def predict():
predict_base_path = "../data/review_relation/predict/content_relation.csv"
food_reviews = []
liren_reviews = []
jiudian_reviews = []
yule_reviews = []
gouwu_reviews = []
with tf.gfile.GFile(predict_base_path, 'r') as reader:
for line in reader:
fields = line.strip().split('\t')
if fields[2] == "美食":
food_reviews.append(line)
elif fields[2] == "丽人":
liren_reviews.append(line)
elif fields[2] == "酒店":
jiudian_reviews.append(line)
elif fields[2] == "休娱":
yule_reviews.append(line)
else:
gouwu_reviews.append(line)
words_path = "../data/review_relation/bert_words.csv"
word_index = load_vocab_ids(words_path, sep='\t')
model_base_path = "../data/review_relation/version2/"
predict_result_base_path = "../data/review_relation/predict_result/"
model_predict(
os.path.join(model_base_path, 'food2.h5'), food_reviews,
os.path.join(predict_result_base_path, 'food_predict_result2.csv'),
word_index)
model_predict(
os.path.join(model_base_path, 'jiudian.h5'), jiudian_reviews,
os.path.join(predict_result_base_path, 'jiudian_predict_result2.csv'),
word_index)
model_predict(
os.path.join(model_base_path, 'liren.h5'), liren_reviews,
os.path.join(predict_result_base_path, 'liren_predict_result2.csv'),
word_index)
model_predict(
os.path.join(model_base_path, 'yule.h5'), yule_reviews,
os.path.join(predict_result_base_path, 'yule_predict_result2.csv'),
word_index)
model_predict(
os.path.join(model_base_path, 'gouwu.h5'), gouwu_reviews,
os.path.join(predict_result_base_path, 'gouwu_predict_result2.csv'),
word_index)
def predict():
word_index = load_vocab_ids(FLAGS.word_path)
data = pd.read_csv(FLAGS.predict_path, header=None)
# shopid = data.values[:, 0]
# textid = data.values[:, 1]
texts = data.values[:, 0]
text_id = [text_to_sequence(text.decode('utf-8'), word_index) for text in texts]
text_id = tf.keras.preprocessing.sequence.pad_sequences(text_id, value=word_index['pad'], padding='post',
maxlen=15)
print text_id[:10]
label = np.zeros((len(texts), 2)).astype(np.int64)
# a = serialize_example(text_id[0], label[0])
# print tf.train.Example.FromString(a)
# iterator = dataset.make_one_shot_iterator()
# next = iterator.get_next()
# with tf.Session() as sess:
# while True:
# try:
# x, y = sess.run(next)
# print
# except tf.errors.OutOfRangeError:
# break
#
config = BaseConfig.from_json_file(FLAGS.model_config_path).to_dict()
model = BilstmModel(config)
model = model.create_model()
model.compile(optimizer=tf.train.AdamOptimizer(learning_rate=0.001), loss='binary_crossentropy',
metrics=['accuracy'])
model = tf.keras.estimator.model_to_estimator(keras_model=model)
# tf.estimator.Estimator.predict()
inference = model.predict(
input_fn=lambda: predict_input_fn(text_id, label),
checkpoint_path=FLAGS.model_path)
result = [0 if ele['dense'][0] > 0.5 else 1 for ele in inference]
with tf.gfile.GFile("/tmp/1.csv", 'w') as writer:
for i in xrange(len(result)):
writer.write("%s\t%d\n" % (texts[i], result[i]))
def predict(text):
words_path = '../data/review_relation/bert_words.csv'
words = load_vocab_ids(words_path, sep='\t')
wordIds = text_to_ids(text, words)
signature_key = "xiaoxiang"
sess_config = tf.ConfigProto()
sess_config.gpu_options.allow_growth = True
with tf.Session(graph=tf.Graph(), config=sess_config) as sess:
meta_graph_def = tf.saved_model.loader.load(
sess, [tf.saved_model.tag_constants.SERVING], "/tmp/saved_model/1")
signature = meta_graph_def.signature_def
text = signature[signature_key].inputs["text"].name
outputs = signature[signature_key].outputs["label"].name
text = sess.graph.get_tensor_by_name(text)
outputs = sess.graph.get_tensor_by_name(outputs)
print(sess.run([outputs], feed_dict={text: [wordIds]}))
def predict2():
predict_base_path = "../data/review_relation/predict/content_relation.csv"
reviews = []
contentids = []
reviewlist = []
with tf.gfile.GFile(predict_base_path, 'r') as reader:
for line in reader:
fields = line.strip().split("\t")
if len(fields) != 2:
continue
contentids.append(fields[0])
reviewlist.append(fields[1])
reviews.append(line.strip())
words_path = "../data/review_relation/bert_words.csv"
word_index = load_vocab_ids(words_path, sep='\t')
model_base_path = "../data/review_relation/version2/"
predict_result_base_path = "../data/review_relation/content_pic_relation/content_relation_res.csv"
food_label = model_predict2(os.path.join(model_base_path, 'food2.h5'),
