def run(self):
time_all = []
bc = BertClient(port=PORT, show_server_config=False)
for _ in range(self.num_repeat):
start_t = time.perf_counter()
bc.encode(self.batch)
time_all.append(time.perf_counter() - start_t)
print(time_all)
self.avg_time = mean(time_all)
class BenchmarkClient(threading.Thread):
def __init__(self, args):
super().__init__()
self.batch = [
''.join(
random.choices(string.ascii_uppercase + string.digits,
k=args.client_seq_len))
for _ in args.client_batch_size
]
self.bc = BertClient()
self.num_repeat = args.num_repeat
self.avg_time = 0
def run(self):
time_all = []
for _ in range(self.num_repeat):
start_t = time.clock()
self.bc.encode(self.batch)
time_all.append(time.clock() - start_t)
self.avg_time = mean(time_all)
def test():
print("Loading test data...")
start_time = time.time()
x_test = process_file_nolabel(test_dir, word_to_id, config.seq_length)
bc = BertClient()
x_test = bc.encode(x_test)
# (test 2364, 80)
session = tf.Session()
session.run(tf.global_variables_initializer())
saver = tf.train.Saver()
saver.restore(sess=session, save_path=save_path) # 读取保存的模型
# print('Testing...')
# loss_test, acc_test = evaluate(session, x_test, y_test)
# msg = 'Test Loss: {0:>6.2}, Test Acc: {1:>7.2%}'
# print(msg.format(loss_test, acc_test))
batch_size = 32
data_len = len(x_test)
num_batch = int((data_len - 1) / batch_size) + 1
# y_test_cls = np.argmax(y_test, 1)
y_pred_cls = np.zeros(shape=len(x_test), dtype=np.int32) # 保存预测结果
for i in range(num_batch): # 逐批次处理
start_id = i * batch_size
end_id = min((i + 1) * batch_size, data_len)
feed_dict = {
model.input_x: x_test[start_id:end_id],
model.keep_prob: 1.0
}
y_pred_cls[start_id:end_id] = session.run(model.y_pred_cls, feed_dict=feed_dict)
# 评估
print("Precision, Recall and F1-Score...")
# print(metrics.classification_report(y_test_cls, y_pred_cls, target_names=categories))
# 混淆矩阵
print("Confusion Matrix...")
# cm = metrics.confusion_matrix(y_test_cls, y_pred_cls)
# print(cm)
time_dif = get_time_dif(start_time)
print("Time usage:", time_dif)
return y_pred_cls
def main():
# 汽车观点提取,分过词的数据
# df_train = pd.read_csv('./data/ccf_car_train.csv')
# df_test = pd.read_csv('./data/ccf_car_test.csv')
# train = df_train['word_seg']
# test_x = df_test['content']
# target = df_train['class']
# train = np.array(train)
# target = np.array(target)
# list = []
# for i in target:
# list.append(i)
bc = BertClient()
train = np.load('./data/train_x.npy')
train = list(train)
train = bc.encode(train)
target = np.load('./data/train_y.npy')
test_x = np.load('./data/test_x.npy')
test_y = np.load('./data/test_y.npy')
test_x = list(test_x)
test_x = bc.encode(test_x)
# 分为训练集和测试的时候加上
# train, test_x, target, test_y = train_test_split(train, target, test_size = 0.15, random_state = 0)
# np.savetxt('./data/train_x', train_X, fmt='%s')
# np.savetxt('./data/train_y', train_y, fmt='%s')
# np.savetxt('./data/test_X', test_X, fmt='%s')
# np.savetxt('./data/test_y', test_y, fmt='%s')
# df_test = pd.read_csv(test_file)
# df_train = df_train.drop(['article'], axis=1)
# df_test = df_test.drop(['article'], axis=1)
# ngram_range：tuple (min_n, max_n) 要提取的不同n-gram的n值范围的下边界和上边界。 将使用n的所有值，使得min_n <= n <= max_n。
# max_features: 构建一个词汇表，该词汇表仅考虑语料库中按术语频率排序的最高max_features
# if feature == 'Count':
# vectoriser = CountVectorizer(ngram_range=(1, 2), min_df = 3)
# elif feature == 'Tfidf':
# vectoriser = TfidfVectorizer(ngram_range=(1, 5), min_df = 3, max_df = 0.7)
# # 构建特征，先训练
# vectoriser.fit(train)
# 训练完进行归一化 总共有315503个词，过滤小于3的，剩下23546, max_df 貌似没用
# (7957, 2082), (1990, 2082) type:crs_matrix
# train_X = vectoriser.transform(train)
# test_X = vectoriser.transform(test_x)
# y_train = df_train['class'] - 1
# train_X = np.array(train_X.data).reshape(train_X.shape)
# 开始构建分类器
if classfier == 'LR':
'''
c：正则化系数λ的倒数，float类型，默认为1.0。必须是正浮点型数。像SVM一样，越小的数值表示越强的正则化
'''
rg = LogisticRegression(C=1)
rg.fit(train, target)
y_pred = rg.predict(test_x)
# elif classfier == 'NB':
# # 使用默认的配置对分类器进行初始化。
# mnb_count = MultinomialNB(alpha=0.2)
# # 使用朴素贝叶斯分类器，对CountVectorizer（不去除停用词）后的训练样本进行参数学习。
# mnb_count.fit(train_X, target)
# y_pred = mnb_count.predict(test_X)
#
# elif classfier =='tree':
# DT = tree.DecisionTreeClassifier()
# DT.fit(train_X, target)
# y_pred = DT.predict(test_X)
'''
