split_video.py 文件源码

def np_2_vExample(vid, labs, rgb, audio):
  nframes = audio.shape[0]

  if False:
      # top 5
      k = 5
      if nframes > 10:
        tk_rgb = my_utils.top_k_along_column(rgb, k)
        tk_audio = my_utils.top_k_along_column(audio, k)
      else:
        tk_rgb = np.repeat(rgb[0].reshape([1, rgb.shape[1]]), k, axis=0)
        tk_audio = np.repeat(audio[0].reshape([1, audio.shape[1]]), k, axis=0)      

  # std of all rgb or audio entries
  s_rgb = np.std(rgb)
  s_aud = np.std(audio)

  rgb_sq = rgb * rgb
  aud_sq = audio * audio

  vExample  = tf.train.Example(features=tf.train.Features(feature={
     'video_id':     my_utils._byteslist_feature([vid]),
     'labels':       my_utils._int64list_feature(labs),
     'mean_rgb':     my_utils._floatlist_feature(np.mean(rgb, axis=0)),
     'mean_audio':   my_utils._floatlist_feature(np.mean(audio, axis=0)),
     'std_rgb':      my_utils._floatlist_feature(np.std(rgb, axis=0)),
     'std_audio':    my_utils._floatlist_feature(np.std(audio, axis=0)),
     'x3_rgb':       my_utils._floatlist_feature(np.cbrt(np.mean(rgb_sq * rgb, axis=0))),
     'x3_audio':     my_utils._floatlist_feature(np.cbrt(np.mean(aud_sq * audio, axis=0))),
     'num_frames':   my_utils._floatlist_feature([(nframes-151.)/300.]),
     'std_all_rgb':    my_utils._floatlist_feature([s_rgb]),
     'std_all_audio':  my_utils._floatlist_feature([s_aud])
     }))      

     #'top_1_rgb':    my_utils._floatlist_feature(tk_rgb[-1]),
     #'top_3_rgb':    my_utils._floatlist_feature(tk_rgb[-3]),
     #'top_5_rgb':    my_utils._floatlist_feature(tk_rgb[-5]),
     #'top_1_audio':  my_utils._floatlist_feature(tk_audio[-1]),
     #'top_3_audio':  my_utils._floatlist_feature(tk_audio[-3]),
     #'top_5_audio':  my_utils._floatlist_feature(tk_audio[-5]),

  return vExample

#%%