video.py 13.7 KB
Newer Older
1
import gc
2
import math
3
import re
4
import warnings
5
from typing import Any, Dict, List, Optional, Tuple, Union
6

7
8
import numpy as np
import torch
Francisco Massa's avatar
Francisco Massa committed
9

10
11
from . import _video_opt
from ._video_opt import VideoMetaData
Francisco Massa's avatar
Francisco Massa committed
12
13


14
15
try:
    import av
16

17
    av.logging.set_level(av.logging.ERROR)
18
19
20
    if not hasattr(av.video.frame.VideoFrame, "pict_type"):
        av = ImportError(
            """\
21
22
23
24
25
Your version of PyAV is too old for the necessary video operations in torchvision.
If you are on Python 3.5, you will have to build from source (the conda-forge
packages are not up-to-date).  See
https://github.com/mikeboers/PyAV#installation for instructions on how to
install PyAV on your system.
26
27
"""
        )
28
except ImportError:
29
30
    av = ImportError(
        """\
31
32
33
PyAV is not installed, and is necessary for the video operations in torchvision.
See https://github.com/mikeboers/PyAV#installation for instructions on how to
install PyAV on your system.
34
35
"""
    )
36
37


38
def _check_av_available() -> None:
39
40
41
42
    if isinstance(av, Exception):
        raise av


43
def _av_available() -> bool:
44
45
46
    return not isinstance(av, Exception)


47
48
# PyAV has some reference cycles
_CALLED_TIMES = 0
49
_GC_COLLECTION_INTERVAL = 10
50
51


52
53
54
55
56
57
def write_video(
    filename: str,
    video_array: torch.Tensor,
    fps: float,
    video_codec: str = "libx264",
    options: Optional[Dict[str, Any]] = None,
58
59
60
61
    audio_array: Optional[torch.Tensor] = None,
    audio_fps: Optional[float] = None,
    audio_codec: Optional[str] = None,
    audio_options: Optional[Dict[str, Any]] = None,
62
) -> None:
63
64
65
    """
    Writes a 4d tensor in [T, H, W, C] format in a video file

66
67
68
69
70
71
72
73
74
75
76
77
    Args:
        filename (str): path where the video will be saved
        video_array (Tensor[T, H, W, C]): tensor containing the individual frames,
            as a uint8 tensor in [T, H, W, C] format
        fps (Number): video frames per second
        video_codec (str): the name of the video codec, i.e. "libx264", "h264", etc.
        options (Dict): dictionary containing options to be passed into the PyAV video stream
        audio_array (Tensor[C, N]): tensor containing the audio, where C is the number of channels
            and N is the number of samples
        audio_fps (Number): audio sample rate, typically 44100 or 48000
        audio_codec (str): the name of the audio codec, i.e. "mp3", "aac", etc.
        audio_options (Dict): dictionary containing options to be passed into the PyAV audio stream
78
79
80
81
    """
    _check_av_available()
    video_array = torch.as_tensor(video_array, dtype=torch.uint8).numpy()

82
83
84
85
86
    # PyAV does not support floating point numbers with decimal point
    # and will throw OverflowException in case this is not the case
    if isinstance(fps, float):
        fps = np.round(fps)

87
88
89
90
91
92
93
    with av.open(filename, mode="w") as container:
        stream = container.add_stream(video_codec, rate=fps)
        stream.width = video_array.shape[2]
        stream.height = video_array.shape[1]
        stream.pix_fmt = "yuv420p" if video_codec != "libx264rgb" else "rgb24"
        stream.options = options or {}

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
        if audio_array is not None:
            audio_format_dtypes = {
                'dbl': '<f8',
                'dblp': '<f8',
                'flt': '<f4',
                'fltp': '<f4',
                's16': '<i2',
                's16p': '<i2',
                's32': '<i4',
                's32p': '<i4',
                'u8': 'u1',
                'u8p': 'u1',
            }
            a_stream = container.add_stream(audio_codec, rate=audio_fps)
            a_stream.options = audio_options or {}

            num_channels = audio_array.shape[0]
            audio_layout = "stereo" if num_channels > 1 else "mono"
            audio_sample_fmt = container.streams.audio[0].format.name

            format_dtype = np.dtype(audio_format_dtypes[audio_sample_fmt])
            audio_array = torch.as_tensor(audio_array).numpy().astype(format_dtype)

            frame = av.AudioFrame.from_ndarray(
                audio_array, format=audio_sample_fmt, layout=audio_layout
            )

            frame.sample_rate = audio_fps

            for packet in a_stream.encode(frame):
                container.mux(packet)

            for packet in a_stream.encode():
                container.mux(packet)

129
130
131
132
133
134
135
136
        for img in video_array:
            frame = av.VideoFrame.from_ndarray(img, format="rgb24")
            frame.pict_type = "NONE"
            for packet in stream.encode(frame):
                container.mux(packet)

