def predict_similarity(self, sa, sb):
q = [[sa, sb, 0]]
x1, mas1, x2, mas2, y2 = prepare_data(q, self.maxlen)
ls = []
ls2 = []
use_noise.set_value(0.)
for j in range(0, len(x1)):
ls.append(embed(x1[j], 'en', W=self.W))
ls2.append(embed(x2[j], 'jp'))
trconv = np.dstack(ls)
trconv2 = np.dstack(ls2)
emb2 = np.swapaxes(trconv2, 1, 2)
emb1 = np.swapaxes(trconv, 1, 2)
return self.f2sim(emb1, mas1, emb2, mas2)
def evaluate(self, data):
x1, mas1, x2, mas2, y2 = prepare_data(data, self.maxlen)
use_noise.set_value(0.)
n_samples = len(data)
ls = [] # Embedding results of xa
ls2 = [] # Embedding results of xb
for j in range(0, n_samples):
ls.append(embed(x1[j], 'en', W=self.W))
ls2.append(embed(x2[j], 'jp'))
# print "ls: (should be the same ref_embed)", ls
rank_results = []
for i in range(0, n_samples):
# NOTE: mas1 and mas2 are verticle matrix, not a normal one!
# ref_ls refers to n_samples(999,EN) of duplicated ls[i]
# So we can compare the ls[i](EN) with other sentences(999,JP)
# to derive the ranking results for this given article ls[i](EN)
# 用一个英语文章比较所有可能为pairs的日语文章（如999篇）求出ranking
# ref_ls 就是一个重复了999（n_samples）次的文章ls[i]
# 而 ls2 就是可能为paris的999篇日语的文章
ref_ls = [ls[i]] * n_samples
# print "ref_embed", ref_embed
ref_mas1 = np.array([
mas1[:, i],
] * n_samples).T
# print "ref_mas", ref_mas
# print "mas1", mas1
# return mas1, ref_mas
trconv = np.dstack(ref_ls)
trconv2 = np.dstack(ls2)
emb2 = np.swapaxes(trconv2, 1, 2)
emb1 = np.swapaxes(trconv, 1, 2)
pred = self.f2sim(emb1, ref_mas1, emb2, mas2)
rank = pd.Series(pred).rank(ascending=False)[i]
rank_results.append(rank)
print "the round", i, "rank:", rank
return rank_results
def get_mse(self, data):
# list saving the projection results (50 dim):
x1, mas1, x2, mas2, y2 = prepare_data(data, self.maxlen)
# print "Finish preparing the data!"
use_noise.set_value(0.)
n_samples = len(data)
ls = [] # Embedding results of xa
ls2 = [] # Embedding results of xb
for j in range(0, n_samples):
ls.append(embed(x1[j], 'en', W=self.W))
ls2.append(embed(x2[j], 'jp'))
# print "Finished embedding,start projecting..."
# start_time = time.time()
# for i in range(0, n_samples):
# print "conducting the", i, "projection"
# loop_time = time.time()
trconv = np.dstack(ls)
trconv2 = np.dstack(ls2)
emb1 = np.swapaxes(trconv, 1, 2)
emb2 = np.swapaxes(trconv2, 1, 2)
# list saving the projection results (50 dim):
# list_projection1 = self.f_proj11(emb1, mas1)
# list_projection2 = self.f_proj11(emb2, mas2)
c = self.f_cost(emb1, mas1, emb2, mas2, y2)
# After projection, compare the distance for possible pairs
# ## SKIP
return c
def chkterr2(self, mydata):
# count = []
num = len(mydata)
px = []
yx = []
use_noise.set_value(0.)
for i in range(0, num, 256):
q = []
x = i + 256
if x > num:
x = num
for j in range(i, x):
q.append(mydata[j])
x1, mas1, x2, mas2, y2 = prepare_data(q, self.maxlen)
ls = []
ls2 = []
for j in range(0, len(q)):
ls.append(embed(x1[j], 'en', W=self.W))
ls2.append(embed(x2[j], 'jp'))
trconv = np.dstack(ls)
trconv2 = np.dstack(ls2)
emb2 = np.swapaxes(trconv2, 1, 2)
emb1 = np.swapaxes(trconv, 1, 2)
pred = (self.f2sim(emb1, mas1, emb2, mas2)) * 4.0 + 1.0
#dm1=np.ones(mas1.shape,dtype=np.float32)
#dm2=np.ones(mas2.shape,dtype=np.float32)
#corr=f_cost(emb1,mas1,emb2,mas2,y2)
for z in range(0, len(q)):
yx.append(y2[z])
px.append(pred[z])
#count.append(corr)
px = np.array(px)
yx = np.array(yx)
#print "average error= "+str(np.mean(acc))
return np.mean(np.square(px - yx)), meas.pearsonr(
px, yx)[0], meas.spearmanr(yx, px)[0]
def train_lstm(self,
train,
max_epochs,
correct,
test_correct,
batchsize=32):
print "Training"
print "the length of the training data is ", len(train)
# test = train
print "Batchsize =", batchsize
print "max_epochs =", max_epochs
lrate = 0.0001 # Learning rate, but Not USED ???
freq = 0 # ???
