def predict(name, command):
command = command.lower()
label_path = path.join(path.dirname(path.realpath(__file__)), "intents",
"config", "labels", "%s_labels.json" % name)
with open(label_path, encoding="utf8") as f:
labels = json.load(f)
word_vocab = Vocabulary()
word_vocab.load("%s_word_vocab.json" % name)
#char embedding
char_vocab = Vocabulary()
char_vocab.load("%s_char_vocab.json" % name)
idx2label = dict((idx, label) for idx, label in enumerate(labels))
preprocessor = Preprocessor(word_vocab, None, char_vocab)
model = BiLSTMCRF(labels, len(word_vocab), len(char_vocab))
model.load_weights('intents/config/weights/%s.hdf5' % name)
sentence = tokenize(command)
features = preprocessor.transform([sentence])
p = model.predict(features)
predicted_labels = []
for pred in p:
predicted_labels.append(idx2label[pred])
for word, label in zip(sentence, predicted_labels):
print('%s: %s' % (word, label))
def single_predict():
vocab_size = 4688
embed_size = 128
units = 64
num_tags = 4
_, _, char_index_dict, index_char_dict = open_file("./data/data.txt")
optimizer = tf.keras.optimizers.Adagrad(learning_rate=0.1)
my_model = BiLSTMCRF(vocab_size, embed_size, units, num_tags)
ckpt = tf.train.Checkpoint(optimizer=optimizer, my_model=my_model)
ckpt.restore(tf.train.latest_checkpoint("./save_checkpoint/"))
text = input_text()
char_index_list = [char_index_dict.get(char, 0) for char in text]
text_list = [char for char in text]
tag_list = ['b', 'm', 'e', 's']
inputs = tf.keras.preprocessing.sequence.pad_sequences([char_index_list],
padding='post')
#predict得到numpy矩阵
logits, inputs_length = my_model.predict(inputs)
#viterbi_decode得到最优路径
path, _ = tfa_crf.viterbi_decode(logits[0], my_model.transition_params)
path_list = [tag_list[index] for index in path]
new_path_list = tag_finetune(path_list)
#衡量标签路径更改的程度
print("标签正常率%.2f%%" %
(100 * sum([i1 == i2 for i1, i2 in zip(path_list, new_path_list)]) /
len(path_list)))
seg_text(text, new_path_list)
def predict(text, config, params, is_export=False):
"""模型预测。"""
# 读取词典
vocab2id, id2vocab = read_vocab(config["vocab_file"])
tag2id, id2tag = read_vocab(config["tag_file"])
# 构建模型
model = BiLSTMCRF(hidden_num=params["hidden_num"],
vocab_size=len(vocab2id),
label_size=len(tag2id),
embedding_size=params["embedding_size"])
model.load_weights(config["ckpt_path"])
# 数据预处理
dataset = tf.keras.preprocessing.sequence.pad_sequences(
[[vocab2id.get(char, 0) for char in text]],
padding='post',
maxlen=params["maxlen"])
# 模型预测
result = model.predict(dataset)[0]
result = np.argmax(result, axis=-1)
result = [id2tag[i] for i in result]
print(result)
# 结果处理
entities_result = format_result(list(text), result)
print(json.dumps(entities_result, indent=4, ensure_ascii=False))
if is_export:
# 导出模型
tf.keras.models.save_model(model,
config["export_dir"],
overwrite=True,
include_optimizer=True,
save_format=None,
options=None)
def predict(text, config, params):
"""模型预测。"""
# 读取词典
vocab2id, id2vocab = read_vocab(config["vocab_file"])
tag2id, id2tag = read_vocab(config["tag_file"])
# 构建模型
model = BiLSTMCRF(
hidden_num=params["hidden_num"], vocab_size=len(vocab2id),
label_size=len(tag2id), embedding_size=params["embedding_size"])
model.load_weights(config["ckpt_path"])
# 数据预处理
dataset = tf.keras.preprocessing.sequence.pad_sequences(
[[vocab2id.get(char, 0) for char in text]], padding='post')
# 模型预测
result = model.predict(dataset)[0]
result = np.argmax(result, axis=-1)
result = [id2tag[i] for i in result]
print(result)
# 结果处理
entities_result = format_result(list(text), result)
print(json.dumps(entities_result, indent=4, ensure_ascii=False))