Deep Neural Network in TensorFlow - type along

from tensorflow.examples.tutorials.mnist import input_data
mnist = input_data.read_data_sets(".", one_hot=True, reshape=False)

import tensorflow as tf

learning_rate = 0.001
training_epochs = 20
batch_size = 128
display_step = 1

n_input = 784
n_classes = 10

n_hidden_layer = 256

weights = {
    'hidden_layer':tf.Variable(tf.random_normal([n_input, n_hidden_layer])),
    'out':tf.Variable(tf.random_normal([n_hidden_layer, n_classes]))
}

biases = {
    'hidden_layer': tf.Variable(tf.random_normal([n_input, n_hidden_layer])),
    'out':tf.Variable(tf.random_normal([n_classes]))
}

x = tf.placeholder("float", [None, 28, 28, 1])
y = tf.placeholder("float", [None, n_classes])

x_flat = tf.reshape(x,[-1, n_input])

layer_1 = tf.add(tf.matmul(x_flat, weights['hidden_layer']), biases['hidden_layer'])
layer_1 = tf.nn.relu(layer_1)

# failed with -
# InvalidArgumentError: Dimensions must be equal, but are 784 and 256 for 'Add_5' (op: 'Add') with input shapes: [784,10], [256,10].
logits = tf.add(tf.matmul(layer_1, weights['out']), biases['out'])

cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=logits, labels = y))
optimizer = tf.train.GradientDescentOptimizer(learning_rate=learning_rate).minimize(cost)

init = tf.global_variables_initializer()
with tf.Session() as sess:
    sess.run(init)
    for epoch in range(training_epochs):
        total_batch = int(mnist.train.num_examples/batch_size)
        for i in range(total_batch):
            batch_x, batch_y = mnist.train.next_batch(batch_size)
            #InvalidArgumentError: Incompatible shapes: [128,256] vs. [784,256]
            sess.run(optimizer, feed_dict = {x:batch_x, y:batch_y})