def test_moviepy(infile, out_dir):
import numpy as np
import pipi
from moviepy.video.io.ffmpeg_reader import FFMPEG_VideoReader
from moviepy.video.io.ffmpeg_writer import FFMPEG_VideoWriter
chunksize = 128
vr = FFMPEG_VideoReader(infile)
w,h = vr.size
chunk = np.zeros((h, w, vr.depth, chunksize), dtype=np.uint8)
with pipi.Timer("mp read..."):
for i in range(chunksize):
frame = vr.read_frame()
chunk[...,i] = frame
vw = FFMPEG_VideoWriter(os.path.join(out_dir, "mp_ov.mkv"), (w,h), 30, codec="libx264", preset="fast", ffmpeg_params=["-crf", "0"])
with pipi.Timer("mp write..."):
for i in range(chunksize):
vw.write_frame(chunk[...,i])
vr.close()
vw.close()
def test_moviepy(infile, out_dir):
import numpy as np
import pipi
from moviepy.video.io.ffmpeg_reader import FFMPEG_VideoReader
from moviepy.video.io.ffmpeg_writer import FFMPEG_VideoWriter
chunksize = 128
vr = FFMPEG_VideoReader(infile)
w,h = vr.size
chunk = np.zeros((h, w, vr.depth, chunksize), dtype=np.uint8)
with pipi.Timer("mp read..."):
for i in range(chunksize):
frame = vr.read_frame()
chunk[...,i] = frame
vw = FFMPEG_VideoWriter(os.path.join(out_dir, "mp_ov.mkv"), (w,h), 30, codec="libx264", preset="fast", ffmpeg_params=["-crf", "0"])
with pipi.Timer("mp write..."):
for i in range(chunksize):
vw.write_frame(chunk[...,i])
vr.close()
vw.close()
def get_video_info(video_path):
# video_fps, video_n_frames, video_duration = video_utils.
cap = FFMPEG_VideoReader(video_path, False)
cap.initialize()
fps = cap.fps
n_frames = cap.nframes
duration = cap.duration
cap.close()
del cap
return fps, n_frames, duration
def test_autorotate():
# This test requires ffmpeg >=2.7
video_file = 'media/ficus_vertical.mp4'
reader = FFMPEG_VideoReader(video_file)
assert reader.infos['video_size'] == [1920, 1080]
assert reader.infos['video_rotation'] == 90
assert reader.size == [1080, 1920]
reader.close()
reader = FFMPEG_VideoReader(video_file, ffmpeg_params=['-noautorotate'])
assert reader.size == [1920, 1080]
assert reader.rotation == 90
reader.close()
def video_uniform_sample_and_save_old(spf,
video_path,
frames_path,
image_name_format,
resize_type,
verbose=False):
if resize_type is not None:
assert resize_type in ['resize', 'resize_crop', 'resize_crop_scaled']
resize_function = None
if resize_type == 'resize':
resize_function = image_utils.resize_frame
elif resize_type == 'resize_crop':
resize_function = image_utils.resize_crop
elif resize_type == 'resize_crop_scaled':
resize_function = image_utils.resize_crop_scaled
cap = FFMPEG_VideoReader(video_path, False)
cap.initialize()
fps = cap.fps
n_frames = cap.nframes
duration = cap.duration
n_samples = int(duration / float(spf))
# check if no samples because the video duration is less than spf
# then at least, get 1 frame of the video
if n_samples == 0:
n_samples = 1
for i in range(n_samples):
num = i + 1
if verbose:
print(' ... reading frame %d/%d' % (num, n_samples))
time_sec = i * spf
frame = cap.get_frame(time_sec)
if resize_type is not None:
# resize frame to fit in the array, it's going to be used by caffe anyway
frame = resize_function(frame)
frame = cv2.cvtColor(frame, cv2.COLOR_RGB2BGR)
image_name = image_name_format % (num, )
frame_path = os.path.join(frames_path, image_name)
cv2.imwrite(frame_path, frame)
# very important, or we'd have memory leakage
cap.close()
return fps, n_frames, duration
else:
writer = FFMPEG_VideoWriter(arguments_strVideoOut, reader.size,
reader.fps * 2)
if exists:
# Write frames that were already completed
print("Re-writing processed frames...")
