sigmoid_softmax.py

# with sigmoid function -> accuracy : 0.9108
W1 = tf.Variable(tf.random_normal([784, neurons]), name='weight1')
b1 = tf.Variable(tf.random_normal([neurons]), name='bias1')
logits1 = tf.matmul(X, W1) + b1
layer1 = tf.sigmoid(logits1)

W2 = tf.Variable(tf.random_normal([neurons, neurons]), name='weight2')
b2 = tf.Variable(tf.random_normal([neurons]), name='bias3')
logits2 = tf.matmul(layer1, W2) + b2
layer2 = tf.sigmoid(logits2)

W3 = tf.Variable(tf.random_normal([neurons, neurons]), name='weight3')
b3 = tf.Variable(tf.random_normal([neurons]), name='bias3')
logits3 = tf.matmul(layer2, W3) + b3
layer3 = tf.sigmoid(logits3)

W4 = tf.Variable(tf.random_normal([neurons, nb_classes]), name='weight4')
b4 = tf.Variable(tf.random_normal([nb_classes]), name='bias4')

logits4 = tf.matmul(layer3, W4) + b4
hypothesis = tf.nn.softmax(logits4)

# with softmax function -> accuracy : 0.488
W1 = tf.Variable(tf.random_normal([784, neurons]), name='weight1')
b1 = tf.Variable(tf.random_normal([neurons]), name='bias1')
logits1 = tf.matmul(X, W1) + b1
layer1 = tf.nn.softmax(logits1)

W2 = tf.Variable(tf.random_normal([neurons, neurons]), name='weight2')
b2 = tf.Variable(tf.random_normal([neurons]), name='bias3')
logits2 = tf.matmul(layer1, W2) + b2
layer2 = tf.nn.softmax(logits2)

W3 = tf.Variable(tf.random_normal([neurons, neurons]), name='weight3')
b3 = tf.Variable(tf.random_normal([neurons]), name='bias3')
logits3 = tf.matmul(layer2, W3) + b3
layer3 = tf.nn.softmax(logits3)

W4 = tf.Variable(tf.random_normal([neurons, nb_classes]), name='weight4')
b4 = tf.Variable(tf.random_normal([nb_classes]), name='bias4')

logits4 = tf.matmul(layer3, W4) + b4
hypothesis = tf.nn.softmax(logits4)