kernel='linear'时，为线性核，C越大分类效果越好，但有可能会过拟合（defaul C=1）。
kernel='rbf'时（default），为高斯核，gamma值越小，分类界面越连续；gamma值越大，分类界面越“散”，分类效果越好，但有可能会过拟合。
'''
# C=0.1 准确率高召回率低 C = 0.8
# elif classfier =='RT':
# sv = RandomForestClassifier(n_estimators=400)
# sv.fit(train_X, target)
# # y_hat = sv.predict(train_X)
# y_pred = sv.predict(test_X)
# scores = cross_val_score(sv, train_X, target, cv=5)
# print(scores)
# 从sklearn.metrics 导入 classification_report。
# 输出更加详细的其他评价分类性能的指标。
print('classifier is : ' + classfier + '\tFeature is : ' + feature)
print(classification_report(test_y, y_pred))
# using BertClient in multicast way
import sys
import threading
from service.client import BertClient
def client_clone(id, idx):
bc = BertClient(port=int(sys.argv[1]),
port_out=int(sys.argv[2]),
identity=id)
for j in bc.listen():
print('clone-client-%d: received %d x %d' %
(idx, j.shape[0], j.shape[1]))
if __name__ == '__main__':
bc = BertClient(port=int(sys.argv[1]), port_out=int(sys.argv[2]))
# start two cloned clients sharing the same identity as bc
for j in range(2):
t = threading.Thread(target=client_clone, args=(bc.identity, j))
t.start()
with open('README.md') as fp:
data = [v for v in fp if v.strip()]
for _ in range(3):
vec = bc.encode(data)
print('bc received %d x %d' % (vec.shape[0], vec.shape[1]))
import sys
import time
from service.client import BertClient
if __name__ == '__main__':
bc = BertClient(ip='localhost', port=int(sys.argv[1]))
# encode a list of strings
with open('README.md', encoding='utf8') as fp:
data = [v for v in fp if v.strip()]
for j in range(1, 200, 10):
start_t = time.time()
tmp = data * j
bc.encode(tmp)
time_t = time.time() - start_t
print('encoding %d strs in %.2fs, speed: %d/s' %
(len(tmp), time_t, int(len(tmp) / time_t)))
# bad example: encode a string
# print(ec.encode('abc'))
from service.client import BertClient
bc = BertClient('ip = localhost')
vec = bc.encode(['hello world', 'good day'])
def RL_tuning_model(config):
emd_file = config.nemd_dir, config.cemd_dir
train_data, test_data, eval_data, vocab, embd = load_data(config.data_dir, emd_file, "BA")
##train_Id = [SOS] + clean_Id, ground_truth = clean_Id + [EOS], clean_len = get_length(ground_truth)
train_noisy_Id, train_noisy_len, train_char_noisy_Id, train_char_noisy_len, train_nemd, train_target_Id, train_input_Id, train_clean_Id, train_clean_len, train_answer_Id, train_answer_len, max_char, max_word = train_data
# noisy_Id, noisy_len, char_noisy_Id, char_noisy_len, ground_truth, train_Id, clean_Id, clean_len, answer_Id, answer_len, max_char, max_word = train_data
test_noisy_Id, test_noisy_len, test_char_noisy_Id, test_char_noisy_len,test_nemd, test_ground_truth, test_train_Id, test_clean_Id, test_clean_len, test_answer_Id, test_answer_len = test_data
# eval_noisy_Id, eval_noisy_len, eval_char_noisy_Id, eval_char_noisy_len,eval_target_Id, eval_input_Id, eval_clean_Id, eval_clean_len, eval_answer_Id, eval_answer_len =eval_data
# config.max_target_length = max_word
L = max_word
embd = np.array(embd)
batch_size =config.batch_size
discount_factor = config.discount_factor
sen_reward_rate = config.sen_reward_rate
lam = config.lam
gamma = config.gamma
beam_width = config.beam_width
update_N = config.update_N
# config = tf.ConfigProto()
# config.gpu_options.allow_growth = True
model_name = "BS_"
values = {}
checkpoint_path = config.seq2seq_ckp_dir
BS_ckp_dir = config.BS_ckp_dir
reader = pywrap_tensorflow.NewCheckpointReader(checkpoint_path)
var_to_shape_map = reader.get_variable_to_shape_map()
for key in var_to_shape_map:
if 'loss_fun' not in key:
values[model_name + key + ':0'] = reader.get_tensor(key)
vocab_size =len(vocab)
G_BeamSearch = tf.Graph()
with G_BeamSearch.as_default():
BeamSearch_seq2seq = Network.BeamSearch_Seq2seq(config, embd, vocab_size, model_name)
seq2seq_len = L
with tf.Session(graph=G_BeamSearch) as sess_beamsearch:
sess_beamsearch.run(tf.global_variables_initializer())
for v in tf.trainable_variables():
if v.name in values.keys():
