基于张量流的批处理矩阵分解中的内存泄漏

Question

假设我有一个速率矩阵R，我想用tensorflow把它分解成矩阵U和V

没有批量大小是一个简单的问题，可以用下面的代码解决:

# define Variables
u = tf.Variable(np.random.rand(R_dim_1, output_dim), dtype=tf.float32, name='u')
v = tf.Variable(np.random.rand(output_dim, R_dim_2), dtype=tf.float32, name='v')
# predict rate by multiplication 
predicted_R = tf.matmul(tf.cast(u, tf.float32), tf.cast(v, tf.float32))
#cost function and train step
cost = tf.reduce_sum(tf.reduce_sum(tf.abs(tf.sub(predicted_R, R))))
train_step = tf.train.AdamOptimizer(learning_rate).minimize(cost)
with tf.Session() as sess:
    init = tf.initialize_all_variables()
    sess.run(init)
    for i in range(no_epochs):
        _, this_cost = sess.run([train_step, cost])
        print 'cost: ', this_cost

我决定用批量更新来解决这个问题，我的解决方案是发送U和V的索引，我想用它们来预测速率矩阵R，并更新那些选定的索引下面是我的代码(如果需要很多时间，请阅读注释):

# define variables
u = tf.Variable(np.random.rand(R_dim_1, output_dim), dtype=tf.float32, name='u')
v = tf.Variable(np.random.rand(output_dim, R_dim_2), dtype=tf.float32, name='v')
idx1 = tf.placeholder(tf.int32, shape=batch_size1, name='idx1')
idx2 = tf.placeholder(tf.int32, shape=batch_size2, name='idx2')
# get current U and current V by slicing U and V
cur_u = tf.Variable(tf.gather(u, idx1), dtype=tf.float32, name='cur_u')
cur_v = tf.transpose(v)
cur_v = tf.gather(cur_v, idx2)
cur_v = tf.Variable(tf.transpose(cur_v), dtype=tf.float32, name='cur_v')
# predict rate by multiplication 
predicted_R = tf.matmul(tf.cast(cur_u, tf.float32), tf.cast(cur_v, tf.float32))
# get needed rate from rate matrix by slicing it
cur_rate = tf.gather(R, idx1)
cur_rate = tf.transpose(cur_rate)
cur_rate = tf.gather(cur_rate, idx2)
cur_rate = tf.transpose(cur_rate)
#cost function and train step
cost = tf.reduce_sum(tf.reduce_sum(tf.abs(tf.sub(predicted_R, cur_rate))))
train_step = tf.train.AdamOptimizer(learning_rate).minimize(cost)
with tf.Session() as sess:
    # initialize variables 
    init_new_vars_op = tf.initialize_variables([v, u])
    sess.run(init_new_vars_op)
    init = tf.initialize_all_variables()
    rand_idx = np.sort(np.random.randint(0, R_dim_1, batch_size1))
    rand_idx2 = np.sort(np.random.randint(0, R_dim_2, batch_size2))
    sess.run(init, feed_dict={idx1: rand_idx, idx2: rand_idx2})
    for i in range(no_epochs):
        with tf.Graph().as_default():
            rand_idx1 = np.random.randint(0, R_dim_1, batch_size1)
            rand_idx2 = np.random.randint(0, R_dim_2, batch_size2)
            _, this_cost, tmp_u, tmp_v, tmp_cur_u, tmp_cur_v = sess.run([train_step, cost, u, v, cur_u, cur_v],feed_dict={idx1: rand_idx1, idx2: rand_idx2})
        print this_cost
            #update U and V with computed current U and current V
            tmp_u = np.array(tmp_u)
            tmp_u[rand_idx] = tmp_cur_u
            u = tf.assign(u, tmp_u)
            tmp_v = np.array(tmp_v)
            tmp_v[:, rand_idx2] = tmp_cur_v
            v = tf.assign(v, tmp_v)

但是我在u = tf.assign(u, tmp_u)和u = tf.assign(u, tmp_u)有内存泄漏我应用了这个，但是没有得到任何结果。
还有另一种解决方案可以将更新应用于U和V的子集，但遇到了许多其他错误，所以请继续讨论如何解决我的内存泄漏问题。
很抱歉我的问题很长，谢谢你的阅读。

Answer 1

我只是通过发送U和V的更新值作为占位符来解决这个问题，然后将U和V分配给这些传递的参数，这样创建的图形将在不同的迭代中保持不变。下面是代码:

# define variables
u = tf.Variable(np.random.rand(R_dim_1, output_dim), dtype=tf.float32, name='u')
v = tf.Variable(np.random.rand(output_dim, R_dim_2), dtype=tf.float32, name='v')
idx1 = tf.placeholder(tf.int32, shape=batch_size1, name='idx1')
idx2 = tf.placeholder(tf.int32, shape=batch_size2, name='idx2')
#define new place holder for changed values of U and V
last_u = tf.placeholder(tf.float32, shape=[R_dim_1, output_dim], name='last_u')
last_v = tf.placeholder(tf.float32, shape=[output_dim, R_dim_2], name='last_v')
#set U and V to updated ones
change_u = tf.assign(u, last_u)
change_v = tf.assign(v, last_v)
# get current U and current V by slicing U and V
cur_u = tf.Variable(tf.gather(u, idx1), dtype=tf.float32, name='cur_u')
cur_v = tf.transpose(v)
cur_v = tf.gather(cur_v, idx2)
cur_v = tf.Variable(tf.transpose(cur_v), dtype=tf.float32, name='cur_v')
# predict rate by multiplication 
predicted_R = tf.matmul(tf.cast(cur_u, tf.float32), tf.cast(cur_v, tf.float32))
# get needed rate from rate matrix by slicing it
cur_rate = tf.gather(R, idx1)
cur_rate = tf.transpose(cur_rate)
cur_rate = tf.gather(cur_rate, idx2)
cur_rate = tf.transpose(cur_rate)
#cost function and train step
cost = tf.reduce_sum(tf.reduce_sum(tf.abs(tf.sub(predicted_R, cur_rate))))
train_step = tf.train.AdamOptimizer(learning_rate).minimize(cost)
with tf.Session() as sess:
    tmp_u = initial_u;
    tmp_v = initial_v;
    # initialize variables  
    init_new_vars_op = tf.initialize_variables([v, u])
    sess.run(init_new_vars_op, feed_dict={last_u: tmp_u, last_v: tmp_v})
    init = tf.initialize_all_variables()
    rand_idx = np.sort(np.random.randint(0, R_dim_1, batch_size1))
    rand_idx2 = np.sort(np.random.randint(0, R_dim_2, batch_size2))
    sess.run(init, feed_dict={idx1: rand_idx, idx2: rand_idx2})
    for i in range(no_epochs):
        with tf.Graph().as_default():
            rand_idx1 = np.random.randint(0, R_dim_1, batch_size1)
            rand_idx2 = np.random.randint(0, R_dim_2, batch_size2)
            _, this_cost, tmp_u, tmp_v, tmp_cur_u, tmp_cur_v, _, _ =
                sess.run([train_step, cost, u, v, cur_u, cur_v, change_u, change_v],
                feed_dict={idx1: rand_idx1, idx2: rand_idx2, last_u: tmp_u, last_v: tmp_v})
        print this_cost
            # find new values of  U and current V but don't assign to them
            tmp_u = np.array(tmp_u)
            tmp_u[rand_idx] = tmp_cur_u
            tmp_v = np.array(tmp_v)
            tmp_v[:, rand_idx2] = tmp_cur_v