本文介绍了一种基于TensorFlow的YouTubeNet推荐系统模型实现。该模型利用用户历史点击行为及商品特征进行商品推荐，通过训练调整参数以提高推荐准确性。模型涵盖了商品ID、品牌、价格等多维度特征，并使用了批量归一化、Dropout等技术。
youtubenet_model
from config import *
import time
class Model():
def __init__(self, item_dict):
self.embedding_size = 128
self.item_count = len(set(item_dict[:, 0]))
self.item_dict = item_dict
self.brand_count = 10
self.price_count = 15
self.cate2_count = 500
self.color_count = 500
self.corp_count = 2800
self.city_count = 38
self.shop_level_cnt = 10
#self.item_feature_cnt = 5
self.build_model()
def build_model(self):
self.hist_item = tf.placeholder(tf.int32, [None, None]) # history click[B, T]
self.hist_brand = tf.placeholder(tf.int32, [None, None]) # history click[B, T]
self.hist_price = tf.placeholder(tf.int32, [None, None]) # history click[B, T]
self.hist_cate2 = tf.placeholder(tf.int32, [None, None]) # history click[B, T]
self.hist_color = tf.placeholder(tf.int32, [None, None]) # history click[B, T]
self.cate3d = tf.placeholder(tf.int32, [None])
self.cate7d = tf.placeholder(tf.int32, [None])
self.order7d = tf.placeholder(tf.int32, [None])
self.city = tf.placeholder(tf.int32, [None])
self.shop_level = tf.placeholder(tf.int32, [None])
self.sample_item = tf.placeholder(tf.int32, [None, None]) # history click[B, T]
self.sample_brand = tf.placeholder(tf.int32, [None, None]) # history click[B, T]
self.sample_price = tf.placeholder(tf.int32, [None, None]) # history click[B, T]
self.sample_cate2 = tf.placeholder(tf.int32, [None, None]) # history click[B, T]
self.sample_color = tf.placeholder(tf.int32, [None, None]) # history click[B, T]
self.sl = tf.placeholder(tf.int32, [None])
self.keep_prob = tf.placeholder(tf.float32, [])
self.lr = tf.placeholder(tf.float64, [])
item_emb_w = tf.get_variable("item_emb_w", [self.item_count, self.embedding_size//2], initializer=tf.initializers.truncated_normal())
brand_emb_w = tf.get_variable("brand_emb_w", [self.brand_count, self.embedding_size//8], initializer=tf.initializers.truncated_normal())
price_emb_w = tf.get_variable("price_emb_w", [self.price_count, self.embedding_size//8], initializer=tf.initializers.truncated_normal())
cate2_emb_w = tf.get_variable("cate2_emb_w", [self.cate2_count, self.embedding_size//8], initializer=tf.initializers.truncated_normal())
color_emb_w = tf.get_variable("color_emb_w", [self.color_count, self.embedding_size//8], initializer=tf.initializers.truncated_normal())
city_emb_w = tf.get_variable("city_emb_w", [self.city_count, self.embedding_size//8], initializer=tf.initializers.truncated_normal())
shop_level_emb_w = tf.get_variable("shop_level_emb_w", [self.shop_level_cnt, 80], initializer=tf.initializers.truncated_normal())
input_b = tf.get_variable("input_b", [self.item_count], initializer=tf.constant_initializer(0.0))
h_emb = tf.concat([tf.nn.embedding_lookup(item_emb_w, self.hist_item),
tf.nn.embedding_lookup(brand_emb_w, self.hist_brand),
tf.nn.embedding_lookup(price_emb_w, self.hist_price),
tf.nn.embedding_lookup(cate2_emb_w, self.hist_cate2),
tf.nn.embedding_lookup(color_emb_w, self.hist_color)], axis=2)
mask = tf.sequence_mask(self.sl, tf.shape(h_emb)[1], dtype=tf.float32) # [B,T]
mask = tf.expand_dims(mask, -1) # [B,T,1]
mask = tf.tile(mask, [1, 1, tf.shape(h_emb)[2]]) # [B, T, 5*e]
h_emb *= mask # [B, T, 5*e]
hist = tf.reduce_sum(h_emb, 1) # [B, 5*e]
hist = tf.div(hist,
#tf.cast(tf.tile(tf.expand_dims(self.sl, 1), [1, self.item_feature_cnt * self.embedding_size]),
tf.cast(tf.tile(tf.expand_dims(self.sl, 1), [1, self.embedding_size]),
tf.float32)) # [B,e]
#hist = tf.reduce_mean(h_emb, 1)
user_info = tf.concat([tf.nn.embedding_lookup(shop_level_emb_w, self.shop_level),
tf.nn.embedding_lookup(city_emb_w, self.city),
tf.nn.embedding_lookup(cate2_emb_w, self.cate3d),
tf.nn.embedding_lookup(cate2_emb_w, self.cate7d),
tf.nn.embedding_lookup(cate2_emb_w, self.order7d)], axis=-1)
#self.input = tf.concat([hist,
# tf.nn.embedding_lookup(city_emb_w, self.city)
# ], axis=-1)
self.input = tf.concat([user_info, hist], axis=-1)
bn = tf.layers.batch_normalization(inputs=self.input, name='b1')
layer_1 = tf.layers.dense(bn, 512, activation=tf.nn.leaky_relu, name='f1')
layer_1 = tf.nn.dropout(layer_1, keep_prob=self.keep_prob)
layer_2 = tf.layers.dense(layer_1, 256, activation=tf.nn.leaky_relu, name='f2')
layer_2 = tf.nn.dropout(layer_2, keep_prob=self.keep_prob)
#layer_3 = tf.layers.dense(layer_2, self.item_feature_cnt * self.embedding_size, activation=tf.nn.relu,
# name='f3')
self.user_emb = tf.layers.dense(layer_2, self.embedding_size, activation=tf.nn.leaky_relu, name='f3')
sample_w = tf.concat([tf.nn.embedding_lookup(item_emb_w, self.sample_item),
tf.nn.embedding_lookup(brand_emb_w, self.sample_brand),
tf.nn.embedding_lookup(price_emb_w, self.sample_price),
tf.nn.embedding_lookup(cate2_emb_w, self.sample_cate2),
tf.nn.embedding_lookup(color_emb_w, self.sample_color),
# tf.tile(tf.expand_dims(self.basic, 1), [1, tf.shape(next_b)[1], 1])
], axis=-1) # [B,5*e]
self.item_emb = sample_w
#bn = tf.layers.batch_normalization(sample_w, name='b2')
#item_layer_1 = tf.layers.dense(bn, 512, activation=tf.nn.leaky_relu, name='f4')
#item_layer_1 = tf.nn.dropout(item_layer_1, keep_prob=self.keep_prob)
#item_layer_2 = tf.layers.dense(item_layer_1, 256, activation=tf.nn.leaky_relu, name='f5')
#item_layer_2 = tf.nn.dropout(item_layer_2, keep_prob=self.keep_prob)
#self.item_emb = tf.layers.dense(item_layer_2, self.embedding_size, activation=tf.nn.leaky_relu, name='f6')
sample_b = tf.nn.embedding_lookup(input_b, self.sample_item) # [B,5e]
user_v = tf.expand_dims(self.user_emb, axis=1)
user_x_item = tf.squeeze(tf.matmul(user_v, self.item_emb, transpose_b=True), axis=1)
user_norm = tf.sqrt(tf.reduce_sum(tf.square(user_v), axis=-1))
sample_norm = tf.sqrt(tf.reduce_sum(tf.square(sample_w), axis=-1))
sim = user_x_item / user_norm / sample_norm * 20
prob = tf.nn.softmax(sim)
self.sim = prob
#self.logits = tf.squeeze(tf.matmul(user_v, next_w), axis=1) + next_b
prob = tf.slice(prob, [0, 0], [-1, 1])
global_step = tf.Variable(0, trainable=False, name='global_step')
initial_learning_rate = self.lr
self.learning_rate = tf.train.exponential_decay(initial_learning_rate,
global_step=global_step,
decay_steps=20000,
decay_rate=0.9)
self.add_global = global_step.assign_add(1)
#self.global_epoch_step = tf.Variable(0, trainable=False, name='global_epoch_step')
#self.global_epoch_step_op = tf.assign(self.global_epoch_step, self.global_epoch_step + 1)
##self.yhat = tf.nn.softmax(self.logits)
##self.loss = tf.reduce_mean(-self.y * tf.log(self.yhat + 1e-24))
#self.loss = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=self.logits, labels=self.y),name='loss')
self.loss = -tf.reduce_sum(tf.log(prob+1e-8))
#trainable_params = tf.trainable_variables()
#self.opt = tf.train.GradientDescentOptimizer(learning_rate=self.lr)
#gradients = tf.gradients(self.loss, trainable_params)
#clip_gradients, _ = tf.clip_by_global_norm(gradients, 5)
#self.train_op = self.opt.apply_gradients(
# zip(clip_gradients, trainable_params), global_step=self.global_step)
#optimizer = tf.train.AdamOptimizer(self.lr, name='adam')
optimizer = tf.train.AdamOptimizer(self.learning_rate, name='adam')
#optimizer = tf.train.GradientDescentOptimizer(self.learning_rate)
self.train_op = optimizer.minimize(self.loss)
self.item_emb = tf.concat([item_emb_w,
tf.nn.embedding_lookup(brand_emb_w, tf.convert_to_tensor(self.item_dict[:, 1], dtype=tf.int32)),
tf.nn.embedding_lookup(price_emb_w, tf.convert_to_tensor(self.item_dict[:, 2], dtype=tf.int32)),
tf.nn.embedding_lookup(cate2_emb_w, tf.convert_to_tensor(self.item_dict[:, 3], dtype=tf.int32)),
tf.nn.embedding_lookup(color_emb_w, tf.convert_to_tensor(self.item_dict[:, 5], dtype=tf.int32)),
], axis=-1)
#self.user_x_item = tf.matmul(layer_3, item_emb, transpose_b=True) + input_b
#self.output = tf.nn.softmax(self.user_item)
def train(self, sess, uij):
loss, _, _, lr = sess.run([self.loss, self.train_op, self.add_global, self.learning_rate], feed_dict={
self.hist_item: uij[1],
self.hist_brand: uij[2],
self.hist_price: uij[3],
self.hist_cate2: uij[4],
self.hist_color: uij[5],
self.city: uij[6],
self.cate3d: uij[7],
self.cate7d: uij[8],
self.order7d: uij[9],
self.shop_level: uij[10],
self.sample_item: uij[14],
self.sample_brand: uij[15],
self.sample_price: uij[16],
self.sample_cate2: uij[17],
self.sample_color: uij[18],
self.sl: uij[19],
self.keep_prob: 0.9,
self.lr: 0.3
})
#for i in tmp:
# print(i)
return loss
def test(self, sess, uij):
user_emb = sess.run(self.user_emb, feed_dict={
self.hist_item: uij[1],
self.hist_brand: uij[2],
self.hist_price: uij[3],
self.hist_cate2: uij[4],
self.hist_color: uij[5],
self.city: uij[6],
self.cate3d: uij[7],
self.cate7d: uij[8],
self.order7d: uij[9],
self.shop_level: uij[10],
self.sl: uij[12],
self.keep_prob: 1.0
})
return user_emb
def save(self, sess, path):
saver = tf.train.Saver()
saver.save(sess, save_path=path)
def restore(self, sess, path):
saver = tf.train.Saver()
saver.restore(sess, save_path=path)
youtubenet_iter_train
import numpy as np
from config import *
item_dict = []
with open(os.path.join(DATA_PATH, "item_dict.new"), encoding='utf-8') as f:
for line in f.readlines():
item_dict.append(line.strip().split('\t'))
item_dict = np.array(item_dict)
id_li = sorted(list(set(item_dict[:, 0])))
brand_li = [x if x!= '' else '无' for x in sorted(list(set(item_dict[:, 1])))]
price_li = [x if x!= '' else '无' for x in sorted(list(set(item_dict[:, 2])))]
cate2_li = [x if x!= '' else '无' for x in sorted(list(set(item_dict[:, 3])))]
cate3_li = [x if x!= '' else '无' for x in sorted(list(set(item_dict[:, 4])))]
color_li = [x if x!= '' else '无' for x in sorted(list(set(item_dict[:, 5])))]
id_dict = {v: k + 1 for k, v in enumerate(id_li)}
reverse_id_dict = {v: k for k, v in id_dict.items()}
brand_dict = {v: k + 1 for k, v in enumerate(brand_li)}
price_dict = {v: k + 1 for k, v in enumerate(price_li)}
cate2_dict = {v: k + 1 for k, v in enumerate(cate2_li)}
cate3_dict = {v: k + 1 for k, v in enumerate(cate3_li)}
color_dict = {v: k + 1 for k, v in enumerate(color_li)}
mapping = {0: id_dict, 1: brand_dict, 2: price_dict, 3: cate2_dict, 4: cate3_dict, 5: color_dict}
all_item = {}
for d in item_dict:
tmp = []
for ind in range(len(d)):
if d[ind] == '':
d[ind] = '无'
tmp.append(mapping[ind][d[ind]])
all_item[d[0]] = tmp
all_item[''] = [0] * 6
for i, d in enumerate(item_dict):
d[0] = id_dict[d[0]]
d[1] = brand_dict[d[1]]
d[2] = price_dict[d[2]]
d[3] = cate2_dict[d[3]]
d[4] = cate3_dict[d[4]]
d[5] = color_dict[d[5]]
item_dict[i] = d
class TrainIterator:
def __init__(self, source, max_batch_size=1000):
self.source = open(source, 'r')
self.source_dicts = []
self.batch_size = batch_size
self.source_buffer = []
self.k = batch_size * max_batch_size
self.end_of_data = False
def __iter__(self):
return self
def reset(self):
self.source.seek(0)
def __next__(self):
if self.end_of_data:
self.end_of_data = False
self.reset()
raise StopIteration
user_id, hist_items, hist_brand, hist_price, hist_cate2, hist_color, city, cate3d, cate7d, order7d, shop_level, \
reg_till_cur, reg_till_order, cur_end_time_diff, sample_items, sample_brand, sample_price, sample_cate2, sample_color, sl = \
[], [], [], [], [], [], [], [], [], [], [], [], [], [], [], [], [], [], [], []
if len(self.source_buffer) == 0:
for k_ in range(self.k):
ss = self.source.readline()
if ss == "":
break
self.source_buffer.append(ss.strip().split("\t"))
if len(self.source_buffer) == 0:
self.end_of_data = False
self.reset()
raise StopIteration
try:
while True:
if len(self.source_buffer) >= self.batch_size:
end = self.batch_size
else:
end = len(self.source_buffer)
buf = np.array(self.source_buffer[:end])
user_id = np.array(buf[:, 0], np.str)
#continuous_feat = np.array(buf[:, 5:9], np.float)
city = [prov_dict[x] for x in buf[:, 1]]
shop_level = [int(x) for x in buf[:, 2]]
cate3d = [int(cate2_dict[x]) for x in buf[:, 3]]
cate7d = [int(cate2_dict[x]) for x in buf[:, 4]]
order7d = [int(cate2_dict[x]) for x in buf[:, 5]]