v.load(values[v.name], sess_beamsearch)
for epo in range(config.epoch):
print(" epoch: {}".format(epo))
idx = np.arange(0, len(train_noisy_Id))
idx = list(np.random.permutation(idx))
Bleu_score1 = []
Bleu_score2 = []
Bleu_score3 = []
Bleu_score4 = []
Rouge_score = []
Meteor_score = []
Cider_score = []
for batch in range(len(train_noisy_Id) / config.batch_size):
source_shuffle, source_len, source_nemd, char_Id, char_len, train_shuffle, target_shuffle, target_len, answer_shuffle, answer_len = next_batch(
train_noisy_Id,
train_noisy_len,
train_nemd,
train_char_noisy_Id,
train_char_noisy_len,
train_input_Id,
train_target_Id,
train_clean_len,
train_answer_Id,
train_answer_len,
batch_size,
batch,
idx)
source_shuffle = tf.keras.preprocessing.sequence.pad_sequences(source_shuffle, maxlen=max_word,
padding='post', truncating="post",
value=EOS)
source_emd = []
for n in source_nemd:
source_emd.append(n[0:source_shuffle.shape[1]])
source_emd = np.asarray(source_emd)
target_shuffle_in = tf.keras.preprocessing.sequence.pad_sequences(target_shuffle, maxlen=seq2seq_len,
padding='post', truncating="post",
value=EOS)
# target_emd = []
# for c in train_cemd:
# target_emd.append(c[0:train_shuffle.shape[1]])
# target_emd = np.asarray(target_emd)
train_shuffle = tf.keras.preprocessing.sequence.pad_sequences(train_shuffle, maxlen=seq2seq_len,
padding='post', truncating="post",
value=EOS)
whole_noisy_char_Id = []
for sen_char in char_Id:
sen_len = len(sen_char)
for i in range(len(source_shuffle[0]) - sen_len):
sen_char.append([0] * max_char) ## fix the char with the length of max_word
whole_noisy_char_Id.append(sen_char)
whole_noisy_char_Id = np.asarray(whole_noisy_char_Id)
whole_char_len = []
for char_List in char_len:
whole_char_len.append(char_List[:len(source_shuffle[0])])
# initial_input_in = [target_shuffle_in[i][-1] for i in range(config.batch_size)]
target_len = np.tile(seq2seq_len, config.batch_size)
# print "the shape of source_shuffle ", np.array(source_shuffle).shape
# print "the shape of source_len ", np.array(source_len).shape
# print "the shape of source_emd ", np.array(source_emd).shape
# print "the shape of whole_noisy_char_Id ", np.array(whole_noisy_char_Id).shape
# print "the shape of whole_char_len ", np.array(whole_char_len).shape
# print "the shape of target_len ", np.array(target_len).shape
# print "the shape of train_shuffle ", np.array(train_shuffle).shape
# print "the shape of target_shuffle_in ", np.array(target_shuffle_in).shape
fd = {BeamSearch_seq2seq.encoder_inputs: source_shuffle,
BeamSearch_seq2seq.encoder_inputs_length: source_len,
BeamSearch_seq2seq.encoder_emb: source_emd,
BeamSearch_seq2seq.encoder_char_ids: whole_noisy_char_Id, BeamSearch_seq2seq.encoder_char_len: whole_char_len,
BeamSearch_seq2seq.decoder_length: target_len,
BeamSearch_seq2seq.decoder_inputs: train_shuffle,
# BeamSearch_seq2seq.target_emb: target_emd,
BeamSearch_seq2seq.decoder_targets: target_shuffle_in,
}
if batch <= 100:
ids, _, loss_seq2seq = sess_beamsearch.run(
[BeamSearch_seq2seq.seq_ids, BeamSearch_seq2seq.train_op,BeamSearch_seq2seq.loss_seq2seq], fd)
if batch % 100 == 0:
print ('Batch {} Seq2seq_loss: '.format(batch),loss_seq2seq)
# noisy_sen = []
# for pre_id in source_shuffle[0]:
# if pre_id != EOS and pre_id != PAD and pre_id != SOS:
# noisy_sen.append(vocab[pre_id])
#
# clean_sen = []
# for clean_id in target_shuffle[0]:
# if clean_id != EOS and clean_id != PAD and clean_id != SOS:
# clean_sen.append(vocab[clean_id])
#
# gene_sen = []
# for gen_id in ids[0]:
# if gen_id != EOS and gen_id != PAD and gen_id != SOS:
# gene_sen.append(vocab[gen_id])
#
# print("question" + str(1) + "\n");
# print("noisy question: " + " ".join(noisy_sen) + "\n")
# print("clean question: " + " ".join(clean_sen) + "\n")
# print("generated question: " + " ".join(gene_sen) + "\n")
elif batch > 100:
RL_logits, policy, values = sess_beamsearch.run([BeamSearch_seq2seq.predicting_logits, BeamSearch_seq2seq.predicting_scores, BeamSearch_seq2seq.value], fd)
