def test_model(sess, pred, satisfaction_output, aspiration_output, t_X, t_Y,
pl_X, scaler, actionlist, ACTION):
print_important('>>>> testing...')
test_predict = []
test_output = []
test_state = []
for step in range(len(t_X)):
prob, output_, state_ = sess.run(
[pred, satisfaction_output, aspiration_output],
feed_dict={
pl_X: [t_X[step]],
ACTION: [actionlist[step]]
})
predict = prob.reshape((-1))
test_predict.extend(predict)
test_output.extend(output_)
test_state.extend(state_)
test_predict = scaler.inverse_transform(
np.array(test_predict).reshape(-1, 1)) # 将归一化的数据还原成原来的数据
test_y = scaler.inverse_transform(np.array(t_Y).reshape(-1, 1))
mae = mean_absolute_error(y_pred=test_predict, y_true=test_y)
rmse = np.sqrt(mean_squared_error(test_predict, test_y))
test_y = np.reshape(np.reshape(test_y, [-1, 15, 1])[:, -1, :],
[-1, 1]) # 只保留15步最后一步的预测值
test_predict = np.reshape(
np.reshape(test_predict, [-1, 15, 1])[:, -1, :],
[-1, 1]) # 只保留15步最后一步的预测值
print('mae:', mae, ' rmse:', rmse)
return test_predict, test_output, test_state, mae, rmse, test_y
def preprocess(data, batch_size=80, time_step=15, percentage=0.7):
print_important('>>>> preprocessing...')
train_end = int(len(data) * percentage) # 分割数据
batch_index = []
scaler_for_x = MinMaxScaler(feature_range=(0, 1)) # 最小最大值标准化，将数据转换成百分比
scaler_for_y = MinMaxScaler(feature_range=(0, 1))
# scaled_x_data = scaler_for_x.fit_transform(data[:, :-1]) # X
# scaled_y_data = scaler_for_y.fit_transform(data[:, -1:]).reshape(-1) # Y
scaled_x_data = data[:, :-1] # X
scaled_y_data = data[:, -1:].reshape(-1) # Y
label_train = scaled_y_data[:train_end]
label_test = scaled_y_data[train_end:]
normalized_train_data = scaled_x_data[: train_end]
normalized_test_data = scaled_x_data[train_end:]
# 以上获得归一化后的x和y的训练和测试集
train_x, train_y = [], []
for i in range(len(normalized_train_data) - time_step):
if i % batch_size == 0:
batch_index.append(i)
x = normalized_train_data[i: i + time_step]
y = label_train[i:i + time_step, np.newaxis]
train_x.append(x.tolist())
train_y.append(y.tolist())
batch_index.append((len(normalized_train_data) - time_step))
# batch_index中存储了分批次存储好的训练数据的index，下同
test_x, test_y = [], []
for i in range(len(normalized_test_data) - time_step):
x = normalized_test_data[i: i + time_step]
y = label_test[i: i + time_step, np.newaxis]
test_x.append(x.tolist())
test_y.append(y.tolist())
# train_x = np.asarray(train_x)
# train_y = np.asarray(train_y)
# test_x = np.asarray(test_x)
# test_y = np.asarray(test_y)
return batch_index, train_x, train_y, test_x, test_y
def lafee_model(X,ACTION):
print_important('>>>> using lafee_model now...')
batch_size = tf.shape(X)[0]
X_state = X[:, :, :DIM_STATE]
X_action = X[:, :, DIM_STATE:]
with tf.variable_scope('rnn_satisfaction'):
satis_cell = c_LaFeeCell(DIM_STATE, activation=math_ops.sigmoid, name='satis_cell')
init_state = satis_cell.zero_state(batch_size, dtype=tf.float32)
rnn_satis_, satisfaction = tf.nn.dynamic_rnn(satis_cell, X_state, initial_state=init_state, dtype=tf.float32)
rnn_satis = tf.reshape(satisfaction, [-1, DIM_STATE]) # !!!!注意这里的变量使用的是后面的
with tf.variable_scope('rnn_aspiration'):
aspir_cell = c_LaFeeCell(DIM_ACTION, activation=math_ops.sigmoid, name='aspir_cell')
init_state2 = aspir_cell.zero_state(batch_size, dtype=tf.float32)
rnn_aspir_, aspiration = tf.nn.dynamic_rnn(aspir_cell, X_action, initial_state=init_state2, dtype=tf.float32)
rnn_aspir = tf.reshape(aspiration, [-1, DIM_ACTION]) # !!!!注意这里的变量使用的是后面的
X_action_new = tf.reshape(X_action[:,-1,:], [-1, DIM_ACTION])
asp_sat = tf.concat([rnn_satis, rnn_aspir], 1) # 将得到的aspiration和satisfaction拼接在一起
aspiration_l1 = tf.layers.dense(asp_sat, 200, tf.nn.relu, name='aspiration_l1') # hidden layer 输入层增加方式
aspiration_l2 = tf.layers.dense(aspiration_l1, 100, tf.nn.relu, name='aspiration_l2') # hidden layer 输入层增加方式
aspiration_output = tf.layers.dense(aspiration_l2, DIM_ACTION, name='aspiration_output') # output layer 输出层增加方式
asp_act = tf.concat([aspiration_output, X_action_new], 1)
tin_l1 = tf.layers.dense(asp_act, 200, tf.nn.relu, name='tin_l1') # hidden layer 输入层增加方式
tin_l2 = tf.layers.dense(tin_l1, 100, tf.nn.relu, name='tin_l2') # hidden layer 输入层增加方式
tin = tf.layers.dense(tin_l2, 1, name='tin') # output layer 输出层增加方式
satisfaction_l1 = tf.layers.dense(asp_sat, 200, tf.nn.relu, name='satisfaction_l1') # hidden layer 输入层增加方式
satisfaction_l2 = tf.layers.dense(satisfaction_l1, 100, tf.nn.relu, name='satisfaction_l2') # hidden layer 输入层增加方式
satisfaction_output = tf.layers.dense(satisfaction_l2, DIM_STATE,
name='satisfaction_output') # output layer 输出层增加方式
tout_l1 = tf.layers.dense(satisfaction_output, 200, tf.nn.relu, name='tout_l1') # hidden layer 输入层增加方式
tout_l2 = tf.layers.dense(tout_l1, 100, tf.nn.relu, name='tout_l2') # hidden layer 输入层增加方式
tout = tf.layers.dense(tout_l2, 1, name='tout') # output layer 输出层增加方式
pred = tf.where(ACTION,tout,tin)
return pred, satisfaction_output, aspiration_output # 返回的是过了神经网络之后的satisfaction和aspiration
def train(batch_size=80, time_step=15, percentage=0.7, model=lafee_model, cell=c_BasicRNNCell,
summary_dir=TIMESTAMP, num=10):
print_important('>>>> training...')
