.float)
y = img[..., 0, :, :]
u = img[..., 1, :, :]
v = img[..., 2, :, :]
y /= 255
u = u / 255 - 0.5
v = v / 255 - 0.5
r = y + 1.14 * v
g = y + -0.396 * u - 0.581 * v
b = y + 2.029 * u
rgb = torch.stack([r, g, b], -1)
return rgb.permute(3, 0, 1, 2)
def _derive_cthw_axis_order(axis_order):
mapping = {"c": 0, "t": 1, "h": 2, "w": 3}
assert set(mapping.keys()) == set(
axis_order
), "please provide 't', 'c', 'h', 'w' in any order"
assert len(axis_order) == 4, "please provide 't', 'c', 'h', 'w' in any order"
result = []
for ch in axis_order:
result.append(mapping[ch])
return tuple(result)
class StridedReader:
def __init__(self, path, stride, frame_window_size, axis_order: str = "cthw"):
self.path = path
self.meta = get_video_meta(path)
self.all_pts = self.meta["all_pts"]
self.stride = stride
self.frame_window_size = frame_window_size
self.axis_order = axis_order
self.cthw_to_axis_order = _derive_cthw_axis_order(axis_order)
if self.stride == 0:
self.stride = self.frame_window_size
def __getitem__(self, _: int) -> torch.Tensor:
raise AssertionError("Not implemented")
def __len__(self):
return int(
math.ceil((len(self.all_pts) - self.frame_window_size) / self.stride)
)
class TorchAudioStreamReader(StridedReader):
def __init__(
self,
path: str,
resize: Optional[int],
crop: Optional[int],
mean: Optional[torch.Tensor],
std: Optional[torch.Tensor],
frame_window_size: int,
stride: int,
gpu_idx: int,
axis_order: str = "cthw",
uint8_scale: bool = False,
resize_with_hardware: Optional[int] = None,
):
"""
NOTE:
- resize_on_hardware uses CUVID's hardware resize operation. I don't
know what interpolation algorithm it uses (can't find documentation).
Use with caution.
"""
super().__init__(path, stride, frame_window_size, axis_order)
self.mean = mean
self.crop = crop
self.std = std
self.resize = resize
self.resize_transform = (
ShortSideScale(self.resize) if self.resize is not None else None
)
self.norm_transform = (
Normalize(mean=self.mean, std=self.std) if self.mean is not None else None
)
self.crop_transform = (
CenterCropVideo(self.crop) if self.crop is not None else None
)
self.resize_on_hardware = resize_with_hardware
self.uint8_scale = uint8_scale
self.create_underlying_cont(gpu_idx)
def create_underlying_cont(self, gpu_id):
self.gpu_id = gpu_id
decoder_basename = self.meta["codec"]
if self.resize_on_hardware:
assert self.gpu_id >= 0
w, h = self.meta["width"], self.meta["height"]
roh = self.resize_on_hardware
rw, rh = roh, roh
if w < h:
rh = int(math.floor((float(h) / w) * roh))
elif w > h:
rw = int(math.floor((float(w) / h) * roh))
decoder_opt = (
{"resize": f"{rw}x{rh}", "gpu": f"{gpu_id}"}
if self.resize is not None
else {"gpu": f"{gpu_id}"}
)
self.conf = {
"decoder": f"{decoder_basename}_cuvid",
"hw_accel": f"cuda:{gpu_id}",
"decoder_option": decoder_opt,
"stream_index": 0,
}
else:
self.conf = {
"decoder": decoder_basename,
"stream_index": 0,
}
self.resize_transform = (
ShortSideScale(self.resize) if self.resize is not None else None
)
self.cont = StreamReader(self.path)
self.cont.add_video_stream(self.frame_window_size, **self.conf)
self.cont.fill_buffer()
def __getitem__(self, idx: int) -> dict:
frame_i = self.stride * idx
frame_j = frame_i + self.frame_window_size - 1
assert frame_i >= 0 and frame_j < len(self.all_pts)
frame_i_pts = self.all_pts[frame_i]
self.cont.seek(float(frame_i_pts * self.meta["tb"]))
fs = None
for fs in self.cont.stream():
break
assert fs is not None
assert len(fs) == 1
ret = _yuv_to_rgb(fs[0]) # TODO: optimize me
assert ret.shape[1] == self.frame_window_size
if self.resize_transform is not None:
ret = self.resize_transform(ret)
if self.crop_transform is not None:
ret = self.crop_transform(ret)
if self.norm_transform is not None:
ret = self.norm_transform(ret)
ret = ret.permute(self.cthw_to_axis_order)
if self.uint8_scale:
ret *= 255
ret = ret.to(dtype=torch.uint8)
return {
"video": ret,
"frame_start_idx": frame_i,
"frame_end_idx": frame_j,
}
class PyAvReader(StridedReader):
def __init__(
self,
path: str,
resize: Optional[int],
crop: Optional[int],
mean: Optional[torch.Tensor],
std: Optional[torch.Tensor],
frame_window_size: int,
stride: int,
gpu_idx: int,
axis_order: str = "cthw",
uint8_scale: bool = False,
):
super().__init__(path, stride, frame_window_size, axis_order)
if gpu_idx >= 0:
print("WARN: GPU decoding not supported for PyAV, using CPU")
self.mean = mean
self.crop = crop
self.std = std
self.resize = resize
self.resize_transform = (
ShortSideScale(self.resize) if self.resize is not None else None
)
self.norm_transform = (
Normalize(mean=self.mean, std=self.std) if self.mean is not None else None
)
self.crop_transform = (
CenterCropVideo(self.crop) if self.crop is not None else None
)
self.uint8_scale = uint8_scale
self.create_underlying_cont(gpu_idx)
def create_underlying_cont(self, _):
self.cont = av.open(self.path)
def __getitem__(self, idx: int) -> dict:
frame_i = self.stride * idx
frame_j = frame_i + self.frame_window_size
assert frame_i >= 0 and frame_j < len(self.all_pts)
frame_i_pts = self.all_pts[frame_i]
frame_j_pts = self.all_pts[frame_j]
self.cont.seek(frame_i_pts, stream=self.cont.streams.video[0])
fs = []
for f in self.cont.decode(video=0):
if f.pts < frame_i_pts:
continue
if f.pts >= frame_j_pts:
break
fs.append(
(
f.pts,
torch.tensor(f.to_ndarray(format="rgb24"), dtype=torch.float32)
/ 255,
)
)
fs = sorted(fs, key=lambda x: x[0])
# NOTE: channel first for the transforms
ret = torch.stack([x[1] for x in fs]).permute(3, 0, 1, 2)
if self.resize_transform is not None:
ret = self.resize_transform(ret)
if self.crop_transform is not None:
ret = self.crop_transform(ret)
if self.norm_transform is not None:
ret = self.norm_transform(ret)
ret = ret.permute(self.cthw_to_axis_order)
if self.uint8_scale:
ret *= 255
ret = ret.to(dtype=torch.uint8)
return {
"video": ret,
"frame_start_idx": frame_i,
"frame_end_idx": frame_j - 1,
}