def extract_frames((vid_file, time, label, vid_idx, im_dir, prev_free_time)): examples = [] with ut.TmpDir() as tmp_dir: free_dur = 0.1 ut.sys_check('ffmpeg -loglevel warning -ss %f -i "%s" -vf scale=%d:%d -t %f -r %d "%s/%%07d.png"' % \ (prev_free_time, vid_file, full_dim, full_dim, free_dur, sample_fps, tmp_dir)) #fname = sorted(ut.glob(pj(tmp_dir, '*.png')))[0] assert len(ut.glob(pj(tmp_dir, '*.png'))), 'no frames for prev_free_time' fname = random.choice(sorted(ut.glob(pj(tmp_dir, '*.png')))) prev_file = pj(im_dir, 'prev_%s_%05d_%d.png' % (vid_idx, 0, label)) ut.sys_check('cp %s %s' % (fname, prev_file)) with ut.TmpDir() as tmp_dir: # ut.sys_check('ffmpeg -i "%s" -vf scale=%d:%d -ss %f -t %f -r %d "%s/%%07d.png"' % \ # (vid_file, full_dim, full_dim, time, # sample_dur_secs, sample_fps, tmp_dir)) ut.sys_check('ffmpeg -loglevel warning -ss %f -i "%s" -vf scale=%d:%d -t %f -r %d "%s/%%07d.png"' % \ (time, vid_file, full_dim, full_dim, sample_dur_secs, sample_fps, tmp_dir)) for frame_idx, fname in enumerate(sorted(ut.glob(pj(tmp_dir, '*.png')))): im_file = pj(im_dir, '%s_%05d_%d.png' % (vid_idx, frame_idx, label)) ut.sys_check('cp %s %s' % (fname, im_file)) examples.append((im_file, prev_file, label, vid_file)) return examples
def run(vid_file, start_time, dur, pr, gpu, buf=0.05, mask=None, arg=None, net=None): print pr dur = dur + buf with ut.TmpDir() as vid_path: height_s = '-vf "scale=-2:\'min(%d,ih)\'"' % arg.max_full_height if arg.max_full_height > 0 else '' ut.sys_check( ut.frm( 'ffmpeg -loglevel error -ss %(start_time)s -i "%(vid_file)s" -safe 0 ' '-t %(dur)s -r %(pr.fps)s -vf scale=256:256 "%(vid_path)s/small_%%04d.png"' )) ut.sys_check( ut.frm( 'ffmpeg -loglevel error -ss %(start_time)s -i "%(vid_file)s" -safe 0 ' '-t %(dur)s -r %(pr.fps)s %(height_s)s "%(vid_path)s/full_%%04d.png"' )) ut.sys_check( ut.frm( 'ffmpeg -loglevel error -ss %(start_time)s -i "%(vid_file)s" -safe 0 ' '-t %(dur)s -ar %(pr.samp_sr)s -ac 2 "%(vid_path)s/sound.wav"') ) if arg.fullres: fulls = map( ig.load, sorted(ut.glob(vid_path, 'full_*.png'))[:pr.sampled_frames]) fulls = np.array(fulls) snd = sound.load_sound(pj(vid_path, 'sound.wav')) samples_orig = snd.normalized().samples samples_orig = samples_orig[:pr.num_samples] samples_src = samples_orig.copy() if samples_src.shape[0] < pr.num_samples: return None ims = map(ig.load, sorted(ut.glob(vid_path, 'small_*.png'))) ims = np.array(ims) d = 224 y = x = ims.shape[1] / 2 - d / 2 ims = ims[:, y:y + d, x:x + d] ims = ims[:pr.sampled_frames] if mask == 'l': ims[:, :, :ims.shape[2] / 2] = 128 if arg.fullres: fulls[:, :, :fulls.shape[2] / 2] = 128 elif mask == 'r': ims[:, :, ims.shape[2] / 2:] = 128 if arg.fullres: fulls[:, :, fulls.shape[2] / 2:] = 128 elif mask is None: pass else: raise RuntimeError() samples_src = mu.normalize_rms_np(samples_src[None], pr.input_rms)[0] net.init() ret = net.predict(ims[None], samples_src[None]) samples_pred_fg = ret['samples_pred_fg'][0][:, None] samples_pred_bg = ret['samples_pred_bg'][0][:, None] spec_pred_fg = ret['spec_pred_fg'][0] spec_pred_bg = ret['spec_pred_bg'][0] print spec_pred_bg.shape spec_mix = ret['spec_mix'][0] if arg.cam: cam, vis = find_cam(fulls, samples_orig, arg) else: if arg.fullres: vis = fulls else: vis = ims return dict(ims=vis, samples_pred_fg=samples_pred_fg, samples_pred_bg=samples_pred_bg, samples_mix=ret['samples_mix'][0], samples_src=samples_src, spec_pred_fg=spec_pred_fg, spec_pred_bg=spec_pred_bg, spec_mix=spec_mix)