Question

我使用一个函数来计算每个时期tf.Session()内的均方误差。但是，如果我在每一步都保存模型，则保存文件的大小会随着步长的增加而增加。

当我使用函数get_mse()时，每个时期之后保存的check_points的大小会增加。但是，当我使用简单的语句来计算mse即MSE时，却不会发生。我在做什么错了？

import os
import numpy as np
import tensorflow as tf
from tensorflow.python.ops.losses.losses_impl import Reduction, compute_weighted_loss
from tensorflow.python.framework import ops
from tensorflow.python.ops import math_ops

def get_mse(
    labels, predictions, weights=1.0, name="mse", scope=None,
    loss_collection=ops.GraphKeys.LOSSES,
    reduction=Reduction.SUM_BY_NONZERO_WEIGHTS):
  with ops.name_scope(scope, name,
                      (predictions, labels, weights)) as scope:
    predictions = math_ops.to_float(predictions)
    labels = math_ops.to_float(labels)
    predictions.get_shape().assert_is_compatible_with(labels.get_shape())
    losses = math_ops.squared_difference(predictions, labels)
    return compute_weighted_loss(
losses, weights, scope, loss_collection, reduction=reduction)

true  = np.random.random((100,1))
pred = np.random.random((100,1))

variable_to_save = tf.Variable(true)
true_ph = tf.placeholder(tf.float32, shape = [None, 1], name='labels')
pred_ph = tf.placeholder(tf.float32, shape = [None, 1], name='predictions')

MSE = tf.reduce_mean(tf.square(pred_ph - true_ph), name='get_mse')

with tf.Session() as sess:
    init = tf.global_variables_initializer()
    sess.run(init)
    saver = tf.train.Saver(max_to_keep=100)
    for epoch in range(10):
        mse = sess.run(get_mse(true_ph, pred_ph), feed_dict={'labels:0': true, 'predictions:0':pred})   #this increases size of checkpoints after each epoch
        #mse = sess.run(MSE, feed_dict={'labels:0': true, 'predictions:0': pred})   #running this does not increases size of check_points
        saver.save(sess, os.getcwd(), global_step=epoch)

Answer 1

每次调用get_mse（）时，都会构建另一堆计算图节点。即每次您创建predictions，labels和losses的另一个版本时。

调用一次get_mse（）并将结果存储到一个变量中，例如mse_computer。然后改用sess.run(mse_computer, feed_dict=....。

希望有帮助

Answer 2

您的代码应如下所示：

import os
import numpy as np
import tensorflow as tf
from tensorflow.python.ops.losses.losses_impl import Reduction, compute_weighted_loss
from tensorflow.python.framework import ops
from tensorflow.python.ops import math_ops

def get_mse(
    labels, predictions, weights=1.0, name="mse", scope=None,
    loss_collection=ops.GraphKeys.LOSSES,
    reduction=Reduction.SUM_BY_NONZERO_WEIGHTS):
    with ops.name_scope(scope, name,
                      (predictions, labels, weights)) as scope:
        predictions.get_shape().assert_is_compatible_with(labels.get_shape())
        losses = math_ops.squared_difference(predictions, labels)
    return compute_weighted_loss(
        losses, weights, scope, loss_collection, reduction=reduction)

true  = np.random.random(size=(100,1))
pred = np.random.random(size=(100,1))
true_ph = tf.placeholder(tf.float32, shape=[None, 1], name='labels')
pred_ph = tf.placeholder(tf.float32, shape=[None, 1], name='predictions')
mse_tensor = get_mse(true_ph, pred_ph)

with tf.Session() as sess:
    init = tf.global_variables_initializer()
    sess.run(init)
    saver = tf.train.Saver(max_to_keep=100)
    for epoch in range(10):
        mse_val = sess.run(mse_tensor, feed_dict={true_ph: true, pred_ph:pred})   #this increases size of checkpoints after each epoch
        saver.save(sess, os.getcwd(), global_step=epoch)

对于您的情况，在每次迭代中，您都需要为图形添加其他操作和张量，因此其大小会增大。

递归使用函数会导致张量流中模型大小的增加

2 个答案: