Sudoku Solver 3 - Appendix 2 #blog #bernard

predict_digits.py

def simple_network():
	x = tf.placeholder(tf.float32, shape=[None, 784])  # Placeholder for input
	y_ = tf.placeholder(tf.float32, shape=[None, 10])  # Placeholder for true labels (used in training)
	hidden_neurons = 16  # Number of neurons in the hidden layer

	# Hidden layer
	w_1 = weights([784, hidden_neurons])
	b_1 = biases([hidden_neurons])
	h_1 = tf.nn.sigmoid(tf.matmul(x, w_1) + b_1)  # Order of x and w_1 matters here purely syntactically

	# Output layer
	w_2 = weights([hidden_neurons, 10])
	b_2 = biases([10])
	y = tf.matmul(h_1, w_2) + b_2  # Note that we don't use sigmoid here because the next step uses softmax

	return x, y, y_


def predict_digits(test_images, model_path):
	x, y, y_ = simple_network()  # Get our network graph and associated variables

	saver = tf.train.Saver()
	with tf.Session() as sess:
		try:
			saver.restore(sess, model_path)
		except:
			print('Could not load saved model.')
			return False

		prediction = tf.argmax(y, 1)
		out = prediction.eval(feed_dict={x: test_images})  # Feed in our test images
	return out  # Return a prediction for each image