printProgressBar(0, readerCont.nframes)
for x in range(0, readerCont.nframes):
writer.write_frame(readerCont.read_frame())
printProgressBar(x + 1, readerCont.nframes)
reader.skip_frames(startFrame)
readerCont.close()
print("Deleting temporary file(s)...")
shutil.rmtree(tempDir)
print("Processing resumed!")
print("")
totalFrames = reader.nframes
nextFrame = reader.read_frame()
startedTime = datetime.now()
for x in range(startFrame, reader.nframes):
start = datetime.now()
firstFrame = nextFrame
nextFrame = reader.read_frame()
tensorInputFirst = torch.FloatTensor(
class VideoFileClip(VideoClip):
"""
A video clip originating from a movie file. For instance: ::
>>> clip = VideoFileClip("myHolidays.mp4")
>>> clip.close()
>>> with VideoFileClip("myMaskVideo.avi") as clip2:
>>> pass # Implicit close called by context manager.
Parameters
----------
filename:
The name of the video file, as a string or a path-like object.
It can have any extension supported by ffmpeg:
.ogv, .mp4, .mpeg, .avi, .mov etc.
has_mask:
Set this to 'True' if there is a mask included in the videofile.
Video files rarely contain masks, but some video codecs enable
that. For instance if you have a MoviePy VideoClip with a mask you
can save it to a videofile with a mask. (see also
``VideoClip.write_videofile`` for more details).
audio:
Set to `False` if the clip doesn't have any audio or if you do not
wish to read the audio.
target_resolution:
Set to (desired_width, desired_height) to have ffmpeg resize the frames
before returning them. This is much faster than streaming in high-res
and then resizing. If either dimension is None, the frames are resized
by keeping the existing aspect ratio.
resize_algorithm:
The algorithm used for resizing. Default: "bicubic", other popular
options include "bilinear" and "fast_bilinear". For more information, see
https://ffmpeg.org/ffmpeg-scaler.html
fps_source:
The fps value to collect from the metadata. Set by default to 'fps', but
can be set to 'tbr', which may be helpful if you are finding that it is reading
the incorrect fps from the file.
pixel_format
Optional: Pixel format for the video to read. If is not specified
'rgb24' will be used as the default format unless ``has_mask`` is set
as ``True``, then 'rgba' will be used.
Attributes
----------
filename:
Name of the original video file.
fps:
Frames per second in the original file.
Read docs for Clip() and VideoClip() for other, more generic, attributes.
Lifetime
--------
Note that this creates subprocesses and locks files. If you construct one
of these instances, you must call close() afterwards, or the subresources
will not be cleaned up until the process ends.
If copies are made, and close() is called on one, it may cause methods on
the other copies to fail.
"""
@convert_path_to_string("filename")
def __init__(
self,
filename,
decode_file=False,
has_mask=False,
audio=True,
audio_buffersize=200000,
target_resolution=None,
resize_algorithm="bicubic",
audio_fps=44100,
audio_nbytes=2,
fps_source="fps",
pixel_format=None,
):
VideoClip.__init__(self)
# Make a reader
if not pixel_format:
pixel_format = "rgba" if has_mask else "rgb24"
self.reader = FFMPEG_VideoReader(
filename,
decode_file=decode_file,
pixel_format=pixel_format,
target_resolution=target_resolution,
resize_algo=resize_algorithm,
fps_source=fps_source,
)
# Make some of the reader's attributes accessible from the clip
self.duration = self.reader.duration
self.end = self.reader.duration
self.fps = self.reader.fps
self.size = self.reader.size
self.rotation = self.reader.rotation
self.filename = filename
if has_mask:
self.make_frame = lambda t: self.reader.get_frame(t)[:, :, :3]
def mask_make_frame(t):
return self.reader.get_frame(t)[:, :, 3] / 255.0
self.mask = VideoClip(is_mask=True,
make_frame=mask_make_frame).with_duration(
self.duration)
self.mask.fps = self.fps
else:
self.make_frame = lambda t: self.reader.get_frame(t)
# Make a reader for the audio, if any.
if audio and self.reader.infos["audio_found"]:
self.audio = AudioFileClip(
filename,
buffersize=audio_buffersize,
fps=audio_fps,
nbytes=audio_nbytes,
)
def __deepcopy__(self, memo):
"""Implements ``copy.deepcopy(clip)`` behaviour as ``copy.copy(clip)``.