hist_ck_items = [x.split(",") for x in buf[:, 9]]
hist_items = [[all_item[y][0] for y in x] for x in hist_ck_items]
hist_items = [x+[0]*(SEQ_LEN-len(x)) if len(x) <= SEQ_LEN else x[:SEQ_LEN] for x in hist_items]
hist_brand = [[all_item[y][1] for y in x] for x in hist_ck_items]
hist_brand = [x+[0]*(SEQ_LEN-len(x)) if len(x) <= SEQ_LEN else x[:SEQ_LEN] for x in hist_brand]
hist_price = [[all_item[y][2] for y in x] for x in hist_ck_items]
hist_price = [x+[0]*(SEQ_LEN-len(x)) if len(x) <= SEQ_LEN else x[:SEQ_LEN] for x in hist_price]
hist_cate2 = [[all_item[y][3] for y in x] for x in hist_ck_items]
hist_cate2 = [x+[0]*(SEQ_LEN-len(x)) if len(x) <= SEQ_LEN else x[:SEQ_LEN] for x in hist_cate2]
hist_color = [[all_item[y][5] for y in x] for x in hist_ck_items]
hist_color = [x+[0]*(SEQ_LEN-len(x)) if len(x) <= SEQ_LEN else x[:SEQ_LEN] for x in hist_color]
next_item = buf[:, 10]
next_brand = [all_item[x][1] for x in next_item]
next_price = [all_item[x][2] for x in next_item]
next_cate2 = [all_item[x][3] for x in next_item]
next_color = [all_item[x][5] for x in next_item]
next_item = [all_item[x][0] for x in next_item]
neg_items = [x.split(",")[:NEG_LEN] for x in buf[:, 11]]
neg_brand = [[all_item[y][1] for y in x] for x in neg_items]
neg_price = [[all_item[y][2] for y in x] for x in neg_items]
neg_cate2 = [[all_item[y][3] for y in x] for x in neg_items]
neg_color = [[all_item[y][5] for y in x] for x in neg_items]
neg_items = [[all_item[y][0] for y in x] for x in neg_items]
sample_items = [[next_item[ind]] + neg_items[ind] for ind in range(len(neg_items))]
sample_brand = [[next_brand[ind]] + neg_brand[ind] for ind in range(len(neg_brand))]
sample_price = [[next_price[ind]] + neg_price[ind] for ind in range(len(neg_price))]
sample_cate2 = [[next_cate2[ind]] + neg_cate2[ind] for ind in range(len(neg_cate2))]
sample_color = [[next_color[ind]] + neg_color[ind] for ind in range(len(neg_color))]
sl = [len(x) for x in hist_ck_items]
del self.source_buffer[:end]
break
except IOError:
self.end_of_data = True
if len(user_id) == 0:
user_id, hist_items, hist_brand, hist_price, hist_cate2, hist_color, city, cate3d, cate7d, order7d, shop_level, \
reg_till_cur, reg_till_order, cur_end_time_diff, sample_items, sample_brand, sample_price, sample_cate2, sample_color, sl = self.next()
return [user_id, hist_items, hist_brand, hist_price, hist_cate2, hist_color, city, cate3d, cate7d, order7d, shop_level, \
reg_till_cur, reg_till_order, cur_end_time_diff, sample_items, sample_brand, sample_price, sample_cate2, sample_color, sl]
daily
from config import *
from youtubenet_model import Model
from youtubenet_iter_train import *
import time
import sys
def train():
train_data = TrainIterator(TRAIN_PATH)
with tf.Session(config=gpu_config) as sess:
model = Model(item_dict)
sess.run(tf.global_variables_initializer())
sess.run(tf.local_variables_initializer())
print("Model built. {}".format(time.strftime('%Y-%m-%d %H:%M:%S', time.localtime(time.time()))))
sys.stdout.flush()