# generated_que_input = np.insert(generated_que, 0, SOS, axis=1)[:, 0:-1]
# target_shuffle = tf.keras.preprocessing.sequence.pad_sequences(target_shuffle, maxlen=None,
# padding='post',truncating="post", value=EOS)
generated_target_len = RL_logits.shape[1]
bc = BertClient(port=5555, port_out=5556)
generated_sen_beam = []
for beam in range(beam_width):
for b in range(batch_size):
sen = []
for l in range(RL_logits.shape[1]):
sen.append(vocab[int(RL_logits[b][l][beam])])
generated_sen_beam.append(" ".join(sen))
bert_emd_beam = bc.encode(generated_sen_beam)
# bert_emd =
# bert_emd_beam.append(bert_emd[:,:generated_target_len,:])
bert_emd_beam = bert_emd_beam[:,1:generated_target_len+1,:]
bert_emd_beam = bert_emd_beam.reshape(batch_size, beam_width, bert_emd_beam.shape[1], bert_emd_beam.shape[2])
sen_len_batch_beam = generated_target_len * np.ones((batch_size,beam_width))
# print "the shape of bert emd: ", bert_emd_beam.shape
#
#
# source_shuffle_beam = np.tile(source_shuffle, (beam_width, 1))
# source_len_beam = np.tile(source_len, (beam_width))
# source_emd_beam = np.tile(source_emd, (beam_width,1,1))
# print "the shape of source_emd_beam", source_emd_beam.shape
# whole_noisy_char_Id_beam = np.tile(whole_noisy_char_Id,(beam_width,1,1))
# whole_char_len_beam = np.tile(whole_char_len,(beam_width,1))
# print "the char_len: ", whole_char_len_beam.shape
# RL_logits_beam = np.transpose(RL_logits,(0,2,1)).reshape(batch_size*beam_width, generated_target_len)
#
# answer_shuffle_beam = np.tile(answer_shuffle,(beam_width,1))
# answer_len_beam = np.tile(answer_len, (beam_width))
#
# print "the shape of source_shuffle: ", source_shuffle_beam.shape
# print "the input of target: ", RL_logits_beam.shape
#
# # for beam in range(beam_width):
# # sen_len_batch_beam = []
# # for b in range(batch_size):
# # sen_len_batch =[]
# # l=0
# # while (l < generated_target_len and RL_logits[b][l][beam] != 1):
# # l = l + 1
# # sen_len_batch.append(l)
# # sen_len_batch_beam.append(sen_len_batch)
#
# Character_data= source_shuffle_beam,source_len_beam,source_emd_beam, whole_noisy_char_Id_beam, whole_char_len_beam, RL_logits_beam, sen_len_batch_beam, generated_target_len, bert_emd_beam
# QA_data = answer_shuffle_beam, answer_len_beam, source_shuffle_beam, source_len_beam, source_emd_beam, RL_logits_beam, sen_len_batch_beam, bert_emd_beam, vocab
Character_data = source_shuffle, source_len, source_emd, whole_noisy_char_Id, whole_char_len, RL_logits, sen_len_batch_beam, generated_target_len, bert_emd_beam
QA_data = answer_shuffle, answer_len, source_shuffle, source_len, source_emd, RL_logits, sen_len_batch_beam, bert_emd_beam, vocab
RL_rewards = Reward(config, Character_data, QA_data, vocab, embd)
values_t = np.transpose(values,(2,0,1))
values_t1 = np.transpose(values[:,1:,:],(2,0,1))
values_t1 = np.insert(values_t1, values_t1.shape[2], values=np.zeros(batch_size), axis=2)
Returns_beam =[]
Deltas_beam = []
for beam in range(len(RL_rewards)):
Returns_batch =[]
Deltas_batch = []
for batch_len in range(len(RL_rewards[1])):
returns, deltas = get_training_data(RL_rewards[beam][batch_len], values_t[beam][batch_len], values_t1[beam][batch_len], lam, gamma)
Returns_batch.append(returns)
Deltas_batch.append(deltas)
Returns_beam.append(Returns_batch)
Deltas_beam.append(Deltas_batch)
Returns_beam = np.transpose(Returns_beam,(1,2,0))
Deltas_beam = np.transpose(Deltas_beam,(1,2,0))
old_policy = policy
for N in range(update_N):
fd = {BeamSearch_seq2seq.returns_ph: Returns_beam, BeamSearch_seq2seq.advantages_ph: Deltas_beam,
BeamSearch_seq2seq.old_log_probs_ph: old_policy,
BeamSearch_seq2seq.encoder_inputs: source_shuffle,
BeamSearch_seq2seq.encoder_inputs_length: source_len,
BeamSearch_seq2seq.encoder_emb: source_emd,
BeamSearch_seq2seq.encoder_char_ids: whole_noisy_char_Id,
BeamSearch_seq2seq.encoder_char_len: whole_char_len,
BeamSearch_seq2seq.decoder_length: target_len,
BeamSearch_seq2seq.decoder_inputs: train_shuffle,
BeamSearch_seq2seq.decoder_targets: target_shuffle_in}
policy_l, entr_l, loss_ppo, ratio, _, policy= sess_beamsearch.run(
[BeamSearch_seq2seq.policy_loss, BeamSearch_seq2seq.entropy,
BeamSearch_seq2seq.loss, BeamSearch_seq2seq.ratio, BeamSearch_seq2seq.optimize_expr, BeamSearch_seq2seq.predicting_scores], fd)
