时代花费越来越多的时间

Question

我可能有一个“臃肿的图表”（Why does tf.assign() slow the execution time?），因为每个时代都需要越来越多的时间，但我在我的代码中看不到它。 你能帮帮我吗，还是一个 Tensorflow 新手。

# NEURAL NETWORK
def MLP(x, weights, biases, is_training):

    # Hiden layer 1
    hLayer1 = tf.add(tf.matmul(x, weights["w1"]), biases["b1"])
    hLayer1 = tf.nn.sigmoid(hLayer1)
    bn1 = batch_norm_wrapper(hLayer1, gamma=weights["gamma1"], beta=weights["beta1"], is_training=is_training, name="1")
    hLayer1 = bn1


    # Hiden layer 2
    hLayer2 = tf.add(tf.matmul(hLayer1, weights["w2"]), biases["b2"])
    hLayer2 = tf.nn.sigmoid(hLayer2)
    bn2 = batch_norm_wrapper(hLayer2, gamma=weights["gamma2"], beta=weights["beta2"], is_training=is_training, name="2")
    hLayer2 = bn2

    # Output layer
    outLayer = tf.add(tf.matmul(hLayer2, weights["wOut"]), biases["bOut"], name="outLayer")

    return outLayer



# Weights and biases
weights = {
    "w1": tf.get_variable(shape=[n_input, n_hLayer1], initializer=tf.keras.initializers.he_normal(seed=5), name="w1", trainable=True),
    "w2": tf.get_variable(shape=[n_hLayer1, n_hLayer2], initializer=tf.keras.initializers.he_normal(seed=5), name="w2", trainable=True),
    "wOut": tf.get_variable(shape=[n_hLayer2, n_classes], initializer=tf.keras.initializers.he_normal(seed=5), name="wOut", trainable=True),

    "gamma1": tf.get_variable(shape=[n_hLayer1], initializer=tf.ones_initializer(), name="gamma1", trainable=True),
    "beta1": tf.get_variable(shape=[n_hLayer1], initializer=tf.zeros_initializer(), name="beta1", trainable=True),

    "gamma2":tf.get_variable(shape=[n_hLayer2], initializer=tf.ones_initializer(), name="gamma2", trainable=True),
    "beta2": tf.get_variable(shape=[n_hLayer2], initializer=tf.zeros_initializer(), name="beta2", trainable=True)
}

biases = {
    "b1": tf.get_variable(shape=[n_hLayer1], initializer=tf.zeros_initializer(), name="b1", trainable=True),
    "b2": tf.get_variable(shape=[n_hLayer2], initializer=tf.zeros_initializer(), name="b2", trainable=True),
    "bOut": tf.get_variable(shape=[n_classes], initializer=tf.zeros_initializer(), name="bOut", trainable=True)
}


def batch_norm_wrapper(inputs, gamma, beta,  is_training, name, decay=0.999):

    pop_mean = tf.Variable(tf.zeros([inputs.get_shape()[-1]]), name="pop_mean{}".format(name), trainable=False)
    pop_var = tf.Variable(tf.zeros([inputs.get_shape()[-1]]), name="pop_var{}".format(name), trainable=False)

    if is_training:
        batch_mean, batch_var = tf.nn.moments(inputs, [0])

        train_mean = tf.assign(pop_mean, pop_mean*decay + batch_mean*(1-decay))
        train_var = tf.assign(pop_var, pop_var*decay + batch_var*(1-decay))


        with tf.control_dependencies([train_mean, train_var]):
            return tf.nn.batch_normalization(x=inputs, mean=batch_mean, variance=batch_var, scale=gamma, offset=beta, variance_epsilon=0.001)

    else:
        return tf.nn.batch_normalization(x=inputs, mean=pop_mean, variance=pop_var, scale=gamma, offset=beta, variance_epsilon=0.001)


# Model
predictions = MLP(next_element[0], weights, biases, is_training=True)


# Loss function and regularization
loss = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=predictions, labels=next_element[1]))
l1_regularizer = tf.reduce_sum(tf.abs(weights["w1"])) + tf.reduce_sum(tf.abs(weights["w2"])) + tf.reduce_sum(tf.abs(weights["wOut"]))
l2_regularizer = tf.reduce_mean(tf.nn.l2_loss(weights["w1"]) + tf.nn.l2_loss(weights["w2"]) + tf.nn.l2_loss(weights["wOut"]))
loss = loss + r*alpha1*l1_regularizer + (1-r)*alpha2*l2_regularizer

# Optimizer
optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate).minimize(loss)



# LAUNCH THE GRAPH
with tf.Session() as sess:

    sess.run(init_op)

    # Training
    for trainEpoch in range(training_epochs):

        sess.run(training_iterator_op)

        while True:

            try:
                value = sess.run(next_element)
                sess.run([loss, optimizer])

            except tf.errors.OutOfRangeError:
                break

我使用数据集 API 来运行我的训练数据。

Answer 1

我想我会在 TensorFlow 2.x.x（在我的例子中是 2.4.1）中分享与您的问题相关的训练问题的发现。 Here 是我在互联网上数小时研究后发现的。

Answer 2

有一些原因会导致在训练期间变慢：

• 这很可能是由于您的训练循环坚持了一些不该做的事情。还要确保您不会在不删除它们的情况下将一些临时计算存储在不断增长的列表中。

• 如果您使用自定义的network/loss 函数，计算也可能会随着您接近最优解而变得更加昂贵。为了追踪这一点，您可以分别获取不同部分的时间：数据加载、网络转发、损失计算、反向传递和参数更新。希望只有一个会增加，您将能够更好地了解正在发生的事情。

参考： Training gets slow down by each batch slowly [in Pytorch]

另请阅读： training slow down [in Keras]