encoder.eval()
with torch.no_grad():
all_ref = []
all_pred = []
#print('to device finish')
for i, (images, batch_captions) in enumerate(BLEU4loader):
if i >= 40:
continue
all_ref.extend(batch_captions)
images = images.to(device)
#all_ref.extend(batch_captions)
# Generate an caption from the image
feature = encoder(images)
all_pred.extend(decoder.beam_search(feature))
#for a_pred in decoder.beam_search(feature):
# print(a_pred[-1] == EDKidx)
# if a_pred[-1] == EDKidx:
# all_pred.append(a_pred[1:-1])
# else:
# all_pred.append(a_pred[1:])
#print('append finished')
#print('pred: {}'.format(all_pred[-1]))
#print('true: {}'.format(all_ref[-1]))
if (i+1) % 10 == 0:
print('model_idx {}, Step [{}/{}], shape: {}, {}'.format(
model_idx, i+1, total_step,
(len(all_ref),len(all_ref[0]),len(all_ref[0][0])),
(len(all_pred),len(all_pred[0]))))
#print(type(all_ref[0]),type(all_pred),type(all_pred[0]))
decoder.load_state_dict(state_dict['decoder_state_dict'])
break
# test
decoder.eval()
encoder.eval()
with torch.no_grad():
# Prepare an image
images, original_images = load_images(img_paths, transform)
images = images.to(device)
# Generate an caption from the image
feature = encoder(images)
print('Encoder finished')
pred_ids = decoder.beam_search(feature)
print('beam search finished')
# Convert word_ids to words
pred_captions = []
for pred_id in pred_ids:
temp = []
for word_id in pred_id:
temp.append(idx2word[word_id])
if temp[-1] == '<end>':
#pred_captions.append(' '.join(temp))
break
if len(temp) > 8:
temp[len(temp) // 2] = temp[len(temp) // 2] + '\n'
pred_captions.append(' '.join(temp))
print('finished caption generation')