print ('Batch {} PPO_loss: '.format(batch),loss_ppo)
# print (" The ratio is :", ratio)
# for t in range(batch_size):
# ref = []
# hyp = []
# for tar_id in target_shuffle[t]:
# if tar_id != EOS and tar_id != PAD:
# ref.append(vocab[tar_id])
# # print vocab[tar_id]
# for pre_id in RL_logits[t]:
# if pre_id != EOS and pre_id != PAD:
# hyp.append(vocab[pre_id])
#
# hyp_sen = u" ".join(hyp).encode('utf-8')
# ref_sen = u" ".join(ref).encode('utf-8')
# dic_hyp = {}
# dic_hyp[0] = [hyp_sen]
# dic_ref = {}
# dic_ref[0] = [ref_sen]
# sen_bleu, _ = Bleu_obj.compute_score(dic_ref, dic_hyp)
# sen_rouge = Rouge_obj.compute_score(dic_ref, dic_hyp)
# sen_meteor, _ = Meteor_obj.compute_score(dic_ref, dic_hyp)
# sen_cider, _ = cIDER_obj.compute_score(dic_ref, dic_hyp)
# Bleu_score1.append(sen_bleu[0])
# Bleu_score2.append(sen_bleu[1])
# Bleu_score3.append(sen_bleu[2])
# Bleu_score4.append(sen_bleu[3])
# Rouge_score.append(sen_rouge[0])
# Meteor_score.append(sen_meteor)
# Cider_score.append(sen_cider)
#
# if batch == 0 or batch % 100 == 0:
# print(' batch {}'.format(batch))
# print(' minibatch reward of training: {}'.format(- loss_ppo))
# #print the result
# for t in range(5): ## five sentences
# print('Question {}'.format(t))
# print("noisy question:")
# print(" ".join(map(lambda i: vocab[i], list(source_shuffle[t]))).strip().replace("<EOS>"," "))
# print("clean question:")
# print(" ".join(map(lambda i: vocab[i], list(target_shuffle[t]))).strip().replace("<EOS>"," "))
# print("the generated first question:")
# pre_sen = []
# for log_id in RL_logits[t, :, 0]:
# if log_id != -1 and log_id!=2 and log_id!=0:
# pre_sen.append(log_id)
# print(" ".join(map(lambda i: vocab[i], pre_sen)).strip())
# print("\n")
# print("\n Bleu_score1, Bleu_score2, Bleu_score3, Bleu_score4, Rouge, Cider\n")
# print(" {:.4f}, {:.4f}, {:.4f}, {:.4f}, {:.4f}, {:.4f}, {:.4f}".format(
# sum(Bleu_score1) / float(len(Bleu_score1)),
# sum(Bleu_score2) / float(len(Bleu_score2)),
# sum(Bleu_score3) / float(len(Bleu_score3)),
# sum(Bleu_score4) / float(len(Bleu_score4)),
# sum(Rouge_score) / float(len(Rouge_score)),
# sum(Meteor_score) / float(len(Meteor_score)),
# sum(Cider_score) / float(len(Cider_score))
# ))
gc.collect()
if (batch % 100 == 0 and batch < 100) or (batch % 20 == 0 and batch >100):
generated_test_sen = []
test_answer_sen = []
test_noisy_sen = []
test_clean_sen = []
# saver.save(sess_beamsearch, BS_ckp_dir)
##testing set result:
for batch_test in range(len(test_noisy_Id) / batch_size):
idx_t = np.arange(0, len(test_noisy_Id))
idx_t = list(np.random.permutation(idx_t))
test_source_shuffle, test_source_len, test_source_nemd, test_char_Id, test_char_len_in, test_train_shuffle, test_target_shuffle, test_target_len, test_answer_shuffle, test_ans_len = next_batch(
test_noisy_Id, test_noisy_len, test_nemd, test_char_noisy_Id, test_char_noisy_len, test_train_Id, test_ground_truth, test_clean_len, test_answer_Id,
test_answer_len, batch_size, batch_test, idx_t)
for an in test_answer_shuffle:
test_answer_sen.append(vocab[anum] for anum in an)
for no in test_source_shuffle:
test_noisy_sen.append(vocab[si] for si in no)
for cl in test_target_shuffle:
test_clean_sen.append(vocab[ci] for ci in cl)
test_source_shuffle_in = tf.keras.preprocessing.sequence.pad_sequences(test_source_shuffle, maxlen = max_word,
padding='post', truncating="post",value=EOS)
test_source_emd = []
for n in test_source_nemd:
test_source_emd.append(n[0 : test_source_shuffle_in.shape[1]])
test_source_emd = np.asarray(test_source_emd)
val_target_shuffle_in=[]
for val in test_target_shuffle:
val_target_shuffle_in.append([val[0]])
val_train_shuffle_in = []
for tra in test_train_shuffle:
val_train_shuffle_in.append([tra[0]])
initial_input_in = [SOS for i in range(batch_size)]
# [SOS for i in range(batch_size)]
val_target_len_in = np.tile(1, batch_size)
test_batch_noisy_char_Id = []
for sen_char in test_char_Id:
sen_len = len(sen_char)
for i in range(len(test_source_shuffle_in[0]) - sen_len):
sen_char.append([0] * max_char) ## fix the char with the length of max_word