avg_loss = 0
step = 1
for _ in range(epoch):
for uij in train_data:
loss_val = model.train(sess, uij)
avg_loss += loss_val
if step % every_k == 0:
print("step: {0}. avg loss: {1}. {2}".format(step, avg_loss/every_k, time.strftime('%Y-%m-%d %H:%M:%S', time.localtime(time.time()))))
sys.stdout.flush()
avg_loss = 0
step += 1
model.save(sess, os.path.join(MODEL_PATH, "youtubenet.ckpt"))
if __name__ == '__main__':
train()
youtube_iter_test
import numpy as np
from config import *
item_dict = []
with open(os.path.join(DATA_PATH, "item_dict.old"), encoding='utf-8') as f:
for line in f.readlines():
item_dict.append(line.strip().split('\t'))
item_dict = np.array(item_dict)
id_li = sorted(list(set(item_dict[:, 0])))
brand_li = [x if x!= '' else '无' for x in sorted(list(set(item_dict[:, 1])))]
price_li = [x if x!= '' else '无' for x in sorted(list(set(item_dict[:, 2])))]
cate2_li = [x if x!= '' else '无' for x in sorted(list(set(item_dict[:, 3])))]
cate3_li = [x if x!= '' else '无' for x in sorted(list(set(item_dict[:, 4])))]
color_li = [x if x!= '' else '无' for x in sorted(list(set(item_dict[:, 5])))]
id_dict = {v: k + 1 for k, v in enumerate(id_li)}
reverse_id_dict = {v: k for k, v in id_dict.items()}
brand_dict = {v: k + 1 for k, v in enumerate(brand_li)}
price_dict = {v: k + 1 for k, v in enumerate(price_li)}
cate2_dict = {v: k + 1 for k, v in enumerate(cate2_li)}
cate3_dict = {v: k + 1 for k, v in enumerate(cate3_li)}
color_dict = {v: k + 1 for k, v in enumerate(color_li)}
mapping = {0: id_dict, 1: brand_dict, 2: price_dict, 3: cate2_dict, 4: cate3_dict, 5: color_dict}
all_item = {}
for d in item_dict:
tmp = []
for ind in range(len(d)):
if d[ind] == '':
d[ind] = '无'
tmp.append(mapping[ind][d[ind]])
all_item[d[0]] = tmp
all_item[''] = [0] * 6
for i, d in enumerate(item_dict):
d[0] = id_dict[d[0]]
d[1] = brand_dict[d[1]]
d[2] = price_dict[d[2]]
d[3] = cate2_dict[d[3]]
d[4] = cate3_dict[d[4]]
d[5] = color_dict[d[5]]
item_dict[i] = d
class TestIterator:
def __init__(self, source, max_batch_size=1000):
self.source = open(source, 'r')
self.source_dicts = []
self.batch_size = batch_size
self.source_buffer = []
self.k = batch_size * max_batch_size
self.end_of_data = False
def __iter__(self):
return self
def reset(self):
self.source.seek(0)
def __next__(self):
if self.end_of_data:
self.end_of_data = False
self.reset()
raise StopIteration
user_id, hist_items, hist_brand, hist_price, hist_cate2, hist_color, city, cate3d, cate7d, order7d, shop_level, continuous_feat, sl = \
[], [], [], [], [], [], [], [], [], [], [], [], []
if len(self.source_buffer) == 0:
for k_ in range(self.k):
ss = self.source.readline()
if ss == "":
break
self.source_buffer.append(ss.strip().split("\t"))
if len(self.source_buffer) == 0:
self.end_of_data = False
self.reset()
raise StopIteration
try:
while True:
if len(self.source_buffer) >= self.batch_size:
end = self.batch_size
else:
end = len(self.source_buffer)
buf = np.array(self.source_buffer[:end])
user_id = np.array(buf[:, 0], np.str)
