parser.add_argument("--features_path", type=str)
parser.add_argument("--annotation_folder", type=str)
args = parser.parse_args()
print("Meshed-Memory Transformer Evaluation")
# Pipeline for image regions
image_field = ImageDetectionsField(detections_path=args.features_path,
max_detections=50,
load_in_tmp=False)
# Pipeline for text
text_field = TextField(
init_token="<bos>",
eos_token="<eos>",
lower=True,
tokenize="spacy",
remove_punctuation=True,
nopoints=False,
)
# Create the dataset
dataset = COCO(
image_field,
text_field,
"coco/images/",
args.annotation_folder,
args.annotation_folder,
)
_, _, test_dataset = dataset.splits
text_field.vocab = pickle.load(open("vocab.pkl", "rb"))
parser.add_argument('--resume_best', action='store_true')
parser.add_argument('--features_path', type=str)
parser.add_argument('--annotation_folder', type=str)
parser.add_argument('--logs_folder', type=str, default='tensorboard_logs')
args = parser.parse_args()
print(args)
print('Meshed-Memory Transformer Training')
writer = SummaryWriter(log_dir=os.path.join(args.logs_folder, args.exp_name))
# Pipeline for image regions
image_field = ImageDetectionsField(detections_path=args.features_path, max_detections=50, load_in_tmp=False)
# Pipeline for text
text_field = TextField(init_token='<bos>', eos_token='<eos>', lower=True, tokenize='spacy',
remove_punctuation=True, nopoints=False)
# Create the dataset
dataset = COCO(image_field, text_field, 'coco/images/', args.annotation_folder, args.annotation_folder)
train_dataset, val_dataset, test_dataset = dataset.splits
if not os.path.isfile('vocab_%s.pkl' % args.exp_name):
print("Building vocabulary")
text_field.build_vocab(train_dataset, val_dataset, min_freq=5)
pickle.dump(text_field.vocab, open('vocab_%s.pkl' % args.exp_name, 'wb'))
else:
text_field.vocab = pickle.load(open('vocab_%s.pkl' % args.exp_name, 'rb'))
# Model and dataloaders
encoder = MemoryAugmentedEncoder(3, 0, attention_module=ScaledDotProductAttentionMemory,
attention_module_kwargs={'m': args.m})
type=str,
default=
"/cluster/sorona/dchen/faster_rcnn_R_101_DC5_3x_ScanNet_feats.hdf5")
args = parser.parse_args()
print('Meshed-Memory Transformer Evaluation')
# Pipeline for image regions
image_field = ScanNetDetectionsField(detections_path=args.features_path,
max_detections=50,
load_in_tmp=False)
# Pipeline for text
text_field = TextField(init_token='<bos>',
eos_token='<eos>',
lower=True,
tokenize='spacy',
remove_punctuation=True,
nopoints=False)
# Create the dataset
dataset = ScanNet(image_field, text_field,
"/cluster/sorona/dchen/ScanNet_frames/",
get_image_ids(args.features_path))
_, _, test_dataset = dataset.splits
text_field.vocab = pickle.load(open('vocab.pkl', 'rb'))
# Model and dataloaders
encoder = MemoryAugmentedEncoder(
3,
0,
attention_module=ScaledDotProductAttentionMemory,