VideoFileClip class instances can't be deeply copied because the locked Thread
of ``proc`` isn't pickleable. Without this override, calls to
``copy.deepcopy(clip)`` would raise a ``TypeError``:
```
TypeError: cannot pickle '_thread.lock' object
```
"""
return self.__copy__()
def close(self):
"""Close the internal reader."""
if self.reader:
self.reader.close()
self.reader = None
try:
if self.audio:
self.audio.close()
self.audio = None
except AttributeError: # pragma: no cover
pass
class VideoFileClip(VideoClip):
"""
A video clip originating from a movie file. For instance: ::
>>> clip = VideoFileClip("myHolidays.mp4")
>>> clip.close()
>>> with VideoFileClip("myMaskVideo.avi") as clip2:
>>> pass # Implicit close called by contex manager.
Parameters
------------
filename:
The name of the video file. It can have any extension supported
by ffmpeg: .ogv, .mp4, .mpeg, .avi, .mov etc.
has_mask:
Set this to 'True' if there is a mask included in the videofile.
Video files rarely contain masks, but some video codecs enable
that. For istance if you have a MoviePy VideoClip with a mask you
can save it to a videofile with a mask. (see also
``VideoClip.write_videofile`` for more details).
audio:
Set to `False` if the clip doesn't have any audio or if you do not
wish to read the audio.
target_resolution:
Set to (desired_height, desired_width) to have ffmpeg resize the frames
before returning them. This is much faster than streaming in high-res
and then resizing. If either dimension is None, the frames are resized
by keeping the existing aspect ratio.
resize_algorithm:
The algorithm used for resizing. Default: "bicubic", other popular
options include "bilinear" and "fast_bilinear". For more information, see
https://ffmpeg.org/ffmpeg-scaler.html
fps_source:
The fps value to collect from the metadata. Set by default to 'tbr', but
can be set to 'fps', which may be helpful if importing slow-motion videos
that get messed up otherwise.
Attributes
-----------
filename:
Name of the original video file.
fps:
Frames per second in the original file.
Read docs for Clip() and VideoClip() for other, more generic, attributes.
Lifetime
--------
Note that this creates subprocesses and locks files. If you construct one of these instances, you must call
close() afterwards, or the subresources will not be cleaned up until the process ends.
If copies are made, and close() is called on one, it may cause methods on the other copies to fail.
"""
def __init__(self, filename, has_mask=False,
audio=True, audio_buffersize = 200000,
target_resolution=None, resize_algorithm='bicubic',
audio_fps=44100, audio_nbytes=2, verbose=False,
fps_source='tbr'):
VideoClip.__init__(self)
# Make a reader
pix_fmt= "rgba" if has_mask else "rgb24"
self.reader = FFMPEG_VideoReader(filename, pix_fmt=pix_fmt,
target_resolution=target_resolution,
resize_algo=resize_algorithm,
fps_source=fps_source)
# Make some of the reader's attributes accessible from the clip
self.duration = self.reader.duration
self.end = self.reader.duration
self.fps = self.reader.fps
self.size = self.reader.size
self.rotation = self.reader.rotation
self.filename = self.reader.filename
if has_mask:
self.make_frame = lambda t: self.reader.get_frame(t)[:,:,:3]
mask_mf = lambda t: self.reader.get_frame(t)[:,:,3]/255.0
self.mask = (VideoClip(ismask = True, make_frame = mask_mf)
.set_duration(self.duration))
self.mask.fps = self.fps
else:
self.make_frame = lambda t: self.reader.get_frame(t)
# Make a reader for the audio, if any.
if audio and self.reader.infos['audio_found']:
self.audio = AudioFileClip(filename,
buffersize= audio_buffersize,
fps = audio_fps,
nbytes = audio_nbytes)
def close(self):
""" Close the internal reader. """
if self.reader:
self.reader.close()
self.reader = None
try:
if self.audio:
self.audio.close()
self.audio = None
except AttributeError:
