def create_video(frames, name, train_step_number, last_index,
episode_reward_sum):
if len(frames):
tmp4_fname = os.path.join(
model_base_filedir, "ATARI_step%010d_%s_%06d_%03d.mp4" %
(train_step_number, name, last_index, episode_reward_sum))
vwrite(tmp4_fname, np.array(frames, dtype=np.uint8))
def play_bson_silhouette(days_folder, start, end, room):
'G:\HouseData-Sphere\4954\backups'
day = '{}-{:02}-{:02}'.format(start.year, start.month, start.day)
bson_path = os.path.join(days_folder, day, 'sphere', 'VID.bson')
if os.path.exists(bson_path):
use_bzip = False
else:
use_bzip = True
bson_path += '.bz2'
bson_data = BsonFile(bson_path)
room_uid = read_room_uid(os.path.dirname(bson_path), room, use_bzip)
json_data = []
for entry in tqdm(bson_data.data):
entry_time = entry['bt']
entry_room = entry['uid']
if entry_room == room_uid and entry_time >= start and entry_time <= end:
json_data.append(entry)
silhouettes = [
bf['e'][0]['v'] for bf in json_data if bf['e'][0]['n'] == 'silhouette'
]
sample = b64_image_to_numpy(silhouettes[0])
video = np.zeros((len(silhouettes), sample.shape[0], sample.shape[1]),
dtype=np.uint8)
for i, silhouette in enumerate(tqdm(silhouettes)):
video[i, ...] = b64_image_to_numpy(silhouette)
video_name = 'bson_{}.mp4'.format(start).replace(':', '.')
vwrite(video_name, video)
print('Frames from {} to {} were saved to {}'.format(
start, end, video_name))
def extract_data(self):
from skvideo.io import vwrite
from os import walk
for (session_dir, _j, scenes) in list(walk(self.sessions_path)):
for scene in scenes:
scene_dict = self.x.extract_signs_from(
osp.join(session_dir, scene))
for key, value in scene_dict.items():
vwrite(osp.join(self.videos_path, key + '.mov'), value)
def gif(filename, image_list):
fname, _ = os.path.splitext(filename)
filename = fname + '.mp4'
np_images = []
for m in image_list:
np_images.append(m.permute(1, 2, 0).numpy() * 255)
vwrite(filename, np_images, verbosity=0)
return
def write_file(colored_frames, fps, output_type, output_wihout_ext,
video_number):
name = "{}_{}{}".format(output_wihout_ext, video_number, output_type) \
if video_number else "{}{}".format(output_wihout_ext, output_type)
fps_string = "{:.2f}".format(fps)
io.vwrite(name,
colored_frames,
inputdict={"-framerate": fps_string},
outputdict={"-r": fps_string}) # WRITE VIDEO
print("\nwritten to {}".format(name))
def read_depth_video(videoname,
framesdir,
shape=(480, 640, 1),
dtype=np.uint16):
vid_data = read_frame_data(
**default_read_z16_args(framesdir, shape=shape, dtype=dtype))
vid_data = grey_to_redblue_codec(np.array(vid_data, dtype=np.long))
skio.vwrite(join(VIDDIR, videoname), vid_data)
def extract(self):
from skvideo.io import vwrite
from os import walk
for (session_dir, _j, scenes) in list(walk(self.sessions_path)):
for scene in scenes:
scene_dict = self.x.extract_signs_from(
osp.join(session_dir, scene))
for key, value in scene_dict.items():
vwrite(osp.join(self.videos_path, key), value)
a = osp.join(self.my_path, 'extracted')
open(a, 'w').close()
def _to_supervisely(dataset, output):
try:
import supervisely_lib as sly
from skvideo.io import vwrite
except ModuleNotFoundError:
raise ModuleNotInstalledException("supervisely")
schema_dict = dataset.schema.dict_
for key, schem in schema_dict.items():
if isinstance(schem, Image):
project_type = "images"
extension = "jpeg"
break
elif isinstance(schem, Video):
project_type = "videos"
extension = "mp4"
break
else:
raise Exception
mode = sly.OpenMode.CREATE
if project_type == "images":
_project = sly.Project
elif project_type == "videos":
_project = sly.VideoProject
else:
raise Exception
pr = _project(output, mode)
meta = pr.meta
meta._project_type = project_type