        # Flush stream
        for packet in stream.encode():
137
138
139
            container.mux(packet)


140
def _read_from_stream(
141
142
143
144
145
146
147
    container: "av.container.Container",
    start_offset: float,
    end_offset: float,
    pts_unit: str,
    stream: "av.stream.Stream",
    stream_name: Dict[str, Optional[Union[int, Tuple[int, ...], List[int]]]],
) -> List["av.frame.Frame"]:
148
149
150
151
152
    global _CALLED_TIMES, _GC_COLLECTION_INTERVAL
    _CALLED_TIMES += 1
    if _CALLED_TIMES % _GC_COLLECTION_INTERVAL == _GC_COLLECTION_INTERVAL - 1:
        gc.collect()

153
    if pts_unit == "sec":
154
155
156
157
        start_offset = int(math.floor(start_offset * (1 / stream.time_base)))
        if end_offset != float("inf"):
            end_offset = int(math.ceil(end_offset * (1 / stream.time_base)))
    else:
158
159
160
161
        warnings.warn(
            "The pts_unit 'pts' gives wrong results and will be removed in a "
            + "follow-up version. Please use pts_unit 'sec'."
        )
162

163
    frames = {}
164
    should_buffer = True
165
166
    max_buffer_size = 5
    if stream.type == "video":
167
        # DivX-style packed B-frames can have out-of-order pts (2 frames in a single pkt)
168
169
        # so need to buffer some extra frames to sort everything
        # properly
170
171
172
173
174
175
176
177
178
179
180
181
        extradata = stream.codec_context.extradata
        # overly complicated way of finding if `divx_packed` is set, following
        # https://github.com/FFmpeg/FFmpeg/commit/d5a21172283572af587b3d939eba0091484d3263
        if extradata and b"DivX" in extradata:
            # can't use regex directly because of some weird characters sometimes...
            pos = extradata.find(b"DivX")
            d = extradata[pos:]
            o = re.search(br"DivX(\d+)Build(\d+)(\w)", d)
            if o is None:
                o = re.search(br"DivX(\d+)b(\d+)(\w)", d)
            if o is not None:
                should_buffer = o.group(3) == b"p"
182
    seek_offset = start_offset
183
184
    # some files don't seek to the right location, so better be safe here
    seek_offset = max(seek_offset - 1, 0)
185
186
187
188
    if should_buffer:
        # FIXME this is kind of a hack, but we will jump to the previous keyframe
        # so this will be safe
        seek_offset = max(seek_offset - max_buffer_size, 0)
189
190
191
192
    try:
        # TODO check if stream needs to always be the video stream here or not
        container.seek(seek_offset, any_frame=False, backward=True, stream=stream)
    except av.AVError:
193
194
        # TODO add some warnings in this case
        # print("Corrupted file?", container.name)
195
        return []
196
    buffer_count = 0
197
    try:
198
        for _idx, frame in enumerate(container.decode(**stream_name)):
199
200
201
202
203
204
205
206
207
            frames[frame.pts] = frame
            if frame.pts >= end_offset:
                if should_buffer and buffer_count < max_buffer_size:
                    buffer_count += 1
                    continue
                break
    except av.AVError:
        # TODO add a warning
        pass
208
    # ensure that the results are sorted wrt the pts
209
210
211
    result = [
        frames[i] for i in sorted(frames) if start_offset <= frames[i].pts <= end_offset
    ]
212
    if len(frames) > 0 and start_offset > 0 and start_offset not in frames:
213
214
215
        # if there is no frame that exactly matches the pts of start_offset
        # add the last frame smaller than start_offset, to guarantee that
        # we will have all the necessary data. This is most useful for audio
216
217
218
219
        preceding_frames = [i for i in frames if i < start_offset]
        if len(preceding_frames) > 0:
            first_frame_pts = max(preceding_frames)
            result.insert(0, frames[first_frame_pts])
220
    return result
221
222


223
224
225
def _align_audio_frames(
    aframes: torch.Tensor, audio_frames: List["av.frame.Frame"], ref_start: int, ref_end: float
) -> torch.Tensor:
226
227
228
229
230
231
232
233
234
235
236
237
    start, end = audio_frames[0].pts, audio_frames[-1].pts
    total_aframes = aframes.shape[1]
    step_per_aframe = (end - start + 1) / total_aframes
    s_idx = 0
    e_idx = total_aframes
    if start < ref_start:
        s_idx = int((ref_start - start) / step_per_aframe)
    if end > ref_end:
        e_idx = int((ref_end - end) / step_per_aframe)
    return aframes[:, s_idx:e_idx]


238
239
240
def read_video(
    filename: str, start_pts: int = 0, end_pts: Optional[float] = None, pts_unit: str = "pts"
) -> Tuple[torch.Tensor, torch.Tensor, Dict[str, Any]]:
241
242
243
244
    """
    Reads a video from a file, returning both the video frames as well as
    the audio frames

245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
    Args:
        filename (str): path to the video file
        start_pts (int if pts_unit = 'pts', float / Fraction if pts_unit = 'sec', optional):
            The start presentation time of the video
        end_pts (int if pts_unit = 'pts', float / Fraction if pts_unit = 'sec', optional):
            The end presentation time
        pts_unit (str, optional): unit in which start_pts and end_pts values will be interpreted,
            either 'pts' or 'sec'. Defaults to 'pts'.