pass
print "n_frames:", movie_reader.nframes
# the 1st frame of the saved video can't be directly read by movie_reader.read_frame(), I don't know why
# maybe it's a bug of anim.save, actually, if we look at the movie we get from anim.save
# we can easilly see that the 1st frame just close very soon.
# so I manually get it at time 0, this is just a trick, I think.
tmp_frame = movie_reader.get_frame(0)
[
movie_writer.write_frame(tmp_frame)
for _ in range(int(new_fps * play_slow_rate))
]
# for the above reason, we should read (movie_reader.nframes-1) frames so that the last frame is not
# read twice (not that get_frame(0) alread read once)
# However, I soon figure out that it should be (movie_reader.nframes-2). The details: we have actually
# 6 frames, but (print movie_reader.nframes) is 7. I read the first frame through movie_reader.get_frame(0)
# then are are 5 left. So I should use movie_reader.nframes - 2. Note that in fig1_pcm_fs2.py
# in the case of: original fps=1
# new_fps = 24, play_slow_rate = 1.5 the result is: 1st frame last 1.8s, others 1.5s, i.e., the 1st frame
# has more duration. This is messy.
for i in range(movie_reader.nframes - 2):
tmp_frame = movie_reader.read_frame()
[
movie_writer.write_frame(tmp_frame)
for _ in range(int(new_fps * play_slow_rate))
]
pass
movie_reader.close()
movie_writer.close()
os.remove(tmp_video_name)
# plt.show()
pass
class MovieEditor(object):
def __init__(self,
input,
output,
width=None,
height=None,
log_level=FFMPEG_LOGLEVEL):
self._reader = FFMPEG_VideoReader(input)
self._fps = self._reader.fps
# self.cur_frame = 1
# self._frame = None
self._querier = FfmpegQuerier()
self._info = self._querier(input)
self._duration = self._querier.duration
self.draw_dict = {}
self._resize = (int(width),
int(height)) if (width and height) else None
self._loglevel = log_level
self._output = output
self._tmp_file = self._make_tmp_file(input) if self._resize else None
self._writer = FFMPEG_VideoWriter(
self._tmp_file if self.need_resize else output,
size=self._reader.size,
fps=self._reader.fps)
@property
def need_resize(self):
return self._resize and list(self._resize) != list(self._reader.size)
# for resizing
def _make_tmp_file(self, input):
return '%s-%s.mp4' % (input, get_time())
def _remove_tmp_file(self):
if os.path.exists(self._tmp_file):
os.remove(self._tmp_file)
# approximate frame count
def get_total_frame(self):
return timestamp2frame(self._duration, self._fps)
# def seek_frame(self, index):
# self._reader.skip_frames(index - self.cur_frame)
# self._frame = self._reader.read_frame()
#
# def seek_timestamp(self, timestamp):
# index = FrameEditor.which_frame(timestamp, self._fps)
# self.seek_frame(index)
#
# def draw_anno(self, anno):
# if self._frame is not None:
# self._frame_editor.frame = self._frame
# self._frame_editor.draw_anno(anno)
def draw_anno_file(self, anno_filename):
# parse anno file
ax = AnnoXml()
ax.read(anno_filename)
ax.parse()
commands = ax.command
fps = self._fps
frame = self._reader.read_frame()
self._writer.write_frame(frame)
# total frame
total_frame = self.get_total_frame()
with click.progressbar(length=total_frame,
label='Processing...') as bar:
# current
frame_index = 1
bar.update(1)
for command in commands:
pageid = command.pageid
command_frame = CommandParser.get_frame_index(command, fps)
if pageid in self.draw_dict:
frame_editor = self.draw_dict[pageid]
else:
frame_editor = FrameEditor(fps=fps, pageid=pageid)
self.draw_dict[pageid] = frame_editor
# before
while frame_index < command_frame:
if not frame_editor.commands:
frame = self.read_frame()
self.write_frame(frame)
else:
frame = self.read_frame()
frame_editor.draw(frame)
self.write_frame(frame_editor.image)
frame_index += 1
bar.update(1)
# append
frame_editor.append_command(command)
# current
frame = self.read_frame()
frame_editor.draw(frame)
self.write_frame(frame_editor.image)
frame_index += 1
bar.update(1)
# write left frames
while frame_index < total_frame:
self.write_frame(self.read_frame())
frame_index += 1
bar.update(1)
# close stream
self.close()
# resize if needed
if self.need_resize:
self.resize()
self._remove_tmp_file()
def read_frame(self):
return self._reader.read_frame()
def write_frame(self, frame):