test_batch_noisy_char_Id.append(sen_char)
test_batch_noisy_char_Id = np.asarray(test_batch_noisy_char_Id)
test_char_len = []
for char_List in test_char_len_in:
test_char_len.append(char_List[:len(test_source_shuffle_in[0])])
# print "the shape of test_noisy_char_Id ", np.array(test_batch_noisy_char_Id).shape
# if np.array(test_batch_noisy_char_Id).shape == (batch_size,):
# continue
# print "the shape of test_char_len ", np.array(test_char_len).shape
# # print "the shape of source_shuffle ", np.array(test_source_shuffle_in).shape
# # print "the shape of source_len ", np.array(test_source_len).shape
# # print "the shape of source_emd ", np.array(test_source_emd).shape
# # print "the shape of whole_noisy_char_Id ", np.array(test_batch_noisy_char_Id).shape
# # print "the shape of whole_char_len ", np.array(test_char_len).shape
fd = {BeamSearch_seq2seq.encoder_inputs: test_source_shuffle_in,
BeamSearch_seq2seq.encoder_inputs_length: test_source_len,
BeamSearch_seq2seq.encoder_emb: test_source_emd,
BeamSearch_seq2seq.encoder_char_ids: test_batch_noisy_char_Id,
BeamSearch_seq2seq.encoder_char_len: test_char_len}
val_id = sess_beamsearch.run([BeamSearch_seq2seq.ids], fd)
for t in range(batch_size):
ref = []
hyp = []
for tar_id in test_target_shuffle[t]:
if tar_id != EOS and tar_id != PAD and tar_id != SOS:
ref.append(vocab[tar_id])
# print vocab[tar_id]
for pre_id in val_id[0][t]:
if pre_id != EOS and pre_id != PAD and pre_id != SOS:
hyp.append(vocab[pre_id])
generated_test_sen.append(hyp)
hyp_sen = u" ".join(hyp).encode('utf-8')
ref_sen = u" ".join(ref).encode('utf-8')
dic_hyp = {}
dic_hyp[0] = [hyp_sen]
dic_ref = {}
dic_ref[0] = [ref_sen]
sen_bleu, _ = Bleu_obj.compute_score(dic_ref, dic_hyp)
sen_rouge = Rouge_obj.compute_score(dic_ref, dic_hyp)
sen_meteor, _ = Meteor_obj.compute_score(dic_ref, dic_hyp)
sen_cider, _ = cIDER_obj.compute_score(dic_ref, dic_hyp)
Bleu_score1.append(sen_bleu[0])
Bleu_score2.append(sen_bleu[1])
Bleu_score3.append(sen_bleu[2])
Bleu_score4.append(sen_bleu[3])
Rouge_score.append(sen_rouge[0])
Meteor_score.append(sen_meteor)
Cider_score.append(sen_cider)
print("\n Bleu_score1, Bleu_score2, Bleu_score3, Bleu_score4, Rouge, Cider\n")
print(" {:.4f}, {:.4f}, {:.4f}, {:.4f}, {:.4f}, {:.4f}, {:.4f}".format(
sum(Bleu_score1) / float(len(Bleu_score1)),
sum(Bleu_score2) / float(len(Bleu_score2)),
sum(Bleu_score3) / float(len(Bleu_score3)),
sum(Bleu_score4) / float(len(Bleu_score4)),
sum(Rouge_score) / float(len(Rouge_score)),
sum(Meteor_score) / float(len(Meteor_score)),
sum(Cider_score) / float(len(Cider_score))
))
bleu_score = sum(Bleu_score1) / float(len(Bleu_score1))
for i in range(len(generated_test_sen[0:2])):
print("question"+str(i) + "\n")
print("answer: "+" ".join(test_answer_sen[i]) + "\n")
print("noisy question: " + " ".join(test_noisy_sen[i]) + "\n")
print("clean question: " + " ".join(test_clean_sen[i]) + "\n")
print("generated question: " + " ".join(generated_test_sen[i]) + "\n")
fname = BS_ckp_dir +str(bleu_score)
f=open(fname, "wb")
f.write("\n Bleu_score1, Bleu_score2, Bleu_score3, Bleu_score4, Rouge, Meteor Cider\n")
f.write(" {:.4f}, {:.4f}, {:.4f}, {:.4f}, {:.4f}, {:.4f}, {:.4f} \n".format(
sum(Bleu_score1) / float(len(Bleu_score1)),
sum(Bleu_score2) / float(len(Bleu_score2)),
sum(Bleu_score3) / float(len(Bleu_score3)),
sum(Bleu_score4) / float(len(Bleu_score4)),
sum(Rouge_score) / float(len(Rouge_score)),
sum(Meteor_score) / float(len(Meteor_score)),
sum(Cider_score) / float(len(Cider_score))
))
for i in range(len(generated_test_sen)):
f.write("question"+str(i) + "\n")
f.write("answer: "+" ".join(test_answer_sen[i]) + "\n")
f.write("noisy question: " + " ".join(test_noisy_sen[i]) + "\n")
f.write("clean question: " + " ".join(test_clean_sen[i]) + "\n")
f.write("generated question: " + " ".join(generated_test_sen[i]) + "\n")
network_vars = tf.get_collection(
tf.GraphKeys.TRAINABLE_VARIABLES, config.PPoscope)
saver = tf.train.Saver(sharded=False)
saver.save(sess_beamsearch, config.BS_ckp_dir)
b = nn.Linear(76, 76)
w = list(b.parameters())
print(b(a) == a @ w[0].transpose(-2, -1) + w[1])
