|
| 1 | +'''Example streaming ffmpeg numpy processing. |
| 2 | +
|
| 3 | +Demonstrates using ffmpeg to decode video input, process the frames in |
| 4 | +python, and then encode video output using ffmpeg. |
| 5 | +
|
| 6 | +This example uses two ffmpeg processes - one to decode the input video |
| 7 | +and one to encode an output video - while the raw frame processing is |
| 8 | +done in python with numpy. |
| 9 | +
|
| 10 | +At a high level, the signal graph looks like this: |
| 11 | +
|
| 12 | + (input video) -> [ffmpeg process 1] -> [python] -> [ffmpeg process 2] -> (output video) |
| 13 | +
|
| 14 | +
F438
This example reads/writes video files on the local filesystem, but the |
| 15 | +same pattern can be used for other kinds of input/output (e.g. webcam, |
| 16 | +rtmp, etc.). |
| 17 | +
|
| 18 | +The simplest processing example simply darkens each frame by |
| 19 | +multiplying the frame's numpy array by a constant value; see |
| 20 | +``process_frame_simple``. |
| 21 | +
|
| 22 | +A more sophisticated example processes each frame with tensorflow using |
| 23 | +the "deep dream" tensorflow tutorial; activate this mode by calling |
| 24 | +the script with the optional `--dream` argument. (Make sure tensorflow |
| 25 | +is installed before running) |
| 26 | +''' |
| 27 | +from __future__ import print_function |
| 28 | +import argparse |
| 29 | +import ffmpeg |
| 30 | +import logging |
| 31 | +import numpy as np |
| 32 | +import os |
| 33 | +import subprocess |
| 34 | +import zipfile |
| 35 | + |
| 36 | + |
| 37 | +parser = argparse.ArgumentParser(description='Example streaming ffmpeg numpy processing') |
| 38 | +parser.add_argument('in_filename', help='Input filename') |
| 39 | +parser.add_argument('out_filename', help='Output filename') |
| 40 | +parser.add_argument( |
| 41 | + '--dream', action='store_true', help='Use DeepDream frame processing (requires tensorflow)') |
| 42 | + |
| 43 | +logger = logging.getLogger(__name__) |
| 44 | +logging.basicConfig(level=logging.INFO) |
| 45 | + |
| 46 | + |
| 47 | +def get_video_size(filename): |
| 48 | + logger.info('Getting video size for {!r}'.format(filename)) |
| 49 | + probe = ffmpeg.probe(filename) |
| 50 | + video_info = next(s for s in probe['streams'] if s['codec_type'] == 'video') |
| 51 | + width = int(video_info['width']) |
| 52 | + height = int(video_info['height']) |
| 53 | + return width, height |
| 54 | + |
| 55 | + |
| 56 | +def start_ffmpeg_process1(in_filename): |
| 57 | + logger.info('Starting ffmpeg process1') |
| 58 | + args = ( |
| 59 | + ffmpeg |
| 60 | + .input(in_filename) |
| 61 | + .output('pipe:', format='rawvideo', pix_fmt='rgb24', vframes=8) |
| 62 | + .compile() |
| 63 | + ) |
| 64 | + return subprocess.Popen(args, stdout=subprocess.PIPE) |
| 65 | + |
| 66 | + |
| 67 | +def start_ffmpeg_process2(out_filename, width, height): |
| 68 | + logger.info('Starting ffmpeg process2') |
| 69 | + args = ( |
| 70 | + ffmpeg |
| 71 | + .input('pipe:', format='rawvideo', pix_fmt='rgb24', s='{}x{}'.format(width, height)) |
| 72 | + .output(out_filename, pix_fmt='yuv420p') |
| 73 | + .overwrite_output() |
| 74 | + .compile() |
| 75 | + ) |
| 76 | + return subprocess.Popen(args, stdin=subprocess.PIPE) |
| 77 | + |
| 78 | + |
| 79 | +def read_frame(process1, width, height): |
| 80 | + logger.debug('Reading frame') |
| 81 | + |
| 82 | + # Note: RGB24 == 3 bytes per pixel. |
