utils.create_dir(args.output)
logger = utils.Logger(os.path.join(args.output, 'args.txt'))
logger.write(args.__repr__())
batch_size = args.batch_size
constructor = 'build_%s' % args.model
model = getattr(base_model, constructor)(train_dset, args.num_hid, args.op,
args.gamma, args.task).cuda()
tfidf = None
weights = None
if args.tfidf:
dict = Dictionary.load_from_file(dict_path)
tfidf, weights = tfidf_from_questions(['train', 'val', 'test2015'],
dict)
model.w_emb.init_embedding(w_emb_path, tfidf, weights)
model = nn.DataParallel(model).cuda()
optim = None
epoch = 0
# load snapshot
if args.input is not None:
print('loading %s' % args.input)
model_data = torch.load(args.input)
model.load_state_dict(model_data.get('model_state', model_data))
optim = torch.optim.Adamax(
filter(lambda p: p.requires_grad, model.parameters()))
pos_emb_dim=args.imp_pos_emb_dim,
dataroot=args.data_folder)
train_dset = VQAFeatureDataset('train',
dictionary,
args.relation_type,
adaptive=args.adaptive,
pos_emb_dim=args.imp_pos_emb_dim,
dataroot=args.data_folder)
model = build_regat(val_dset, args).to(device)
tfidf = None
weights = None
if args.tfidf:
tfidf, weights = tfidf_from_questions(['train', 'val', 'test2015'],
dictionary,
dataroot=args.data_folder)
model.w_emb.init_embedding(
join(args.data_folder, 'glove/glove6b_init_300d.npy'), tfidf, weights)
model = nn.DataParallel(model).to(device)
if args.checkpoint != "":
print("Loading weights from %s" % (args.checkpoint))
if not os.path.exists(args.checkpoint):
raise ValueError("No such checkpoint exists!")
checkpoint = torch.load(args.checkpoint)
state_dict = checkpoint.get('model_state', checkpoint)
matched_state_dict = {}
unexpected_keys = set()
missing_keys = set()