# probably here we can create multiple datasets
out_ds = pr.create_dataset(output)
try:
fn_key = "filename"
dataset[fn_key]
except KeyError:
fn_key = None
zeroes = len(str(len(dataset)))
for idx, view in enumerate(dataset):
obj = view[key].compute()
if fn_key:
fn = view[fn_key].compute()
else:
fn = f"{idx:0{zeroes}}"
fn = "{}.{}".format(fn, extension)
# strangely supervisely prevents from using this method on videos
try:
out_ds.add_item_np(fn, obj)
except RuntimeError:
# fix with in-memory file
path = "{}/{}".format(out_ds.item_dir, fn)
vwrite(path, obj)
out_ds._item_to_ann[fn] = fn + ".json"
out_ds.set_ann(fn, out_ds._get_empty_annotaion(path))
pr.set_meta(meta)
return pr
def frames_tensor_to_video(frames_tensor, video_write_path):
"""
Writes a tensor of frames of size [n_frames x n_channels x height x width]
to a video file at video_write_path.
Args:
frames_tensor (PyTorch Variable / tensor or numpy array): tensor of frames
video_write_path (str): path to write the video
"""
assert type(video_write_path) == str, 'Video write path must be a string.'
frames_tensor = convert_tensor(frames_tensor)
vid.vwrite(video_write_path, frames_tensor)
def video(self):
BIG_MAP_TIME = np.zeros((self.TIME_DURATION, self.X, self.Y))
for i in range(self.TIME_DURATION):
A = tuple(map(tuple, self.DATA[0][i].T))
P = tuple(map(tuple, self.DATA[1][i].T))
BIG_MAP_TIME[i][A] = 100
BIG_MAP_TIME[i][P] = 200
# export to video
self.VIDEO = BIG_MAP_TIME
FILENAME = os.getcwd(
) + os.path.sep + 'OUT' + os.path.sep + 'MAP_LYFE.mp4'
skio.vwrite(FILENAME, self.VIDEO)
def save_summary(src_video, label, out_video):
logging.info(f'loading {src_video}')
video = vread(src_video)
summary = video[np.where(label)]
logging.info(f'writing {out_video}')
vwrite(out_video,
summary,
outputdict={
'-vcodec': 'libx264',
'-crf': '20',
'-pix_fmt': 'yuv420p'
})
def save_images_to_gif(images, save_path):
'''
Input: images: Batch_Size x 3 x H x W [Torch.Tensor]
'''
os.makedirs(os.path.dirname(save_path), exist_ok=True)
images = images.clamp(min=0, max=1.0)
imgs = images.detach().cpu().numpy().transpose((0, 2, 3, 1)) * 255
imgs = imgs.astype('uint8')
imageio.imwrite(save_path.replace('.gif', '.png'), imgs[0])
imageio.mimsave(save_path, imgs)
vwrite(save_path.replace('.gif', '.mp4'), imgs, outputdict={'-r': '30'})
def arr2vid(arr, path, fps = '20', speed_mult = '2'):
assert(len(arr.shape) == 4)
# Compress output vid? makes writing to a video signifigantly faster
compress = False
speed = 'setpts=' + speed_mult + '*PTS'
if compress:
outdict = {'-r': fps, '-vcodec' : 'libx265', '-crf' : '30', '-filter:v' : speed}
else:
outdict = {'-r': fps, '-filter:v' : speed}
skv.vwrite(path, arr, outputdict = outdict)
def images2avi(images, vidfile):
# save video
skv.vwrite(vidfile,
images,
inputdict={
'-vcodec': 'rawvideo',
'-pix_fmt': 'rgb24'
},
outputdict={
'-vcodec': 'rawvideo',
'-pix_fmt': 'rgb24'
})
def predict_video(num_frame, device, video_idx):
# device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
device = torch.device(device)
model_parameters = torch.load('../model/new_video_vangogh' +
str(video_idx) + '.pth')
path = '../../2019_NCL_Brand_Essence_Good_to_be_Free.mp4'
model = transferNet.TransferNet()
model.load_state_dict(model_parameters)
model.to(device)
cap = cv2.VideoCapture(path)
# fourcc = cv2.VideoWriter_fourcc(*args.fourcc)
# out_writer = cv2.VideoWriter("../outputImage/outputvideo_new.avi",
# fourcc, 24.0, (640, 480))
mean = [0.485, 0.456, 0.406]
std = [0.229, 0.224, 0.225]
frame_lst = []
count = 0
while (True):
torch.cuda.empty_cache()
ret, frame = cap.read()
count = count + 1
print(count)
if ret and count <= num_frame:
current_frame = torch.Tensor(frame.transpose(2, 0, 1)) / 255.