    Returns:
        vframes (Tensor[T, H, W, C]): the `T` video frames
        aframes (Tensor[K, L]): the audio frames, where `K` is the number of channels and `L` is the
            number of points
        info (Dict): metadata for the video and audio. Can contain the fields video_fps (float)
            and audio_fps (int)
260
    """
Francisco Massa's avatar
Francisco Massa committed
261
262

    from torchvision import get_video_backend
263

Francisco Massa's avatar
Francisco Massa committed
264
265
266
    if get_video_backend() != "pyav":
        return _video_opt._read_video(filename, start_pts, end_pts, pts_unit)

267
268
269
270
271
272
    _check_av_available()

    if end_pts is None:
        end_pts = float("inf")

    if end_pts < start_pts:
273
274
275
276
        raise ValueError(
            "end_pts should be larger than start_pts, got "
            "start_pts={} and end_pts={}".format(start_pts, end_pts)
        )
277
278
279
280
281

    info = {}
    video_frames = []
    audio_frames = []

282
    try:
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
        with av.open(filename, metadata_errors="ignore") as container:
            if container.streams.video:
                video_frames = _read_from_stream(
                    container,
                    start_pts,
                    end_pts,
                    pts_unit,
                    container.streams.video[0],
                    {"video": 0},
                )
                video_fps = container.streams.video[0].average_rate
                # guard against potentially corrupted files
                if video_fps is not None:
                    info["video_fps"] = float(video_fps)

            if container.streams.audio:
                audio_frames = _read_from_stream(
                    container,
                    start_pts,
                    end_pts,
                    pts_unit,
                    container.streams.audio[0],
                    {"audio": 0},
                )
                info["audio_fps"] = container.streams.audio[0].rate

309
310
311
    except av.AVError:
        # TODO raise a warning?
        pass
312

313
314
    vframes_list = [frame.to_rgb().to_ndarray() for frame in video_frames]
    aframes_list = [frame.to_ndarray() for frame in audio_frames]
315

316
317
    if vframes_list:
        vframes = torch.as_tensor(np.stack(vframes_list))
318
319
320
    else:
        vframes = torch.empty((0, 1, 1, 3), dtype=torch.uint8)

321
322
    if aframes_list:
        aframes = np.concatenate(aframes_list, 1)
323
324
325
326
327
328
329
330
        aframes = torch.as_tensor(aframes)
        aframes = _align_audio_frames(aframes, audio_frames, start_pts, end_pts)
    else:
        aframes = torch.empty((1, 0), dtype=torch.float32)

    return vframes, aframes, info


331
def _can_read_timestamps_from_packets(container: "av.container.Container") -> bool:
332
333
334
335
336
337
338
339
    extradata = container.streams[0].codec_context.extradata
    if extradata is None:
        return False
    if b"Lavc" in extradata:
        return True
    return False


340
def _decode_video_timestamps(container: "av.container.Container") -> List[int]:
341
342
343
344
345
346
347
    if _can_read_timestamps_from_packets(container):
        # fast path
        return [x.pts for x in container.demux(video=0) if x.pts is not None]
    else:
        return [x.pts for x in container.decode(video=0) if x.pts is not None]


348
def read_video_timestamps(filename: str, pts_unit: str = "pts") -> Tuple[List[int], Optional[float]]:
349
350
351
352
353
    """
    List the video frames timestamps.

    Note that the function decodes the whole video frame-by-frame.

354
355
356
357
358
359
360
361
362
    Args:
        filename (str): path to the video file
        pts_unit (str, optional): unit in which timestamp values will be returned
            either 'pts' or 'sec'. Defaults to 'pts'.

    Returns:
        pts (List[int] if pts_unit = 'pts', List[Fraction] if pts_unit = 'sec'):
            presentation timestamps for each one of the frames in the video.
        video_fps (float, optional): the frame rate for the video
363
364

    """
Francisco Massa's avatar
Francisco Massa committed
365
    from torchvision import get_video_backend
366

Francisco Massa's avatar
Francisco Massa committed
367
368
369
    if get_video_backend() != "pyav":
        return _video_opt._read_video_timestamps(filename, pts_unit)

370
    _check_av_available()
371

372
    video_fps = None
373
    pts = []
374
375

    try:
376
377
378
379
380
381
382
383
384
        with av.open(filename, metadata_errors="ignore") as container:
            if container.streams.video:
                video_stream = container.streams.video[0]
                video_time_base = video_stream.time_base
                try:
                    pts = _decode_video_timestamps(container)
                except av.AVError:
                    warnings.warn(f"Failed decoding frames for file {filename}")
                video_fps = float(video_stream.average_rate)
385
386
387
388
    except av.AVError:
        # TODO add a warning
        pass

389
    pts.sort()
390

391
    if pts_unit == "sec":
392
393
394
        pts = [x * video_time_base for x in pts]

    return pts, video_fps