# %% Cosine distance for Various Vectors.
from scipy import spatial
from service.client import BertClient
bc = BertClient(ip='xxx', port=3004, port_out=3005)
sentences = [
'나 는 너 를 사랑 하다 여',
'나 는 너 를 사랑 하다 였 다',
'사랑 누 가 말하다 였 나',
]
(q_length, q_tokens, q_embedding, q_ids) = bc.encode(sentences)
love_1 = q_embedding[0][5]
love_2 = q_embedding[1][5]
love_3 = q_embedding[2][1]
print(spatial.distance.cdist([love_2, love_3], [love_1], metric='cosine'))
print(spatial.distance.cdist([love_2 + love_3], [love_1], metric='cosine'))
print(spatial.distance.cdist([love_2 * 10 + love_3], [love_1],
metric='cosine'))
print(spatial.distance.cdist([love_2 + love_3 * 10], [love_1],
metric='cosine'))
print(
spatial.distance.cdist([(love_2 + love_3) * 10], [love_1],
metric='cosine'))
import os
import GPUtil
import tensorflow as tf
from service.client import BertClient
os.environ['CUDA_VISIBLE_DEVICES'] = str(GPUtil.getFirstAvailable()[0])
tf.logging.set_verbosity(tf.logging.INFO)
with open('README.md') as fp:
data = [v for v in fp if v.strip()]
bc = BertClient()
list_vec = bc.encode(data)
list_label = [0 for _ in data] # a dummy list of all-zero labels
# write tfrecords
with tf.python_io.TFRecordWriter('tmp.tfrecord') as writer:
def create_float_feature(values):
return tf.train.Feature(float_list=tf.train.FloatList(value=values))
def create_int_feature(values):
return tf.train.Feature(int64_list=tf.train.Int64List(
value=list(values)))
for (vec, label) in zip(list_vec, list_label):
features = {
'features': create_float_feature(vec),
'labels': create_int_feature([label])
def Reward(config, Character_data, QA_data, vocab, embd):
batch_size = config.batch_size
vocab_size = len(vocab)
num_units = config.num_units
beam_width = config.beam_width
Bidirection = config.Bidirection
Embd_train = config.Embd_train
Attention = config.Attention
discount_factor = config.discount_factor
dropout = config.dropout
char_num = config.char_num
char_dim = config.char_dim
char_hidden_units = config.char_hidden_units
Seq2Seq_pretain_ckp = config.seq2seq_ckp_dir
QA_pretain_ckp = config.QA_ckp_dir
model_cond = "training"
G_Seq2Seq = tf.Graph()
sess_word_rw = tf.Session(graph=G_Seq2Seq)
with G_Seq2Seq.as_default():
Seq2Seq_model = CharacterBA.Char_Seq2Seq(batch_size, vocab_size,
num_units, embd, model_cond,
Bidirection, Embd_train,
char_hidden_units, Attention,
char_num, char_dim)
saver_word_rw = tf.train.Saver()
saver_word_rw.restore(sess_word_rw, Seq2Seq_pretain_ckp)
model_type = "testing"
G_QA_similiarity = tf.Graph()
sess_QA_rw = tf.Session(graph=G_QA_similiarity)
with G_QA_similiarity.as_default():
QA_simi_model = QA_similiarity.QA_similiarity(batch_size, num_units,
embd, model_type)
saver_sen_rw = tf.train.Saver()
saver_sen_rw.restore(sess_QA_rw, QA_pretain_ckp)
source_shuffle, source_len, source_emd, whole_noisy_char_Id, char_len, generated_sen, sen_len_batch_beam, generated_target_len, bert_emd = Character_data
logits_reward_beam = []
# print "the generated_sen shape: ", generated_sen.shape
# print "sen_len_batch_beam: ", sen_len_batch_beam.shape
# print "bert_emd: ", bert_emd.shape
for i in range(beam_width):
fd_seq = {
Seq2Seq_model.encoder_inputs:
source_shuffle,
Seq2Seq_model.encoder_inputs_length:
source_len,
Seq2Seq_model.encoder_emb:
source_emd,
Seq2Seq_model.encoder_char_ids:
whole_noisy_char_Id,
Seq2Seq_model.encoder_char_len:
char_len,
Seq2Seq_model.decoder_inputs:
generated_sen[:, :, i],
Seq2Seq_model.decoder_length:
sen_len_batch_beam[:, i],
# Seq2Seq_model.target_emb: bert_emd,
Seq2Seq_model.dropout_rate:
dropout
}
logits_reward_beam.append(
sess_word_rw.run(Seq2Seq_model.softmax_logits, fd_seq))
answer_shuffle, answer_len, source_shuffle, source_len, source_emd, generated_sen, sen_len_batch_beam, bert_emd, vocab = QA_data
EOS = 1
answer_shuffle = tf.keras.preprocessing.sequence.pad_sequences(
answer_shuffle, maxlen=None, padding='post', value=EOS)
QA_reward_beam = []
for i in range(beam_width):
fd_qa = {
QA_simi_model.answer_inputs: answer_shuffle,
QA_simi_model.answer_inputs_length: answer_len,
QA_simi_model.question1_inputs: source_shuffle,
QA_simi_model.question1_inputs_length: source_len,
QA_simi_model.question2_inputs: generated_sen[:, :, i],
QA_simi_model.question2_inputs_length: sen_len_batch_beam[:, i],
QA_simi_model.question1_emd: source_emd,
QA_simi_model.question2_emd: bert_emd[:, i, :, :]