reviews, word_index)
jiudian_label = model_predict2(os.path.join(model_base_path, 'jiudian.h5'),
reviews, word_index)
liren_label = model_predict2(os.path.join(model_base_path, 'liren.h5'),
reviews, word_index)
yule_label = model_predict2(os.path.join(model_base_path, 'yule.h5'),
reviews, word_index)
gouwu_label = model_predict2(os.path.join(model_base_path, 'gouwu.h5'),
reviews, word_index)
with tf.gfile.GFile(predict_result_base_path, 'w') as writer:
writer.write(
"contentid\tfood\tjiudian\tliren\tyule\tgouwu\tcontentbody\n")
for i in xrange(len(contentids)):
writer.write("%s\t%s\t%s\t%s\t%s\t%s\t%s\n" %
(contentids[i], str(food_label[i]),
str(jiudian_label[i]), str(liren_label[i]),
str(yule_label[i]), str(gouwu_label[i]), reviews[i]))
def input_fn(path, vocab, config, batch_size):
dataset = tf.data.TextLineDataset(path)
dataset = dataset.map(lambda line: parse_line(line, vocab, config),
num_parallel_calls=4)
return dataset.batch(batch_size=batch_size)
if __name__ == "__main__":
config = BaseConfig.from_json_file(
"../data/dish/dish_similarity/config.json").to_dict()
path = "/tmp/valid.csv"
word_path = "../data/dish/dish_similarity/words.csv"
word_index = load_vocab_ids(word_path)
signature_key = 'predict_label'
sess_config = tf.ConfigProto()
sess_config.gpu_options.allow_growth = True
with tf.Session(graph=tf.Graph(), config=sess_config) as sess:
dataset = input_fn(path, word_index, config, 10)
iterator = dataset.make_one_shot_iterator()
next_element = iterator.get_next()
meta_graph_def = tf.saved_model.loader.load(
sess, [tf.saved_model.tag_constants.SERVING],
"../data/dish/dish_similarity/saved_model/4")
signature = meta_graph_def.signature_def
text_a = signature[signature_key].inputs["text_a"].name
text_b = signature[signature_key].inputs["text_b"].name
padding='pre',
maxlen=max_len)
tag_list = tf.keras.preprocessing.sequence.pad_sequences(
tag_list,
value=word_index.get('[PAD]'.encode('utf-8')),
padding='pre',
maxlen=max_len)
return feature_list, tag_list
if __name__ == '__main__':
# #处理训练数据
words_path = "/Users/lionel/Desktop/data/review_relation/bert_words.csv"
word_index = load_vocab_ids(words_path, sep='\t')
path = '/Users/lionel/Downloads/review_dish.csv'
config = {
'batch_size': 100,
'max_length': 100,
'vocab_size': 21128,
'embedding_size': 100,
'units': 100,
'num_tags': 3
}
feature_list, tag_list = process_data(path, word_index,
config['max_length'])
# index = feature_list.shape[0] // config['batch_size'] * config['batch_size']
#
def train():
train_path = '../train.csv'
valid_path = '../valid.csv'
words_path = '../data/review_relation/bert_words.csv'
words = load_vocab_ids(words_path, sep='\t')
train_text = []
train_label = []
with tf.gfile.GFile(train_path, 'r') as reader:
for line in reader:
fields = line.strip().split('\t')
if len(fields) != 2:
continue
train_text.append(fields[0])
train_label.append(int(fields[1]))
valid_text = []
valid_label = []
with tf.gfile.GFile(valid_path, 'r') as reader:
for line in reader:
fields = line.strip().split('\t')
if len(fields) != 2:
continue
valid_text.append(fields[0])
valid_label.append(int(fields[1]))
train_data = [text_to_ids(ele, words) for ele in train_text]
train_label = tf.keras.utils.to_categorical(train_label, 2)
valid_data = [text_to_ids(ele, words) for ele in valid_text]
valid_label = tf.keras.utils.to_categorical(valid_label, 2)
import numpy as np
train_data = np.array(train_data)
train_label = np.array(train_label)
valid_data = np.array(valid_data)
valid_label = np.array(valid_label)
print train_data.shape
config = {
'max_sequence_length': 50,
'vocab_size': 25000,
'embedding_size': 200,
'hidden_size': 100,
'drop_out': 0.2,
'num_classes': 2,
'epoch': 10,
'batch_size': 100,
'model_path': '/tmp/100'
}
model = LstmModel(config).create_model()
model.summary()
model.compile(loss=tf.keras.losses.categorical_crossentropy,
optimizer=tf.keras.optimizers.Adam(lr=0.01),
metrics=['accuracy'])
model.fit(x=train_data,
y=train_label,
epochs=config['epoch'],
batch_size=config['batch_size'],
validation_data=(valid_data, valid_label),
verbose=1)
model.save(config['model_path'])
def train(config):
"""
训练模型
:param config: 超参数配置文件
:return: 预测结果
"""
word_index = load_vocab_ids(os.path.join(FLAGS.data_dir, 'words.csv'))
if not tf.gfile.Exists(os.path.join(FLAGS.data_dir, 'model')):
tf.gfile.MkDir(tf.gfile.Exists(os.path.join(FLAGS.data_dir, 'model')))
