model.eval()
# init PCA model
pca = PCAWrapper(n_components=cfg.n_components,
batch_size=cfg.pca_batch_size)
pca.load_params(filepath=cfg.pca_model)
# data loader for frames in ingle video
data_loader = get_dataloader(dataset="VideoFrame",
path=cfg.video_file,
num_frames=cfg.num_frames,
batch_size=cfg.batch_size)
# init writer
writer = RecordWriter(filepath="data/test.tfrecord", level="frame")
# extract features by inception_v3
feats = None
for step, frames in enumerate(data_loader):
print("extracting feature [{}/{}]".format(step + 1, len(data_loader)))
feat = model(make_variable(frames))
feat_np = feat.data.cpu().numpy()
# recude dimensions by PCA
feat_ = pca.transform(feat_np)
feats = concat_feat(feats, feat_)
# write features into TFRecord
vid = os.path.splitext(os.path.basename(cfg.video_file))[0]
writer.write(vid=vid, feat_rgb=feats)
writer.close()
model = make_cuda(
inception_v3(pretrained=True, transform_input=True, extract_feat=True))
pca = PCAWrapper(n_components=cfg.n_components)
model.eval()
# data loader for frames in ingle video
# data_loader = get_dataloader(dataset="VideoFrame",
# path=cfg.video_file,
# num_frames=cfg.num_frames,
# batch_size=cfg.batch_size)
# data loader for frames decoded from several videos
data_loader = get_dataloader(dataset="FrameImage",
path=cfg.frame_root,
batch_size=cfg.batch_size)
# extract features by inception_v3
feats = None
for step, frames in enumerate(data_loader):
print("extracting feature [{}/{}]".format(step + 1, len(data_loader)))
feat = model(make_variable(frames))
feats = concat_feat(feats, feat.data.cpu())
# recude dimensions by PCA
X = feats.numpy()
pca.fit(X)
X_ = pca.transform(X)
print("reduce X {} to X_ {}".format(X.shape, X_.shape))
# sabe PCA params
pca.save_params(filepath=cfg.pca_model)
for record in tf.python_io.tf_record_iterator(
os.path.join(cfg.record_root, record_file)):
result = Reader(record)
if not os.path.exists(
cfg.inception_v3_feats_path.format(result.vid)):
print("skipping {}".format(result.vid))
continue
else:
print("extracting {} - labels {}".format(
result.vid, result.labels))
print("--> load fatures")
feats = torch.load(
cfg.inception_v3_feats_path.format(result.vid))
print("--> recude dimensions by PCA")
feats_ = pca.transform(feats.numpy())
print("--> write to tfrecord")
writer.write(vid=result.vid,
feat_rgb=feats_,
labels=result.labels)
feat_counter += 1
if (feat_counter + 1) % cfg.feats_per_file == 0:
print(">>> saving tfrecord: {}".format(
cfg.extract_feat_path.format(save_counter)))
writer.close()
save_counter += 1
feat_counter = 0
writer = RecordWriter(
filepath=cfg.extract_feat_path.format(save_counter),
level="frame")