} #[]
QA_reward_beam.append(
sess_QA_rw.run([QA_simi_model.two_distance], fd_qa))
# print "the shape logits_reward_beam", np.array(logits_reward_beam).shape
logits_reward_beam = np.array(logits_reward_beam).reshape(
batch_size * beam_width, generated_target_len, vocab_size)
QA_reward_beam = np.array(QA_reward_beam).reshape(batch_size * beam_width)
generated_sen = np.array(generated_sen).reshape(batch_size * beam_width,
generated_target_len)
# print "the shape of generated_sen ", generated_sen.shape
logits_batch_beam = []
for b in range(batch_size * beam_width):
logits_batch = []
for l in range(generated_target_len):
id = generated_sen[b][l]
logits_v = logits_reward_beam[b][l][id]
logits_batch.append(logits_v)
logits_batch_beam.append(logits_batch)
bc = BertClient(port=4445, port_out=4446)
gen_sen = []
for idx in range(batch_size * beam_width):
sen = []
for l in range(generated_target_len):
sen.append(vocab[int(generated_sen[idx][l])])
gen_sen.append(" ".join(sen))
bert_pro_beam = bc.encode(gen_sen)
bert_pro_beam = bert_pro_beam[:, 1:generated_target_len + 1, :]
bert_weight_path = config.bert_weight_path
if os.path.isfile(bert_weight_path):
f = open(bert_weight_path, "rb")
data = pickle.load(f)
output_weights = data["bert_W"]
output_bias = data["bert_b"]
output_weights = np.transpose(output_weights)
logits = np.matmul(bert_pro_beam, output_weights)
logits = np.add(logits, output_bias)
log_probs = np.divide(
np.exp(logits),
np.tile(np.sum(np.exp(logits), axis=2),
output_bias.shape[0]).reshape(logits.shape[0],
logits.shape[1],
logits.shape[2]))
# print "the size of log_probs", log_probs.shape
# tf.nn.log_softmax(logits, axis=-1)
bert_Id, bert_vocab_size = bert_vocab_map_Id(generated_sen, config,
vocab)
bert_reward_beam = []
for i in range(batch_size * beam_width):
bert_reward = []
for j in range(generated_target_len):
bert_reward.append(log_probs[i][j][bert_Id[i][j]])
bert_reward_beam.append(bert_reward)
# bert_reward_beam =[]
# for i in range(beam_width):
# gen_sen =[]
# for b in range(batch_size):
# sen = []
# for l in range(generated_sen.shape[1]):
# sen.append(vocab[int(generated_sen[b][l][i])])
# gen_sen.append(" ".join(sen))
# bert_pro = bc.encode(gen_sen)
# bert_pro = bert_pro[:,:generated_target_len,:]
#
# bert_weight_path = config.bert_weight_path
# if os.path.isfile(bert_weight_path):
# f=open(bert_weight_path,"rb")
# data = pickle.load(f)
# output_weights = data["bert_W"]
# output_bias = data["bert_b"]
# output_weights = np.transpose(output_weights)
# logits = np.matmul(bert_pro, output_weights)
# logits = np.add(logits, output_bias)
# log_probs = np.divide(np.exp(logits),np.tile(np.sum(np.exp(logits), axis=2),output_bias.shape[0]).reshape(logits.shape[0],logits.shape[1],logits.shape[2]))
# # tf.nn.log_softmax(logits, axis=-1)
#
# bert_Id = bert_vocab_map_Id(generated_sen, config, vocab)
#
# bert_reward_beam = []
# for i in range(batch_size * beam_width):
# bert_reward = []
# for j in range(generated_target_len):
# bert_reward.append(log_probs[i][j][bert_Id[i][j]])
# bert_reward_beam.append(bert_reward)
RL_reward_beam_dis = []
for beam in range(beam_width * batch_size):
RL_reward = []
for idx in range(generated_target_len):
if idx != len(logits_batch_beam[beam]) - 1:
RL_reward.append(vocab_size * logits_batch_beam[beam][idx] +
bert_vocab_size * bert_reward_beam[beam][idx])
elif idx == len(logits_batch_beam[beam]) - 1:
RL_reward.append(vocab_size * logits_batch_beam[beam][idx] +
bert_vocab_size *
bert_reward_beam[beam][idx] +
10 * QA_reward_beam[beam])
RL_reward_batch_dis = discounted_rewards_cal(RL_reward,
discount_factor)
RL_reward_beam_dis.append(RL_reward_batch_dis)
RL_reward_beam_dis = np.array(RL_reward_beam_dis).reshape(
batch_size, beam_width, generated_target_len)
RL_reward_beam_dis_mm = RL_reward_beam_dis - np.tile(
np.mean(RL_reward_beam_dis, axis=1),
(beam_width, 1)).reshape(batch_size, beam_width, generated_target_len)
# print "the RL_reward_dis: ", np.mean(RL_reward_beam_dis_mm)
RL_reward_beam_dis_mm = np.transpose(RL_reward_beam_dis_mm, (1, 0, 2))
return RL_reward_beam_dis_mm
from service.client import BertClient bc = BertClient() s = ['你好。', '嗨！', '今日我本来打算去公园。'] a = bc.encode(s) # a = bc.encode(['Today is good.']) print(a.shape) # print(s.type) print(a[0][1].shape) print(a[0][1]) print(a[0][4]) print(a[1][3]) print(a[0][5])