normalized = normalize(current_frame, mean, std)
torch.cuda.empty_cache()
new_frame = normalized.reshape((1, *normalized.shape)).to(device)
out = model(new_frame)
new_out = recover_image(out[1].data.cpu().numpy())
# out_writer.write(new_out)
# new_out = new_out.reshape((1, *new_out.shape))
frame_lst.append(new_out)
else:
break
cap.release()
# out_writer.release()
cv2.destroyAllWindows()
out_video = np.concatenate(frame_lst)
io.vwrite("../outputImage/new_video_vangogh" + str(video_idx) + ".mp4",
out_video,
outputdict={'-pix_fmt': 'yuv420p'})
def required_video(video_info):
try:
video_path = video_info[0]
print(video_path)
start_time = video_info[1]
end_time = video_info[2]
capture = cv2.VideoCapture(video_path)
fps = capture.get(cv2.CAP_PROP_FPS)
video = vread(video_path)
video = video[math.floor(start_time * fps):math.floor(end_time * fps)]
vwrite(videos_required_path + video_path.split('/')[-1], video)
except:
f = open(log_file, 'a')
f.write(video_path + '\n')
f.close()
def create_video(self):
'''
composes video from .fformat files, requires ffmpeg
'''
fileList = glob.glob(os.path.join(self.directory, '*' + self.format))
total_movie = []
fileList.sort(key=lambda x: os.path.basename(x))
print("Merging up the files. This might take a moment...")
for filename in fileList:
print(filename)
img = ski.imread(filename)
for i in range(self.framerate):
total_movie.append(img)
total_movie = np.array(total_movie)
skv.vwrite(os.path.join(self.directory, 'movie.mkv'), total_movie)
self.do_cleanup(fileList)
def play_clip(frames):
# torch.Tensor -> Numpy
np_frames = np.transpose(frames.numpy(), (0, 2, 3, 1)) * 255
np_frames = np_frames.astype(np.uint8)
# write frames to temp file
temp_file = tempfile.NamedTemporaryFile(suffix='.mp4')
sk.vwrite(temp_file.name, np_frames)
# play video clip
cmd = 'ffplay -i ' + temp_file.name + ' -loop 0'
os.system(cmd)
# close & delete temp file
temp_file.close()
return
def dataset_to_mp4(s, m, mp4_path):
"""Converts the given series to an mp4 video. If the mask is given, adds an outline around each neuron.
# Arguments
s: imaging series as a (time x height x width) numpy array.
m: neuron masks as a (no. neurons x height x width) numpy array.
mp4_path: path where the mp4 file should be saved.
# Returns
Nothing
"""
from skvideo.io import vwrite
logger = logging.getLogger(funcname())
logger.info('Preparing video %s.' % mp4_path)
s = s.astype(np.float32)