| 83 | + frame_size = width * height * 3 |
| 84 | + in_bytes = process1.stdout.read(frame_size) |
| 85 | + if len(in_bytes) == 0: |
| 86 | + frame = None |
| 87 | + else: |
| 88 | + assert len(in_bytes) == frame_size |
| 89 | + frame = ( |
| 90 | + np |
| 91 | + .frombuffer(in_bytes, np.uint8) |
| 92 | + .reshape([height, width, 3]) |
| 93 | + ) |
| 94 | + return frame |
| 95 | + |
| 96 | + |
| 97 | +def process_frame_simple(frame): |
| 98 | + '''Simple processing example: darken frame.''' |
| 99 | + return frame * 0.3 |
| 100 | + |
| 101 | + |
| 102 | +def write_frame(process2, frame): |
| 103 | + logger.debug('Writing frame') |
| 104 | + process2.stdin.write( |
| 105 | + frame |
| 106 | + .astype(np.uint8) |
| 107 | + .tobytes() |
| 108 | + ) |
| 109 | + |
| 110 | + |
| 111 | +def run(in_filename, out_filename, process_frame): |
| 112 | + width, height = get_video_size(in_filename) |
| 113 | + process1 = start_ffmpeg_process1(in_filename) |
| 114 | + process2 = start_ffmpeg_process2(out_filename, width, height) |
| 115 | + while True: |
| 116 | + frame = read_frame(process1, width, height) |
| 117 | + if frame is None: |
| 118 | + logger.info('End of input stream') |
| 119 | + break |
| 120 | + |
| 121 | + logger.debug('Processing frame') |
| 122 | + frame = process_frame(frame) |
| 123 | + write_frame(process2, frame) |
| 124 | + |
| 125 | + logger.info('Waiting for ffmpeg process1') |
| 126 | + process1.wait() |
| 127 | + |
| 128 | + logger.info('Waiting for ffmpeg process2') |
| 129 | + process2.stdin.close() |
| 130 | + process2.wait() |
| 131 | + |
| 132 | + logger.info('Done') |
| 133 | + |
| 134 | + |
| 135 | +class DeepDream(object): |
| 136 | + '''DeepDream implementation, adapted from official tensorflow deepdream tutorial: |
| 137 | + https://github.com/tensorflow/tensorflow/tree/master/tensorflow/examples/tutorials/deepdream |
| 138 | +
|
| 139 | + Credit: Alexander Mordvintsev |
| 140 | + ''' |
| 141 | + |
| 142 | + _DOWNLOAD_URL = 'https://storage.googleapis.com/download.tensorflow.org/models/inception5h.zip' |
| 143 | + _ZIP_FILENAME = 'deepdream_model.zip' |
| 144 | + _MODEL_FILENAME = 'tensorflow_inception_graph.pb' |
| 145 | + |
| 146 | + @staticmethod |
| 147 | + def _download_model(): |
| 148 | + logger.info('Downloading deepdream model...') |
| 149 | + try: |
| 150 | + from urllib.request import urlretrieve # python 3 |
| 151 | + except ImportError: |
| 152 | + from urllib import urlretrieve # python 2 |
| 153 | + urlretrieve(DeepDream._DOWNLOAD_URL, DeepDream._ZIP_FILENAME) |
| 154 | + |
| 155 | + logger.info('Extracting deepdream model...') |
| 156 | + zipfile.ZipFile(DeepDream._ZIP_FILENAME, 'r').extractall('.') |
| 157 | + |
| 158 | + @staticmethod |
| 159 | + def _tffunc(*argtypes): |
| 160 | + '''Helper that transforms TF-graph generating function into a regular one. |
| 161 | + See `_resize` function below. |
| 162 | + ''' |
| 163 | + placeholders = list(map(tf.placeholder, argtypes)) |
| 164 | + def wrap(f): |
| 165 | + out = f(*placeholders) |
| 166 | + def wrapper(*args, **kw): |
| 167 | + return out.eval(dict(zip(placeholders, args)), session=kw.get('session')) |
| 168 | + return wrapper |
| 169 | + return wrap |
| 170 | + |
| 171 | + @staticmethod |
| 172 | + def _base_resize(img, size): |
| 173 | + '''Helper function that uses TF to resize an image''' |
| 174 | + img = tf.expand_dims(img, 0) |
