def make_normal_labels(mat_files_dir, dst_lab_files_dir, vocab_file):
'''
Creates png files with surface normal labels. Used for training models.
'''
# f = h5py.File(vocab_file, 'r')
vocab = scipy.io.loadmat(vocab_file)
codebook = vocab['vocabs'][0][0]['normals'][0][0]
codebook = codebook.astype(np.float64)
normal_info = load_pickle(
'/data/bw462/nyu/surfacenormal_metadata/all_normals.pklz')
assert NUM_NORMALS_CLASSES == codebook.shape[0]
# whether to create Left-Right Flipped ground truth. This is beta tested only.
flip = False # flip = True is beta
iterator = zip(normal_info['all_normals'],
normal_info['all_valid_depth_masks'],
normal_info['all_filenames'])
for ind, (raw_normals, valid_depth, file_name) in enumerate(iterator):
normal_labels = assign_normals_to_codebook(raw_normals, codebook)
assert normal_labels.min() >= 0 and normal_labels.max(
) < NUM_NORMALS_CLASSES
# now account for valid_depth mask
normal_labels[valid_depth == 0] = IGNORE_LABEL_NORMALS
dst_name = file_name + '_SN%d.png' % (NUM_NORMALS_CLASSES)
dst_file_path = os.path.join(dst_lab_files_dir, dst_name)
im = Image.fromarray(normal_labels.astype(np.uint8))
im.save(dst_file_path)
if ind % 100 == 0:
print('Wrote %s' % (dst_file_path))
if flip:
# first we do a horizontal flip
flipped_normals = raw_normals[:, ::-1, :]
# now we need to do a rotation
# flip the "X" axis
flipped_normals[:, :, 0] *= -1.
def load_normals_file(self):
normals_file = self.normals_file
normals_data = load_pickle(normals_file)
filename_to_id = \
{normals_data['all_filenames'][x]: x for x in range(len(normals_data['all_filenames']))}
def get_normal_and_valid_depth(filename):
id = filename_to_id[filename]
return normals_data['all_normals'][id], normals_data['all_valid_depth_masks'][id]
return get_normal_and_valid_depth
model = load_infersent_model(MODEL_PATH,
bsize=args.bsize,
word_emb_dim=args.word_emb_dim,
enc_lstm_dim=args.enc_lstm_dim,
version=args.version)
sentence_embeddings = []
prefix = os.path.dirname(args.filepath)
if args.streaming:
output_path = os.path.join(
prefix, f'sentence_embeddings_{args.version}_streaming.pkl')
else:
output_path = os.path.join(prefix,
f'sentence_embeddings_{args.version}.pkl')
sentences = load_pickle(args.filepath)
flattened_sentences = [
utterance for conversation in sentences for utterance in conversation
]
print('Encoding sentences ...')
flattened_embeddings = model.encode(flattened_sentences,
tokenize=True,
bsize=64)
print('InferSent encoding done.')
idx = 0
sent_idx = 0
for conversation in sentences:
idx += 1
conversation_embeddings = []
def __init__(self, vocab_file, gt_file=None):
super(NormalsEvaluator, self).__init__(vocab_file)
if gt_file is not None:
self.gt_data = load_pickle(gt_file)
self.metrics = None
argparser.add_argument('--step',
type=int,
default=100,
help="Number of steps for saving output")
argparser.add_argument('--version',
type=int,
default=1,
help="Which model version of inferSent to use. "
"V1 has been trained on GloVe. "
"V2 has been trained on fastText.")
args = argparser.parse_args()
train_dir = os.path.join(args.dataset, 'train')
train_file_path = os.path.join(
train_dir, f'sentence_embeddings_{args.version}_streaming.pkl')
train_embeddings = load_pickle(train_file_path)
flattened_train_embeddings = [
utterance for conversation in train_embeddings
for utterance in conversation
]
if args.fixeddim:
pca_model_file_path = os.path.join(
train_dir, f'v{args.version}_PCA_model_{args.ndim}.pkl')
if args.savepca:
pca_embeddings = fit_pca(flattened_train_embeddings, args.ndim)
dump_pickle(pca_embeddings, pca_model_file_path)
else:
pca_embeddings = load_pickle(pca_model_file_path)
else:
pca_model_file_path = os.path.join(