def infer():
def process(inputs):
inputs = clear_punc(inputs)
x = []
for word in jieba.cut(inputs):
x.append(token2idx.get(word, token2idx['<UNK>']))
x.append(token2idx['</S>'])
x = x + [token2idx['<PAD>']] * (model_params.maxlen - len(x))
x = [x]
return x
model_params = Params()
idx2token, token2idx = load_vocab(model_params.idx2token_path,model_params.token2idx_path)
model = Transformer(model_params)
model.eval()
with tf.Session() as sess:
saver = tf.train.Saver()
last_ckpt = tf.train.latest_checkpoint(model_params.model_save)
saver.restore(sess,last_ckpt)
while True:
x = input('{}：>>'.format('笑给我看'))
# todo 古诗模式
if x == '对古诗':
pass
x = process(x)
feed_dict = {model.xs: x}
y_hat = sess.run(model.y_hat,
feed_dict=feed_dict)
result = ''
for word in y_hat[0]:
if word == token2idx['<UNK>']:
result += '*××'
elif word != 3:
result += idx2token[word]
else:
break
if result == '==':
result = "= ="
elif result == '<UNK>':
result = '哎呀，我不知道啊！'
print('傻逼一号：>>',result,'\n')
feedback_size=2048,
num_head=8,
dropout=0.1,
device=device)
model = model.to(device)
model.load_state_dict(torch.load(args.model_path))
dataload = DataLoader(Dataload(add_(args.bin_path) + 'test',
src=args.src_lang,
trg=args.tgt_lang),
batch_size=32,
collate_fn=collate_fn)
real = []
predict = []
pbtr = tqdm(total=len(dataload))
with torch.no_grad():
model.eval()
for src, trg in dataload:
src = src.to(device)
predicts = beamsearch(model, src, 1, 100, device=device)
for i in range(len(predicts)):
while 0 in predicts[i]:
predicts[i].remove(0)
predict = predict + predicts
real = real + trg
tmp_bleu = corpus_bleu(list_of_references=[[each] for each in trg],
hypotheses=predicts)
pbtr.set_postfix({'bleu': tmp_bleu})
print(tmp_bleu)
pbtr.update(1)
print("Bleu score:{}".format(