model_path = os.path.join(FLAGS.data_dir, 'model/%s' % FLAGS.model_name)
train_examples = None
valid_examples = None
test_examples = None
if FLAGS.do_train and FLAGS.do_valid:
if FLAGS.input_format == 1:
train_examples = OneInputDataProcessor().get_train_examples(
FLAGS.data_dir)
valid_examples = OneInputDataProcessor().get_valid_examples(
FLAGS.data_dir)
if FLAGS.input_format == 2:
train_examples = TwoInputDataProcessor().get_train_examples(
FLAGS.data_dir)
valid_examples = TwoInputDataProcessor().get_valid_examples(
FLAGS.data_dir)
train_text_a, train_text_b, train_label_ids = features_labels_digitalize(
train_examples, word_index, config['max_sequence_length'])
valid_text_a, valid_text_b, valid_label_ids = features_labels_digitalize(
valid_examples, word_index, config['max_sequence_length'])
model = None
if FLAGS.input_format == 2:
model = text_similarity_model.BilstmModel(config,
merge_mode='multiply')
if FLAGS.input_format == 1:
model = text_classification_model.BilstmModel(config)
model = model.create_model()
model.compile(
optimizer=tf.keras.optimizers.Adam(lr=config['learning_rate']),
loss='binary_crossentropy',
metrics=['accuracy'])
start = time.time()
if train_text_b is not None:
model.fit(x=[train_text_a, train_text_b],
y=train_label_ids,
epochs=config['epoch'],
batch_size=config['batch_size'],
validation_data=([valid_text_a,
valid_text_b], valid_label_ids),
callbacks=[keras.callbacks.EarlyStopping(patience=2)])
else:
model.fit(x=train_text_a,
y=train_label_ids,
epochs=config['epoch'],
batch_size=config['batch_size'],
validation_data=(valid_text_a, valid_label_ids),
callbacks=[keras.callbacks.EarlyStopping(patience=2)])
end = time.time()
tf.logging.info("Train time is %ds", end - start)
model.save(model_path, overwrite=True)
if FLAGS.do_test:
if FLAGS.input_format == 1:
test_examples = OneInputDataProcessor().get_test_examples(
FLAGS.data_dir)
if FLAGS.input_format == 2:
test_examples = TwoInputDataProcessor().get_test_examples(
FLAGS.data_dir)
test_text_a, test_text_b, test_label_ids = features_labels_digitalize(
test_examples, word_index, config['max_sequence_length'])
model = keras.models.load_model(model_path)
if FLAGS.do_statistic:
if test_text_b is not None:
result = model.predict([test_text_a, test_text_b])
else:
result = model.predict(test_text_a)
print metrics.classification_report(
np.argmax(test_label_ids, axis=1), np.argmax(result,
axis=1))
else:
count = len(test_text_a)
n = count // config['batch_size'] + 1
with tf.gfile.GFile(
os.path.join(FLAGS.data_dir, 'predict.csv'), 'w') as f:
for i in xrange(n):
if (i + 1) * config['batch_size'] >= count:
x_a = test_text_a[i * config['batch_size']:]
x_b = test_text_b[i * config['batch_size']:]
else:
x_a = test_text_a[i *
config['batch_size']:(i + 1) *
config['batch_size']]
x_b = test_text_b[i *
config['batch_size']:(i + 1) *
config['batch_size']]
predictions = model.predict_on_batch([x_a, x_b])
result = np.argmax(predictions, axis=1)
for j in xrange(len(result)):
index = i * config['batch_size'] + j
f.write('%s\t%s\t%s\n' %
(test_examples[index].text_a,
test_examples[index].text_b, result[j]))
if FLAGS.do_export:
model = keras.models.load_model(model_path)
features = dict()
x = model.input
y = model.output
args = FLAGS.params
if isinstance(x, list):
for i in range(len(x)):
features[args[i]] = x[i]
else:
features[args[0]] = x
labels = dict()
if isinstance(y, list):
for i in range(len(y)):
labels[args[len(features)]] = y[i]
else:
labels[args[len(features)]] = y
sess = tf.keras.backend.get_session()
prediction_signature = tf.saved_model.signature_def_utils.predict_signature_def(
inputs=features, outputs=labels)
valid_prediction_signature = tf.saved_model.signature_def_utils.is_valid_signature(
prediction_signature)
if not valid_prediction_signature:
raise ValueError("Error: Prediction signature not valid!")
saved_model_path = os.path.join(
FLAGS.data_dir, 'saved_model/%d' % (FLAGS.saved_model_version))
if tf.gfile.Exists(saved_model_path):
tf.gfile.DeleteRecursively(saved_model_path)
builder = tf.saved_model.builder.SavedModelBuilder(
saved_model_path)
legacy_init_op = tf.group(tf.tables_initializer(),
name='legacy_init_op')
builder.add_meta_graph_and_variables(
sess, [tf.saved_model.tag_constants.SERVING],
signature_def_map={FLAGS.signature_name: prediction_signature},
legacy_init_op=legacy_init_op)
builder.save()
