02b_tf_mnist.py

import numpy as np
import tensorflow as tf

img_placeholder = tf.placeholder(
    tf.float32,
    shape = (None, 28, 28),
    name = "img"
)

rescaled_image = (img_placeholder - 128) / 128

flattened_img = tf.reshape(
    rescaled_image,
    (-1, 28 * 28)
)

weight_matrix1 = tf.Variable(
    tf.random_normal(
        shape = (28 * 28, 128),
        stddev = np.sqrt(1/(28 *28))
    ),
    name = 'weight_matrix1',
)

bias_vector1 = tf.Variable(
    tf.zeros(shape = (128,)),
    name = 'bias_vector1',
)

z1 = tf.matmul(flattened_img, weight_matrix1) + bias_vector1
h1 = tf.nn.sigmoid(z1)

weight_matrix2 = tf.Variable(
    tf.random_normal(
        shape = (128, 10),
        stddev = np.sqrt(1/128)
    ),
    name = 'weight_matrix2',
)

bias_vector2 = tf.Variable(
    tf.zeros(shape = (10,)),
    name = 'bias_vector2',
)

z2 = tf.matmul(h1, weight_matrix2) + bias_vector2
h2 = tf.nn.softmax(z2)

correct_y = tf.placeholder(tf.int32, shape = (None,))
one_hot_correct_y = tf.one_hot(correct_y, 10)

eps = 1e-6
error = tf.reduce_mean(
    tf.reduce_sum(one_hot_correct_y * -tf.log(h2 + eps), axis = 1)
)

predictions = tf.argmax(
    h2,
    axis = 1,
    output_type = tf.int32
)

accuracy = tf.reduce_mean(
    tf.cast(tf.equal(correct_y, predictions), tf.float32)
)

optimizer = tf.train.GradientDescentOptimizer(
    learning_rate = 1e-2
)

train_step = optimizer.minimize(error)

session = tf.Session()
session.run(tf.global_variables_initializer())

BATCH_SIZE = 128

from keras.datasets import mnist
(x_train, y_train), (x_test, y_test) = mnist.load_data()
for epoch_idx in range(1, 101):
    for batch_idx, batch_start in enumerate(range(0, x_train.shape[0], BATCH_SIZE)):
        x_batch = x_train[batch_start:(batch_start + BATCH_SIZE)]
        y_batch = y_train[batch_start:(batch_start + BATCH_SIZE)]

        _, e, a = session.run(
            [train_step, error, accuracy],
            feed_dict = {
                img_placeholder: x_batch,
                correct_y: y_batch
            }
        )

        if batch_idx % 100 == 0:
            print(f'epoch: {epoch_idx:04d} | batch {batch_idx:04d} | err: {e:0.1f} | acc: {a:0.2f}')

    e, a = session.run(
        [error, accuracy],
        feed_dict = {
            img_placeholder: x_test,
            correct_y: y_test
        }
    )

    print(f'>> epoch: {epoch_idx:04d} | err: {e:0.1f} | acc: {a:0.2f}')