oracle = Oracle(
no_words = oracle_args['vocab_size'],
no_words_feat = oracle_args['embeddings']['no_words_feat'],
no_categories = oracle_args['embeddings']['no_categories'],
no_category_feat = oracle_args['embeddings']['no_category_feat'],
no_hidden_encoder = oracle_args['lstm']['no_hidden_encoder'],
mlp_layer_sizes = oracle_args['mlp']['layer_sizes'],
no_visual_feat = oracle_args['inputs']['no_visual_feat'],
no_crop_feat = oracle_args['inputs']['no_crop_feat'],
dropout = oracle_args['lstm']['dropout'],
inputs_config = oracle_args['inputs'],
scale_visual_to = oracle_args['inputs']['scale_visual_to']
)
oracle = load_model(oracle, oracle_args['bin_file'], use_dataparallel=use_dataparallel)
oracle.eval()
print(model)
print(oracle)
if args.resnet:
cnn = ResNet()
if use_cuda:
cnn.cuda()
cnn = DataParallel(cnn)
cnn.eval()
softmax = nn.Softmax(dim=-1)
if args.resnet:
for split, dataset in zip(exp_config['splits'], [dataset_train, dataset_validation]):
accuracy = []
dataloader = DataLoader(
dataset=dataset,
batch_size=optimizer_config['batch_size'],
shuffle=True,
num_workers=multiprocessing.cpu_count(),
pin_memory=exp_config['use_cuda']
)
if split == 'train':
model.train()
else:
model.eval()
for i_batch, sample in enumerate(dataloader):
# Get Batch
questions, answers, crop_features, visual_features, spatials, obj_categories, lengths = \
sample['question'], sample['answer'], sample['crop_features'], sample['img_features'], sample['spatial'], sample['obj_cat'], sample['length']
# Forward pass
pred_answer = model(Variable(questions),
Variable(obj_categories),
Variable(spatials),
Variable(crop_features),
Variable(visual_features),
Variable(lengths)
)