city = [prov_dict[x] for x in buf[:, 1]]
shop_level = [int(x) if x != '' else 1 for x in buf[:, 2]]
cate3d = [cate2_dict.get(x, 0) for x in buf[:, 3]]
cate7d = [cate2_dict.get(x, 0) for x in buf[:, 4]]
order7d = [cate2_dict.get(x, 0) for x in buf[:, 5]]
hist_ck_items = [x.split(",") for x in buf[:, 9]]
hist_items = [[all_item[y][0] for y in x if y in all_item] for x in hist_ck_items]
hist_items = [x+[0]*(SEQ_LEN-len(x)) if len(x) <= SEQ_LEN else x[:SEQ_LEN] for x in hist_items]
hist_brand = [[all_item[y][1] for y in x if y in all_item] for x in hist_ck_items]
hist_brand = [x+[0]*(SEQ_LEN-len(x)) if len(x) <= SEQ_LEN else x[:SEQ_LEN] for x in hist_brand]
hist_price = [[all_item[y][2] for y in x if y in all_item] for x in hist_ck_items]
hist_price = [x+[0]*(SEQ_LEN-len(x)) if len(x) <= SEQ_LEN else x[:SEQ_LEN] for x in hist_price]
hist_cate2 = [[all_item[y][3] for y in x if y in all_item] for x in hist_ck_items]
hist_cate2 = [x+[0]*(SEQ_LEN-len(x)) if len(x) <= SEQ_LEN else x[:SEQ_LEN] for x in hist_cate2]
hist_color = [[all_item[y][5] for y in x if y in all_item] for x in hist_ck_items]
hist_color = [x+[0]*(SEQ_LEN-len(x)) if len(x) <= SEQ_LEN else x[:SEQ_LEN] for x in hist_color]
sl = [len(x) for x in hist_ck_items]
del self.source_buffer[:end]
break
except IOError:
self.end_of_data = True
if len(user_id) == 0:
user_id, hist_items, hist_brand, hist_price, hist_cate2, hist_color, city, cate3d, cate7d, order7d, shop_level, continuous_feat, sl = self.next()
return [user_id, hist_items, hist_brand, hist_price, hist_cate2, hist_color, city, cate3d, cate7d, order7d, shop_level, continuous_feat, sl]
hourly
import time
from config import *
from youtubenet_model import Model
from youtubenet_iter_test import *
def dump_emb(emb, filename):
with open(filename, "w") as f:
for k, v in emb.items():
f.write("{0} {1}\n".format(k, ",".join(v)))
def update():
model = Model(item_dict)
with tf.Session(config=gpu_config) as sess:
model.restore(sess, os.path.join(MODEL_PATH, "youtubenet.ckpt"))
cnt = 0
user_res = dict()
test_data = TestIterator(TEST_PATH)
for uij in test_data:
output = model.test(sess, uij)
for ind, u in enumerate(uij[0]):
user_res[u] = list(map(lambda x: str(x), output[ind]))
cnt += 1
if cnt % 100000 == 0:
print("{0} user embedding done. {1}".format(cnt, time.strftime('%Y-%m-%d %H:%M:%S', time.localtime(time.time()))))
sys.stdout.flush()
dump_emb(user_res, os.path.join(DATA_PATH, "user_res"))
online_item = set([line.strip() for line in open(os.path.join(DATA_PATH, "online_item"))])
item_emb = sess.run(model.item_emb)
item_all = dict()
for ind, vec in enumerate(item_emb):
item_id = reverse_id_dict[ind+1]
vec = list(map(lambda x: str(x), vec))
item_all[item_id] = vec
common_item = set(item_all.keys()).intersection(online_item)
item_res = {x: item_all[x] for x in common_item}
dump_emb(item_res, os.path.join(DATA_PATH, "item_res"))
if __name__ == '__main__':
update()
扫一扫
1564
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
