Tensorflow Многослойный персептрон. Ранняя остановка обучения

learning_rate = 0.01 
training_epochs = 25 
batch_size = 100 
display_step = 1 
 
n_hidden_1 = 256
n_hidden_2 = 128
n_hidden_3 = 10
 
x = tf.placeholder("float", [None, 784])
y = tf.placeholder("float", [None, 10])
 
h = tf.Variable(tf.random_normal([784, n_hidden_1])) 
bias_layer_1 = tf.Variable(tf.random_normal([n_hidden_1])) 
layer_1 = tf.nn.sigmoid(tf.add(tf.matmul(x, h), bias_layer_1))
 
w = tf.Variable(tf.random_normal([n_hidden_1, n_hidden_2])) 
bias_layer_2 = tf.Variable(tf.random_normal([n_hidden_2])) 
layer_2 = tf.nn.sigmoid(tf.add(tf.matmul(layer_1, w), bias_layer_2)) 
 
t = tf.Variable(tf.random_normal([n_hidden_2, n_hidden_3])) 
bias_layer_3 = tf.Variable(tf.random_normal([n_hidden_3])) 
layer_3 = tf.nn.sigmoid(tf.add(tf.matmul(layer_2, t), bias_layer_3)) 
 
output = tf.Variable(tf.random_normal([n_hidden_3, 10]))
bias_output = tf.Variable(tf.random_normal([10]))
output_layer = tf.matmul(layer_3, output) + bias_output
 
cost = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits = output_layer, labels = y))
 
optimizer = tf.train.AdamOptimizer(learning_rate = learning_rate).minimize(cost) 
 
with tf.Session() as sess:
    init = tf.global_variables_initializer()
    sess.run(init)
    for epoch in range(training_epochs):
        avg_cost = 0.
        total_batch = int(mnist.train.num_examples / batch_size)
 
        for i in range(total_batch):
            batch_xs, batch_ys = mnist.train.next_batch(batch_size)
            sess.run(optimizer, feed_dict = {x: batch_xs, y: batch_ys}) 
            avg_cost += sess.run(cost, feed_dict = {x: batch_xs, y: batch_ys}) / total_batch
            
        if epoch % display_step == 0:
            print ("Epoch:", '%02d' % (epoch+1), "cost=", "{:.3f}".format(avg_cost)) 
            avg_set.append(avg_cost) 
            epoch_set.append(epoch+1)
    print ("END")
 
    correct_prediction = tf.equal(tf.argmax(output_layer, 1), tf.argmax(y, 1)) 
    accuracy = tf.reduce_mean(tf.cast(correct_prediction, "float"))

with tf.Session() as sess:
    init = tf.global_variables_initializer()
    sess.run(init)
    
    for epoch in range(training_epochs):
        
        avg_cost = 0.
        total_batch = int(mnist.train.num_examples / batch_size)
        
        batch_xs, batch_ys = mnist.train.next_batch(batch_size)
        oldErr = sess.run(cost, feed_dict = {x: batch_xs, y: batch_ys}) / total_batch
        
        for i in range(total_batch):
            batch_xs, batch_ys = mnist.train.next_batch(batch_size)
            sess.run(optimizer, feed_dict = {x: batch_xs, y: batch_ys}) 
            avg_cost += sess.run(cost, feed_dict = {x: batch_xs, y: batch_ys}) / total_batch
            
        if epoch % display_step == 0:
            print ("Epoch:", '%02d' % (epoch+1), "cost=", "{:.3f}".format(avg_cost)) 
            avg_set.append(avg_cost) 
            epoch_set.append(epoch+1)
            test_error = sess.run(cost, feed_dict = {x: batch_xs, y: batch_ys}) / total_batch
            if oldErr < avg_cost:
                print("BREAK")
                break
            oldErr += avg_cost
    print ("END")