s = (s - np.min(s)) / (np.max(s) - np.min(s)) * 255
# If mask is given make a color video with neurons outlined.
if m is not None:
video = np.zeros(s.shape + (3, ), dtype=np.uint8)
video[:, :, :, 0] = s
video[:, :, :, 1] = s
video[:, :, :, 2] = s
outlines = np.zeros(s.shape[1:] + (3, ), )
for i in range(m.shape[0]):
reg = one(list(zip(*np.where(m[i] == 1))))
outlines += reg.mask(dims=m.shape[1:],
fill=None,
stroke='white',
background='black')
yy, xx, _ = np.where(outlines != 0)
video[:, yy, xx, :] = [102, 255, 255]
else:
video = s.astype(np.uint8)
vwrite(mp4_path, video)
logger.info('Saved video %s.' % mp4_path)
def write_images():
root = '../UCF-14/'
classes = open('../dataset/classInd_14.txt', 'r')
var1 = {}
for line in classes:
words = line.split(" ")
var1[words[1].split("\n")[0]] = words[0]
for path, subdirs, files in os.walk(root):
for filename in files:
print filename
if ".DS_Store" not in filename:
folder = 'images' + '/' + filename.split('.')[0] + '/'
if not os.path.isdir(folder):
os.mkdir(folder)
else:
shutil.rmtree(folder)
os.mkdir(folder)
try:
cnt = 0
full_path = path + '/' + filename
cap = vreader(full_path)
fcnt = 1
for frame in cap:
vid_name = filename.split('.')[0]
img_path = folder + vid_name + '_{}.jpg'.format(cnt +
1)
img_name = vid_name + '_{}'.format(cnt + 1)
if fcnt % 5 == 0:
vwrite(img_path, frame)
cnt = cnt + 1
fcnt += 1
if cnt:
with open("count.txt", "w") as txt:
text = str(cnt) + " " + img_name.split(
'.')[0] + "\n"
txt.write(text)
except (RuntimeError, TypeError, NameError):
print "Some Error happened"
def save_images(self, webpage, visuals, image_path, aspect_ratio=1.0):
image_dir = webpage.get_image_dir()
short_path = ntpath.basename(image_path[0])
name = os.path.splitext(short_path)[0]
webpage.add_header(name)
ims = []
txts = []
links = []
if self.no_img:
for label, im in visuals.items():
self.save_to_mat(webpage, name, label, im)
else:
for label, im in visuals.items():
self.save_to_mat(webpage, name, label, im) # save a matfile
if self.video_mode:
vid_name = '%s_%s.mp4' % (name, label)
vid_path = os.path.join(image_dir, vid_name)
# video_numpy = util.rescale_vid(video_numpy, 2)
vwrite(vid_path, im, inputdict=self.ffmpeg_params)
ims.append(vid_name)
links.append(vid_path)
else:
image_name = '%s_%s.png' % (name, label)
save_path = os.path.join(image_dir, image_name)
h, w, _ = im.shape
if aspect_ratio > 1.0:
# im = imresize(im, (h, int(w * aspect_ratio)), interp='bicubic')
im = np.array(
Image.fromarray(im).resize(
(h, int(w * aspect_ratio)), Image.BICUBIC))
if aspect_ratio < 1.0:
# im = imresize(im, (int(h / aspect_ratio), w), interp='bicubic')
im = np.array(
Image.fromarray(im).resize(
(int(h / aspect_ratio), w), Image.BICUBIC))
util.save_image(im, save_path)
ims.append(image_name)
links.append(image_name)
txts.append(label)
webpage.add_images(ims, txts, links, width=self.win_size)
def make_video_frames(datadir, outdir):
# train video data
for i in range(10000):
if not os.path.exists(os.path.join(outdir, 'video%s' % i)):
os.makedirs(os.path.join(outdir, 'video%s' % i))
filename = ('video%s.mp4' % i)
videopath = os.path.join(datadir, filename)
cap = sk.vreader(videopath)
metadata = sk.ffprobe(videopath)
# print json.dumps(metadata["video"], indent=4)
"""
fps : @r_frame_rate
length : @duration
frames : @nb_frames
"""
length = float(
json.dumps(metadata["video"]["@duration"]).replace('"', ''))
frames = float(
json.dumps(metadata["video"]["@nb_frames"]).replace('"', ''))
fps = int(frames / length)
