def main():
args = parse_args()
if args.dataset == 'breakfast':
train_segments, _, _ = breakfast.get_data(split='train')
else:
raise ValueError('no such dataset')
video_names = []
for segment in train_segments:
video_name = segment['video-name']
video_names.append(video_name)
video_names = np.unique(video_names).reshape(-1)
video_dirs = [os.path.join(breakfast.EXTRACTED_IMAGES_DIR, video) for video in video_names]
mean, std = compute_mean_std(video_dirs=video_dirs)
np.save(breakfast.MEAN_FILE, mean)
np.save(breakfast.STD_FILE, std)
def _parse_split_data(split, feat_len):
segments, labels, logits = breakfast.get_data(split)
valid_segments = []
valid_labels = []
valid_logits = []
for i, segment in enumerate(tqdm(segments)):
start, end = segment['start'], segment['end']
i3d_feats = breakfast.read_i3d_data(segment['video-name'],
window=[start, end],
i3d_length=feat_len)
if len(i3d_feats
) > 0 and 48 > logits[i] > 0: # remove walk in and walk out.
segment['end'] = segment['start'] + len(i3d_feats)
valid_segments.append(segment)
valid_labels.append(labels[i])
valid_logits.append(logits[i])
return [valid_segments, valid_labels, valid_logits]
def _parse_split_data(split):
segments, labels, logits = breakfast.get_data(split)
valid_segments = []
valid_labels = []
valid_logits = []
for i, segment in enumerate(tqdm(segments)):
start, end = segment['start'], segment['end']
video_name = segment['video-name']
video_dir = os.path.join(breakfast.EXTRACTED_IMAGES_DIR, video_name)
n_video_frames = len(os.listdir(video_dir))
end = min(n_video_frames, end)
n_frames = end - start
if n_frames > 0 and 48 > logits[i] > 0: # remove walk in and walk out.
segment['end'] = end
valid_segments.append(segment)
valid_labels.append(labels[i])
valid_logits.append(logits[i])
return [valid_segments, valid_labels, valid_logits]
def _display_distribution(labels, n_labels):
for label in labels:
assert label < n_labels, 'some labels are not smaller than {}'.format(
n_labels)
fig = plt.hist(x=labels, bins=n_labels)
plt.title('Test Label Distribution')
plt.xlabel("Labels")
plt.ylabel("Frequency")
plt.show()
if __name__ == '__main__':
_, labels, logits = breakfast.get_data(split='test')
logits = [logit for logit in logits if logit]
logits = [logit for logit in logits if logit < 48]
# todo: uncomment to view sorted version
# logit_counter = Counter(logits)
# most_common = logit_counter.most_common()
# remap = {}
# for i in range(len(most_common)):
# mc_tuple = most_common[i]
# remap[mc_tuple[0]] = i
# logits = [remap[logit] for logit in logits]
logits = np.array(logits).reshape(-1)
_display_distribution(logits, n_labels=max(logits) + 1)