def check_image(letter):
letter = np.asarray(letter.convert('LA'))[:, :, -1] / 255
conv_layer1 = ConvolutionLayer.init_with_weights_from_file(
'resources/weights_kernels_t0', 5, 2)
conv_layer2 = ConvolutionLayer.init_with_weights_from_file(
'resources/weights_kernels_t1', 5, 2)
network_conv = Network([
Layer.init_with_weights(
read_matrix_from_file('resources/weights_conv_t' + str(i)),
sigmoid, sigmoid_prime) for i in range(2)
], target_func_for_tests, j_cross_entropy_derivative)
network = ConvolutionNetwork([conv_layer1, conv_layer2], network_conv)
res_matrix = network.get_result_matrix(np.array([[letter]]))
return np.argmax(res_matrix.flatten())
def run_simple_net():
network = Network([
Layer.init_with_weights(
read_matrix_from_file('resources/weights' + str(i)), sigmoid,
sigmoid_prime) for i in range(2)
], j_cross_entropy, j_cross_entropy_derivative)
# data = transform_input_imgs_to_data(read_char_images_from_dir('resources/test', False, 36, 27))
data, answers = read_all_char_examples_with_answers(
'resources/train', False)
data = transform_input_imgs_to_data(data)
print(network.sgd(data, answers, 110, 10, 10, eps=1e-7, visualize=True))
for i, l in enumerate(network.layers):
save_matrix_to_file('resources/weights' + str(i), l.get_weights())
print(network.get_result_matrix(data))
print(network.process_input(data, answers))
def run_conv_net():
conv_layer1 = ConvolutionLayer.init_with_weights_from_file(
'resources/weights_kernels0', 5, 2)
conv_layer2 = ConvolutionLayer.init_with_weights_from_file(
'resources/weights_kernels1', 5, 2)
# conv_layer1 = ConvolutionLayer.init_random_convolution_layer(10, 5, 2)
# conv_layer2 = ConvolutionLayer.init_random_convolution_layer(20, 5, 2)
layer1 = Layer.layer_with_random_weights(50, 320, sigmoid, sigmoid_prime)
layer2 = Layer.layer_with_random_weights(10, 50, sigmoid, sigmoid_prime)
# network_conv = Network([layer1, layer2], j_cross_entropy, j_cross_entropy_derivative)
network_conv = Network([
Layer.init_with_weights(
read_matrix_from_file('resources/weights_conv' + str(i)), sigmoid,
sigmoid_prime) for i in range(2)
], j_cross_entropy, j_cross_entropy_derivative)
network = ConvolutionNetwork([conv_layer1, conv_layer2], network_conv)
images, answers = get_example_batch(2000)
network.sgd(images, answers, 200, 1, 1, 200, 1e-9, visualize=False)
# for i, l in enumerate(network.fully_connected_net.layers):
# save_matrix_to_file('resources/weights_conv' + str(i), l.get_weights())
#
# for i, l in enumerate(network.layers):
# save_matrix_to_file('resources/weights_kernels' + str(i), map(lambda x: x.flatten(), l.get_weights()))
for i, l in enumerate(network.fully_connected_net.layers):
save_matrix_to_file('resources/weights_conv_t' + str(i),
l.get_weights())
for i, l in enumerate(network.layers):
save_matrix_to_file('resources/weights_kernels_t' + str(i),
map(lambda x: x.flatten(), l.get_weights()))
print(network.process_input(images, answers))
def run_on_test_data():
conv_layer1 = ConvolutionLayer.init_with_weights_from_file(
'resources/weights_kernels_t0', 5, 2)
conv_layer2 = ConvolutionLayer.init_with_weights_from_file(
'resources/weights_kernels_t1', 5, 2)
network_conv = Network([
Layer.init_with_weights(
read_matrix_from_file('resources/weights_conv_t' + str(i)),
sigmoid, sigmoid_prime) for i in range(2)
], target_func_for_tests, j_cross_entropy_derivative)
network = ConvolutionNetwork([conv_layer1, conv_layer2], network_conv)
mndata = MNIST('resources/train/mnist')
images, labels = mndata.load_testing()
images = np.array(images).reshape(len(images), 1, 28, 28) / 255
answers = transform_labels_to_vectors(labels)
print("\nQuality is {:.2f}%\n".format(
network.process_input(images[:1000], answers[:1000]) * 100 / 1000))
test_data = cv2.imread("/home/saul/Documents/2.png",
cv2.IMREAD_GRAYSCALE) / 255
network.process_input(np.array([[abs(1 - test_data)]]),
np.array([[0, 0, 1, 0, 0, 0, 0, 0, 0, 0]]))