def vertical_attr(self, lenth):
attr = [0.] * lenth
num_examples = testdata_size
for i in range(num_examples // self.batchsize):
data, compaign_data, labels, seqlen, time = loadCriteo.loadCriteoBatch_AMTA(
batchsize=self.batchsize,
max_input=self.feature_number + 2,
max_seq_len=self.max_seq_len,
fin=f_test)
feed_dict = {
self.x: data,
self.y: labels,
self.time: time,
self.seqlen: seqlen,
# self.channel: compaign_data
}
fetches = [self.attribution]
attribution = self.sess.run(fetches, feed_dict)[0]
for i in range(len(labels)):
if seqlen[i] == lenth and labels[i] == 1:
for j in range(seqlen[i]):
# if click_label[i][j] == 1:
v = attribution[j][i]
attr[j] += v
print attr
def train_one_epoch(self):
num_examples = traindata_size
print num_examples
total_loss = []
total_loss_0 = []
total_loss_1 = []
total_loss_2 = []
pred = []
label = []
f_train = open(train_path)
pbar = tqdm.tqdm(total=traindata_size // self.batchsize)
for i in range(num_examples // self.batchsize):
# print("batch {}:".format(i))
pbar.update(1)
data, compaign_data, labels, seqlen, time = loadCriteo.loadCriteoBatch_AMTA(
batchsize=self.batchsize,
max_input=self.feature_number + 2,
max_seq_len=self.max_seq_len,
fin=f_train)
feed_dict = {
self.x: data,
self.y: labels,
self.time: time,
self.seqlen: seqlen,
# self.channel: compaign_data
}
# print(data)
# print(labels)
# print(time)
# print(seqlen)
fetches = [
self.train_step, self.cross_entropy, self.prediction, self.h,
self.loss_0, self.loss_1, self.loss_2
]
result = self.sess.run(fetches, feed_dict=feed_dict)
_, loss, prediction, h, loss_0, loss_1, loss_2 = result
total_loss.append(loss)
total_loss_0.append(loss_0)
total_loss_1.append(loss_1)
total_loss_2.append(loss_2)
pred += prediction.tolist()
label += labels
# print(loss)
print("training finished")
mean_loss = np.mean(total_loss)
mean_loss_0 = np.mean(total_loss_0)
mean_loss_1 = np.mean(total_loss_1)
mean_loss_2 = np.mean(total_loss_2)
auc = roc_auc_score(label, pred)
print("{:.4f} {:.4f} {:.4f} {:.4f} {:.4f}".format(
mean_loss, mean_loss_0, mean_loss_1, mean_loss_2, auc))
f_train.close()
f_log.write("{:.4f} {:.4f} {:.4f} {:.4f} {:.4f}\n".format(
mean_loss, mean_loss_0, mean_loss_1, mean_loss_2, auc))
return mean_loss, auc, mean_loss_0, mean_loss_1, mean_loss_2
def test(self, epoch=0):
num_examples = testdata_size
print num_examples
total_loss = []
total_loss_0 = []
total_loss_1 = []
total_loss_2 = []
pred = []
label = []
f_test = open(test_path)
for i in range(num_examples // self.batchsize):
data, compaign_data, labels, seqlen, time = loadCriteo.loadCriteoBatch_AMTA(
batchsize=self.batchsize,
max_input=self.feature_number + 2,
max_seq_len=self.max_seq_len,
fin=f_test)
feed_dict = {
self.x: data,
self.y: labels,
self.time: time,
self.seqlen: seqlen,
# self.channel: compaign_data
}
fetches = [
self.cross_entropy, self.prediction, self.h, self.loss_0,
self.loss_1, self.loss_2
]
result = self.sess.run(fetches, feed_dict=feed_dict)
loss, prediction, h, loss_0, loss_1, loss_2 = result
total_loss.append(loss)
total_loss_0.append(loss_0)
total_loss_1.append(loss_1)
total_loss_2.append(loss_2)
pred += prediction.tolist()
label += labels
print("testing {}-th epoch finished".format(epoch))
mean_loss = np.mean(total_loss)
mean_loss_0 = np.mean(total_loss_0)
mean_loss_1 = np.mean(total_loss_1)
mean_loss_2 = np.mean(total_loss_2)
auc = roc_auc_score(label, pred)
print("{:.4f} {:.4f} {:.4f} {:.4f} {:.4f}".format(
mean_loss, mean_loss_0, mean_loss_1, mean_loss_2, auc))
f_log.write("{:.4f} {:.4f} {:.4f} {:.4f} {:.4f}\n".format(
mean_loss, mean_loss_0, mean_loss_1, mean_loss_2, auc))
self.log(epoch, [auc], 'test')
f_test.close()
def attr(self):
num_examples = testdata_size
f_test = open(test_path)
outfile = open('./attribute_criteo_s1/AMTA.txt', 'w')
filec = open('index2channel.pkl', 'rb')
Channel = pkl.load(filec)
Channel_value = {}
Channel_time = {}
for i in range(num_examples // self.batchsize):
data, compaign_data, labels, seqlen, time = loadCriteo.loadCriteoBatch_AMTA(
batchsize=self.batchsize,
max_input=self.feature_number + 2,
max_seq_len=self.max_seq_len,
fin=f_test)
feed_dict = {
self.x: data,
self.y: labels,
self.time: time,
self.seqlen: seqlen,
# self.channel: compaign_data
}
fetches = [self.attribution]
attribution = self.sess.run(fetches, feed_dict)[0]
for i in range(self.batchsize):
if labels[i] != 0:
for j in range(seqlen[i]):
# if click_label[i][j] == 1:
index = Channel[str(data[i][j][0])]
v = attribution[j][i]
if index in Channel_value:
Channel_value[index] += v
Channel_time[index] += 1
else:
Channel_value[index] = v
Channel_time[index] = 1
for key in Channel_value:
outfile.write(key + '\t' +
str(Channel_value[key] / Channel_time[key]) + '\n')