print('%sth video' % i)
print('length : %d / frames : %d / fps : %d' % (length, frames, fps))
cent = np.linspace(0, length, 7)[1:-1]
for x in range(len(cent)):
cent[x] = int(cent[x])
frames = cent * fps
idx = 0
filenum = 0
for frame in cap:
if idx in frames:
frame = cv2.resize(frame, dsize=(224, 224))
sk.vwrite(outdir + '/video%s/frame%s.png' % (i, filenum),
frame)
filenum += 1
idx += 1
if i % 1000 == 0:
print('%sth video processed...' % i)
def main():
sess = tf.Session()
in_image = tf.placeholder(tf.float32, [None, 16, None, None, 4])
gt_image = tf.placeholder(tf.float32, [None, 16, None, None, 3])
out_image = []
out_image += network(in_image),
out_image += network(in_image, 3, prefix='1'),
out_image += network(in_image, 2, prefix='2'),
out_image += network(in_image, 3, 16, prefix='3'),
from skvideo.io import vread, vwrite
import numpy as np
vid = np.load('./0_data/raw/test_data/gt_input/001_00_0001.npy')
vid = np.expand_dims(np.float32(np.minimum((vid[:16, :256, :256, :] / vid.mean() / 5), 1.0)), axis=0)
sess.run(tf.global_variables_initializer())
for i, out in enumerate(out_image):
print out.shape
output = sess.run(out, feed_dict={in_image: vid})
output = (np.minimum(output, 1.0) * 255).astype('uint8')
print output[0].shape
vwrite(str(i) + '.mp4', output[0])
def np_array_to_mp4(array, file_name):
"""
Parameters
----------
array : np.ndarray
Video data. Dimensions = (t, y, x, color)
file_name : str
Name of mp4 file to write the data to
Examples
--------
Example 1:: # black and white
import numpy as np
array = np.random.random(size=(50, 480, 680, 1)) * 255
array = array.astype(np.uint8)
file_name="outputvideo_BW.mp4"
np_array_to_mp4(array,file_name=file_name)
Example 2:: # color
import numpy as np
array = np.random.random(size=(50, 480, 680, 3)) * 255
array = array.astype(np.uint8)
file_name="outputvideo_color.mp4"
np_array_to_mp4(array,file_name=file_name)
References
----------
* http://www.scikit-video.org/stable/io.html
"""
import skvideo.io as _skio
print("work in progress")
if not file_name.endswith('.mp4'):
file_name += '.mp4'
_skio.vwrite(file_name, array)
def main():
filename = 'videos/walking.MOV'
video = io.vread(filename, as_grey=True)
depths = []
for i in range(70, 340):
filename = 'imgs/{}_cam-depth_array_.png'.format(i)
d = cv2.imread(filename, -1)
d = cv2.GaussianBlur(d, (9, 7), 0)
depths.append(d)
imgs = []
offset = 5
failure = 0
for i, frame in enumerate(video):
print(i)
frame = frame[:, :, 0]
if i < offset:
res = np.zeros(frame.shape)
res2 = np.zeros(frame.shape)
res3 = np.zeros(frame.shape)
else:
try:
detector = Detector(video[i - offset], frame,
depths[i - offset], depths[i])
res = detector.detect()
kp1, kp2, matches = detector.match()
res2 = detector.detect_2D(kp1, kp2, matches)
res3 = detector.detect_3D(kp1, kp2, matches)
except:
failure += 1
res = np.zeros(frame.shape)
res2 = np.zeros(frame.shape)
res3 = np.zeros(frame.shape)
top = np.hstack((frame, res))
bottom = np.hstack((res2, res3))
img = np.vstack((top, bottom))
imgs.append(img)
print("failure: ", failure)
io.vwrite('walking_detected.MOV', imgs)
def generate_video(base_dir,
step_number,
frames,
reward,
name='',
results=[]): #, resize=True):
print('converting to np array uint8')
frames_for_video = np.array(frames).astype(np.uint8)
if len(frames_for_video[0].shape) == 2:
name += 'gray'
else:
name += 'color'
mp4_fname = os.path.join(
base_dir,
"ATARI_step%010d_r%04d_%s.mp4" % (step_number, int(reward), name))
for i in range(0, len(frames_for_video), 500):