batchsize = 64
dfreq = 21 #display frequency
self.mse = [] # MSE of train1 + train2
self.rank = []
self.tops = {}
self.mse_test = [] # MSE of test1
self.mse_train = [] # MSE of train1
self.rank_test = []
self.tops_test = {}
self.top_keys = [1, 5, 10]
print "Before trianing, the error is:"
# print self.chkterr2(train) # MSE check
cst_all = self.chkterr2(train)[0] / 16
self.mse.append(cst_all)
cst_test = self.chkterr2(test_correct)[0] / 16
self.mse_test.append(cst_test)
cst_train = self.chkterr2(correct)[0] / 16
self.mse_train.append(cst_train)
# 【注意】内存不足时使用chkterr2但是会慢，内存足够时使用 , self.get_mse(train)
# 【注意】不要直接使用cst变量作为cost，因为这里的cst是最后一个batch的cost而已，不是全部的
print "Training error:", cst_all #, "==", self.get_mse(train)
print "Training_correct error", cst_train
print "Testing_correct error:", cst_test
# Saving (Initialization) the ranking and top1,5,10 information (Trianing data)
rank_results_train, n_tops = self.evaluate2(
correct, tops=self.top_keys) # Similairty check
# print "[debug]", n_tops
for top_key in self.top_keys:
# print "[debug]", n_tops[top_key]
self.tops[top_key] = []
self.tops[top_key].append(n_tops[top_key])
print "top-", top_key, "=", self.tops[top_key], ":", n_tops[
top_key]
print "Discription of evaluation (ranking) for training data:"
print pd.Series(rank_results_train).describe()
# Saving (Initialization) the ranking and top1,5,10 information (Testing data)
rank_results_test, n_tops_test = self.evaluate2(
test_correct, tops=self.top_keys) # Similairty check
# print "[debug]", n_tops
for top_key in self.top_keys:
# print "[debug]", n_tops[top_key]
self.tops_test[top_key] = []
self.tops_test[top_key].append(n_tops_test[top_key])
print "top-", top_key, "=", self.tops_test[
top_key], ":", n_tops_test[top_key]
print "Discription of evaluation (ranking) for testing data:"
print pd.Series(rank_results_test).describe()
# eidx -> index of epoch
for eidx in xrange(0, max_epochs):
sta = time.time()
print ""
print 'Epoch', eidx, '...'
num = len(train) # length of training data
#---------------------Shuffle the data------------------------------#
# 为何不直接用shuffle函数？
# generates a list with length of num from the population xrange(num)
# Used for shuffling the training data each time for each epoches
# [5,2,6,.11,...] length -> len(train)
rnd = random.sample(xrange(num), num)
# i would be (0,32,64,...)
# Iterate all batches
for i in range(0, num, batchsize):
q = []
x = i + batchsize
if x > num:
x = num
# Shuffle data
# Iterate samples inside each batch
# i -> start index of the batch
# x -> end index of the batch
for z in range(i, x):
# shuffling the training data to the list q
q.append(train[rnd[z]])
#---------------------------------------------------------------------#
"""
Mask for LSTM is prepared by sentence module
x1 = np.array([["我"，"很"，"好"，"，"，"，"，"，"][...]...])
len(x1) => 文档的总数
mas1 = np.array([[1,1,1,0,0,0,0,0,0,0][...]...])
"""
x1, mas1, x2, mas2, y2 = prepare_data(q, self.maxlen)
ls = []
ls2 = []
freq += 1
use_noise.set_value(1.)
for j in range(0, len(x1)):
ls.append(embed(x1[j], 'en', W=self.W))
ls2.append(embed(x2[j], 'jp'))
trconv = np.dstack(ls)
trconv2 = np.dstack(ls2)
emb2 = np.swapaxes(trconv2, 1, 2)
emb1 = np.swapaxes(trconv, 1, 2)
cst = self.f_grad_shared(emb2, mas2, emb1, mas1, y2)
s = self.f_update(lrate) # Not USED ???
if np.mod(freq, dfreq) == 0:
print 'Epoch ', eidx, 'Update ', freq, 'Cost ', cst
# print 'Epoch ', eidx, 'Update ', freq, 'Cost ', cst
# Evalution
# print self.chkterr2(train) # MSE check
cst_all = self.chkterr2(train)[0] / 16
self.mse.append(cst_all)
cst_test = self.chkterr2(test_correct)[0] / 16
self.mse_test.append(cst_test)
cst_train = self.chkterr2(correct)[0] / 16
self.mse_train.append(cst_train)
# 【注意】内存不足时使用chkterr2但是会慢，内存足够时使用 , self.get_mse(train)
# 【注意】不要直接使用cst变量作为cost，因为这里的cst是最后一个batch的cost而已，不是全部的
# 错误用法： print "Training error:", cst, "=", self.chkterr2(train)[0]/16, "==", self.get_mse(train)
print "Training error:", cst_all #, "==", self.get_mse(train)
print "Training_correct error", cst_train
print "Testing_correct error:", cst_test
# Saving the ranking and top1,5,10 information
rank_results_train, n_tops = self.evaluate2(
correct, tops=self.top_keys) # Similairty check
self.rank.append(rank_results_train)
for top_key in self.top_keys:
self.tops[top_key].append(n_tops[top_key])
print "top-", top_key, "=", self.tops[top_key], ":", n_tops[
top_key]
print "Discription of evaluation (ranking) for training data:"
print pd.Series(rank_results_train).describe()
# Saving the ranking and top1,5,10 information
rank_results_test, n_tops_test = self.evaluate2(
test_correct, tops=self.top_keys) # Similairty check
self.rank_test.append(rank_results_test)
for top_key in self.top_keys:
self.tops_test[top_key].append(n_tops_test[top_key])
print "top-", top_key, "=", self.tops_test[
top_key], ":", n_tops_test[top_key]
print "Discription of evaluation (ranking) for testing data:"
print pd.Series(rank_results_test).describe()
# Saving the present weights:
self.save_model(name=self.model_name + "_" + str(eidx) + ".p")
sto = time.time()
self.time_saver = sto - sta
print "epoch took:", self.time_saver