# 训练数据格式：15X15X27 15X15X1
# 测试数据格式：1X15X27 1X15X1
X = tf.placeholder(tf.float32, shape=[None, time_step, INPUT_SIZE])
# Y = tf.placeholder(tf.float32, shape=[None, time_step, OUTPUT_SIZE])
Y = tf.placeholder(tf.float32, shape=[None,OUTPUT_SIZE])
ACTION= tf.placeholder(tf.bool,shape=[None,1]) # 判断行为是哪个
batch_index, train_x, train_y, test_x, test_y, scaler_for_y, action_list, result_list, id_list,is_logout_test,is_logout_train = getData(
num=1) # 从库中选择num个文件合并训练
is_logout_test = np.asarray(is_logout_test).reshape([-1,1])
is_logout_train = np.asarray(is_logout_train).reshape([-1,1])
train_y = np.reshape(train_y[:,-1,:],[-1,OUTPUT_SIZE])
test_y = np.reshape(test_y[:, -1, :], [-1, OUTPUT_SIZE])
pred, satisfaction_output, aspiration_output = model(X,ACTION)
loss = tf.reduce_mean(tf.square(tf.reshape(pred, [-1]) - tf.reshape(Y, [-1]))) # 求预测值与标签值的均方误差
loss_summary = tf.summary.scalar('loss', loss)
train_op = tf.train.AdamOptimizer(LEARNING_RATE).minimize(loss)
with tf.Session() as sess:
train_log_dir = 'logs/train/' + summary_dir
test_log_dir = 'logs/test/' + summary_dir
os.mkdir('origin_logs/' + summary_dir)
f1 = open("origin_logs/" + summary_dir + "action.csv", 'w', newline='')
f5 = open("origin_logs/" + summary_dir + "time.csv", 'w', newline='')
csv_write = csv.writer(f1)
for data in action_list.tolist():
csv_write.writerow(data)
csv_write = csv.writer(f5)
for data in np.asarray(result_list).reshape([-1, time_step]).tolist():
csv_write.writerow(data)
f1.close()
f5.close()
f3 = open("origin_logs/" + summary_dir + "aspiration.csv", 'w', newline='')
f2 = open("origin_logs/" + summary_dir + "satisfaction.csv", 'w', newline='')
f4 = open("origin_logs/" + summary_dir + "error.csv", 'w', newline='')
f6 = open("origin_logs/" + summary_dir + "loss.csv", 'w', newline='')
f7 = open("origin_logs/" + summary_dir + "user.csv", 'w', newline='')
csv_write2 = csv.writer(f2)
csv_write3 = csv.writer(f3)
csv_write4 = csv.writer(f4)
csv_write5 = csv.writer(f6)
csv_write6 = csv.writer(f7)
csv_write6.writerow(id_list) # 写入id
f7.close()
merged = tf.summary.merge([loss_summary])
train_writer = tf.summary.FileWriter(train_log_dir, sess.graph)
test_writer = tf.summary.FileWriter(test_log_dir)
sess.run(tf.global_variables_initializer())
# 并没有对batch_index进行打乱
for i in range(ITER_TIME):
loss_ = None
summary_train = None
for step in range(len(batch_index) - 1):
if batch_index[step] != batch_index[step + 1]: # 注意这行，如果两个值相同意味着来了新用户
summary_train, _, loss_, pred_ = sess.run(
[merged, train_op, loss, pred],
feed_dict={X: train_x[batch_index[step]: batch_index[step + 1]],
Y: train_y[batch_index[step]: batch_index[step + 1]],
ACTION:is_logout_train[batch_index[step]: batch_index[step + 1]]})
csv_write5.writerow([loss_])
if i % 100 == 0:
print('iter:', i, 'loss:', loss_)
test_predict, test_output, test_state, mae, rmse = test_model(sess, pred, satisfaction_output, aspiration_output, test_x, test_y, X,
scaler_for_y,is_logout_test,ACTION)
# print("test_predict:", test_predict)
csv_write4.writerow([mae, rmse]) # 记录error
for data in np.asarray(test_output).tolist():
csv_write2.writerow(data)
for data in np.asarray(test_state).tolist():
csv_write3.writerow(data)
# print("test_state(aspiration):", test_state)
train_writer.add_summary(summary_train, i)
f2.close()
f3.close()
train_writer.close()
test_writer.close()
return test_predict, test_output, test_state