end = min(i + 500, len(frames_for_video))
gname = mp4_fname.replace('.mp4', '%05d.mp4' % i)
print("WRITING video", gname, len(frames_for_video))
vwrite(mp4_fname, frames_for_video[i:end])
def read_mesh_video(videoname,
framesdir,
shape=(480, 640),
dtypergb=np.uint8,
dtypedepth=np.uint16):
rgb_data = read_frame_data(framesdir,
framenofunction=rgb_std_naming_frameno,
shape=shape + (3, ),
dtype=dtypergb,
framesregex="\d+\.rgb") * 0.5
depth_data = read_frame_data(framesdir,
framenofunction=z16_std_naming_frameno,
shape=shape + (1, ),
dtype=dtypedepth,
framesregex="\d+\.z16") * 0.5
depth_data = grey_to_redblue_codec(np.array(depth_data, dtype=np.long))
vid_data = depth_data + rgb_data
skio.vwrite(join(VIDDIR, videoname), vid_data)
def convert_videos(self, square_size, save_folder, reduce_frames=True):
x_paths = self.get_paths()
failed = []
for row in x_paths.itertuples():
filepath = row.filepath
obf_id = row.Index
# load
video = skv.vread(filepath, as_grey=False)
self.height = video.shape[1]
self.width = video.shape[2]
# fill video if neccessary
if video.shape[0] < self.num_frames:
video = self._fill_video(video)
# reduce
if reduce_frames:
frames = np.arange(0, video.shape[0], 2)
try:
video = video[frames, :, :] #.squeeze()
except IndexError:
if verbose:
print("FAILED TO REDUCE:", filepath)
print("id:\t", obf_id)
failed.append(obf_id)
#reshape
video = np.asarray([imresize(x, (224, 224, 3), interp='bilinear') for x in video])
#save
save_path = os.path.join(save_folder, obf_id)
skv.vwrite(save_path, video)
return failed
def plot_trajectory_on_video(vid_path, out_path, ball_trajectory, radius=10):
"""
Mark each detected ball with a red circle in a video
:param vid_path: path to the input video
:param out_path: path where to save the video
:param ball_trajectory: array of ball detections (as outputted by detect_ball function)
:param radius: radius of the circle marking the ball
"""
cap = cv2.VideoCapture(vid_path)
arr = []
i = 0
count = 0
while True:
ret, img = cap.read()
if ret == False:
break
ori_img = img.copy()
out_img = img.copy()
if count < len(ball_trajectory) and ball_trajectory[count, 2] == i:
scale = int(255 / (count + 1))
for j in range(count + 1):
weight = count - j
# print(j, weight*scale)
center = tuple(ball_trajectory[j, :2].astype(int))
img = cv2.circle(ori_img,
center,
int(radius),
[255, weight * scale, weight * scale],
thickness=2)
# print(j, weight, anti_weight)
count += 1
#plt.figure(figsize = (10,10))
#plt.imshow(img[30:, 20:])
#plt.show()
i += 1
# append multiple times to have slow motion like video
for _ in range(5):
arr.append(img[30:, 20:])
io.vwrite(out_path, arr)
def test_from_supervisely_video():
import supervisely_lib as sly
from skvideo.io import vwrite
data_path = "./data/test_supervisely/video_from"
if os.path.exists(data_path):
shutil.rmtree(data_path)
project_name = "minuscule_videos/"
project_path = os.path.join(data_path, project_name)
project = sly.VideoProject(project_path, sly.OpenMode.CREATE)
project.meta._project_type = "videos"
item_name = "item.mp4"
np.random.seed(0)
for name in ["foofoo", "bar"]:
ds = project.create_dataset(name)
item_path = os.path.join(ds.item_dir, item_name)
vwrite(item_path,
(np.random.rand(len(name), 2, 2, 3) * 255).astype("uint8"))
ds._item_to_ann[item_name] = item_name + ".json"
ds.set_ann(item_name, ds._get_empty_annotaion(item_path))
project.set_meta(project.meta)
trans = hub.Dataset.from_supervisely(os.path.join(data_path, project_name))