# if self._resize:
# if not isinstance(frame, Image.Image):
# frame = Image.fromarray(frame.astype('uint8'))
# Here's a bug.
# frame = frame.resize(self._resize, Image.ANTIALIAS)
self._writer.write_frame(frame)
def resize(self):
parameters = [
FFMPEG_FILE, '-i', self._tmp_file, '-s',
'{w}*{h}'.format(w=self._resize[0], h=self._resize[1]),
'-loglevel', self._loglevel, '-y', self._output
]
subprocess.call(parameters)
def close(self):
self._reader.close()
self._writer.close()
class VideoFileClip(VideoClip):
"""
A video clip originating from a movie file. For instance: ::
>>> clip = VideoFileClip("myHolidays.mp4")
>>> clip.close()
>>> with VideoFileClip("myMaskVideo.avi") as clip2:
>>> pass # Implicit close called by context manager.
Parameters
------------
filename:
The name of the video file. It can have any extension supported
by ffmpeg: .ogv, .mp4, .mpeg, .avi, .mov etc.
has_mask:
Set this to 'True' if there is a mask included in the videofile.
Video files rarely contain masks, but some video codecs enable
that. For istance if you have a MoviePy VideoClip with a mask you
can save it to a videofile with a mask. (see also
``VideoClip.write_videofile`` for more details).
audio:
Set to `False` if the clip doesn't have any audio or if you do not
wish to read the audio.
target_resolution:
Set to (desired_height, desired_width) to have ffmpeg resize the frames
before returning them. This is much faster than streaming in high-res
and then resizing. If either dimension is None, the frames are resized
by keeping the existing aspect ratio.
resize_algorithm:
The algorithm used for resizing. Default: "bicubic", other popular
options include "bilinear" and "fast_bilinear". For more information, see
https://ffmpeg.org/ffmpeg-scaler.html
fps_source:
The fps value to collect from the metadata. Set by default to 'tbr', but
can be set to 'fps', which may be helpful if importing slow-motion videos
that get messed up otherwise.
Attributes
-----------
filename:
Name of the original video file.
fps:
Frames per second in the original file.
Read docs for Clip() and VideoClip() for other, more generic, attributes.
Life time
--------
Note that this creates sub-processes and locks files. If you construct one of these instances, you must call
close() afterwards, or the sub-resources will not be cleaned up until the process ends.
If copies are made, and close() is called on one, it may cause methods on the other copies to fail.
"""
def __init__(self,
filename,
has_mask=False,
audio=True,
audio_buffersize=200000,
target_resolution=None,
resize_algorithm='bicubic',
audio_fps=44100,
audio_nbytes=2,
verbose=False,
fps_source='tbr'):
VideoClip.__init__(self)
# Make a reader
pix_fmt = "rgba" if has_mask else "rgb24"
self.reader = FFMPEG_VideoReader(filename,
pix_fmt=pix_fmt,
target_resolution=target_resolution,
resize_algo=resize_algorithm,
fps_source=fps_source)
# Make some of the reader's attributes accessible from the clip
self.duration = self.reader.duration
self.end = self.reader.duration
self.fps = self.reader.fps
self.size = self.reader.size
self.rotation = self.reader.rotation
self.filename = self.reader.filename
if has_mask:
self.make_frame = lambda t: self.reader.get_frame(t)[:, :, :3]
mask_mf = lambda t: self.reader.get_frame(t)[:, :, 3] / 255.0
self.mask = (VideoClip(
ismask=True, make_frame=mask_mf).set_duration(self.duration))
self.mask.fps = self.fps
else:
self.make_frame = lambda t: self.reader.get_frame(t)
# Make a reader for the audio, if any.
if audio and self.reader.infos['audio_found']:
self.audio = AudioFileClip(filename,
buffersize=audio_buffersize,
fps=audio_fps,
nbytes=audio_nbytes)
def close(self):
""" Close the internal reader. """
if self.reader:
self.reader.close()
self.reader = None
try:
if self.audio:
self.audio.close()
self.audio = None
except AttributeError:
pass