| 175 | + return tf.image.resize_bilinear(img, size)[0,:,:,:] |
| 176 | + |
| 177 | + def __init__(self): |
| 178 | + if not os.path.exists(DeepDream._MODEL_FILENAME): |
| 179 | + self._download_model() |
| 180 | + |
| 181 | + self._graph = tf.Graph() |
| 182 | + self._session = tf.InteractiveSession(graph=self._graph) |
| 183 | + self._resize = self._tffunc(np.float32, np.int32)(self._base_resize) |
| 184 | + with tf.gfile.FastGFile(DeepDream._MODEL_FILENAME, 'rb') as f: |
| 185 | + graph_def = tf.GraphDef() |
| 186 | + graph_def.ParseFromString(f.read()) |
| 187 | + self._t_input = tf.placeholder(np.float32, name='input') # define the input tensor |
| 188 | + imagenet_mean = 117.0 |
| 189 | + t_preprocessed = tf.expand_dims(self._t_input-imagenet_mean, 0) |
| 190 | + tf.import_graph_def(graph_def, {'input':t_preprocessed}) |
| 191 | + |
| 192 | + self.t_obj = self.T('mixed4d_3x3_bottleneck_pre_relu')[:,:,:,139] |
| 193 | + #self.t_obj = tf.square(self.T('mixed4c')) |
| 194 | + |
| 195 | + def T(self, layer_name): |
| 196 | + '''Helper for getting layer output tensor''' |
| 197 | + return self._graph.get_tensor_by_name('import/%s:0'%layer_name) |
| 198 | + |
| 199 | + def _calc_grad_tiled(self, img, t_grad, tile_size=512): |
| 200 | + '''Compute the value of tensor t_grad over the image in a tiled way. |
| 201 | + Random shifts are applied to the image to blur tile boundaries over |
| 202 | + multiple iterations.''' |
| 203 | + sz = tile_size |
| 204 | + h, w = img.shape[:2] |
| 205 | + sx, sy = np.random.randint(sz, size=2) |
| 206 | + img_shift = np.roll(np.roll(img, sx, 1), sy, 0) |
| 207 | + grad = np.zeros_like(img) |
| 208 | + for y in range(0, max(h-sz//2, sz),sz): |
| 209 | + for x in range(0, max(w-sz//2, sz),sz): |
| 210 | + sub = img_shift[y:y+sz,x:x+sz] |
| 211 | + g = self._session.run(t_grad, {self._t_input:sub}) |
| 212 | + grad[y:y+sz,x:x+sz] = g |
| 213 | + return np.roll(np.roll(grad, -sx, 1), -sy, 0) |
| 214 | + |
| 215 | + def process_frame(self, frame, iter_n=10, step=1.5, octave_n=4, octave_scale=1.4): |
| 216 | + t_score = tf.reduce_mean(self.t_obj) # defining the optimization objective |
| 217 | + t_grad = tf.gradients(t_score, self._t_input)[0] # behold the power of automatic differentiation! |
| 218 | + |
| 219 | + # split the image into a number of octaves |
| 220 | + img = frame |
| 221 | + octaves = [] |
| 222 | + for i in range(octave_n-1): |
| 223 | + hw = img.shape[:2] |
| 224 | + lo = self._resize(img, np.int32(np.float32(hw)/octave_scale)) |
| 225 | + hi = img-self._resize(lo, hw) |
| 226 | + img = lo |
| 227 | + octaves.append(hi) |
| 228 | + |
| 229 | + # generate details octave by octave |
| 230 | + for octave in range(octave_n): |
| 231 | + if octave>0: |
| 232 | + hi = octaves[-octave] |
| 233 | + img = self._resize(img, hi.shape[:2])+hi |
| 234 | + for i in range(iter_n): |
| 235 | + g = self._calc_grad_tiled(img, t_grad) |
| 236 | + img += g*(step / (np.abs(g).mean()+1e-7)) |
| 237 | + #print('.',end = ' ') |
| 238 | + return img |
| 239 | + |
| 240 | + |
| 241 | +if __name__ == '__main__': |
| 242 | + args = parser.parse_args() |
| 243 | + if args.dream: |
| 244 | + import tensorflow as tf |
| 245 | + process_frame = DeepDream().process_frame |
| 246 | + else: |
| 247 | + process_frame = process_frame_simple |
| 248 | + run(args.in_filename, args.out_filename, process_frame) |
0 commit comments