def test_calc_weight_deltas():
learning_rate = .1
i = numpy.matrix('0.4 0.9')
net = [(numpy.matrix('1 1.2; 0.5 0.5'), numpy.matrix('0 0.5')),
(numpy.matrix('0.1; -0.8'), numpy.matrix('-1.3'))]
t = numpy.matrix('0.1')
zs = feed_forward(i, net)
errors = compute_errors(t, zs, net)
deltas = calc_weight_deltas(learning_rate, errors, i, zs, net)
assert deltas[0][1].shape == (1, 2)
assert deltas[1][0].item((0, 0)) == pytest.approx(-.000265, abs=.00001)
assert deltas[1][0].item((1, 0)) == pytest.approx(-.000306, abs=.00001)
assert deltas[1][1].item((0, 0)) == pytest.approx(-0.000379, abs=.00001)
assert deltas[1][1].shape == (1, 1)
new_w2 = numpy.add(net[1][0], deltas[1][0])
assert new_w2.item((0, 0)) == pytest.approx(0.09973, abs=.0001)
def run_softmax(N, image_size, patch_size):
# open training data
print "Training data!"
file_train = "data/pickles/train.pickle"
file_train_labels = "data/pickles/labels_train.pickle"
images_train = helper.unpickle_data(file_train)[:, :, :N]
labels_train = helper.unpickle_data(file_train_labels)[:N]
print "Test data!"
file_test = "data/pickles/test.pickle"
file_test_labels = "data/pickles/labels_test.pickle"
images_test = helper.unpickle_data(file_test)
labels_test = helper.unpickle_data(file_test_labels)
# open saved theta parameters
theta = np.load('weights_learned/weights.out')
# get representations for training data
train_map = model.feed_forward(theta, images_train, patch_size, image_size, images_train.shape[2], 2)
train_features = (train_map['z3'] * train_map['z3_mask']).reshape(image_size ** 2, N)
# get representations for test data
test_map = model.feed_forward(theta, images_test, patch_size, image_size, images_test.shape[2], 2)
test_features = (test_map['z3'] * test_map['z3_mask']).reshape(image_size ** 2, images_test.shape[2])
# print "Check gradients!"
# lambda_ = 0.1
# num_labels = 10
theta = 0.005 * np.random.randn(num_labels * image_size ** 2)
l_cost, l_grad = softmax_cost(theta, num_labels, image_size ** 2, lambda_, train_features, labels_train)
J = lambda x: softmax_cost(x, num_labels, image_size ** 2, lambda_, train_features, labels_train)
# gradient_check.compute_grad(J, theta, l_grad)
# run softmax function
opt_theta, input_size, num_classes = softmax_train(image_size ** 2, num_labels, lambda_, train_features, labels_train, options_)
predictions = softmax_predict((opt_theta, image_size ** 2, num_labels), train_features)
labels_train[labels_train == 10] = 0
print "Accuracy (train): {0:.2f}%".format(100 * np.sum(predictions == labels_train, dtype=np.float64) / labels_train.shape[0])
predictions = softmax_predict((opt_theta, image_size ** 2, num_labels), test_features)
labels_test[labels_test == 10] = 0
print "Accuracy (test): {0:.2f}%".format(100 * np.sum(predictions == labels_test, dtype=np.float64) / labels_test.shape[0])
def test_feed_forward():
i = numpy.matrix('1 2')
net = [
( numpy.matrix('.5 .5; .5 .5'), numpy.matrix('.1 .1') ),
( numpy.matrix('.25 .75; .75 .25'), numpy.matrix('.2 .2') )
]
[ z1, z2 ] = feed_forward(i, net)
assert z2.item((0,0)) == pytest.approx(.737, abs=.001)
assert z2.item((0,1)) == pytest.approx(.737, abs=.001)
def test_another_forward():
i = numpy.matrix('0.4 0.9')
net = [
( numpy.matrix('1 1.2; 0.5 0.5'), numpy.matrix('0 0.5') ),
( numpy.matrix('0.1; -0.8'), numpy.matrix('-1.3') )
]
(z1, z2) = feed_forward(i, net)
assert z2.item((0,0)) == pytest.approx(0.133, abs=.001)
assert z1.item((0,0)) == pytest.approx(0.700, abs=.001)
assert z1.item((0,1)) == pytest.approx(0.807, abs=.001)
def test_compute_errors():
i = numpy.matrix('0.4 0.9')
t = numpy.matrix('0.1')
net = [(numpy.matrix('1 1.2; 0.5 0.5'), numpy.matrix('0 0.5')),
(numpy.matrix('0.1; -0.8'), numpy.matrix('-1.3'))]
zs = feed_forward(i, net)
errors = compute_errors(t, zs, net)
assert errors[1].item((0, 0)) == pytest.approx(-0.0037, abs=.0001)
assert errors[0].item((0, 0)) == pytest.approx(-0.000079, abs=.00001)
assert errors[0].item((0, 1)) == pytest.approx(0.000469, abs=.0001)
def test_calc_errors():
i = numpy.matrix('0.4 0.9')
net = [(numpy.matrix('1 1.2; 0.5 0.5'), numpy.matrix('0 0.5')),
(numpy.matrix('0.1; -0.8'), numpy.matrix('-1.3'))]
t = numpy.matrix('0.1')
(z1, z2) = feed_forward(i, net)
output_error = calc_output_error(t, z2)
hidden_error = calc_hidden_error(output_error, z1, net[1][0])
assert output_error.item((0, 0)) == pytest.approx(-0.0037, abs=.0001)
assert hidden_error.item((0, 0)) == pytest.approx(-0.000079, abs=.00001)
assert hidden_error.item((0, 1)) == pytest.approx(0.000469, abs=.0001)
def write_mesh_output(weights_path, image_path, out_mesh_path):
Ws = {}
w_file = h5py.File(weights_path, 'r')
for name in model.weight_names:
Ws[name] = tf.constant(w_file[name])
image = Image.open(image_path)
input_image = np.empty([1, 128, 128, 3])
input_image[0] = np.array(image.resize((128, 128)), dtype=np.float32) / 255
X = tf.constant(input_image, dtype='float32')
ff = model.feed_forward(X, Ws)
sess = tf.Session()
sess.run(tf.global_variables_initializer())
depth = np.array(ff.eval(session=sess))
depth_image = depth[0, :, :, 0]
scaled_image = get_scaled_image(image)
scaled_width = scaled_image.width
scaled_height = scaled_image.height
resized_depth = resize_depth(depth_image, scaled_width, scaled_height)
plt.imshow(resized_depth, cmap=plt.get_cmap('gray'))
plt.show()
vertices, faces = create_mesh(resized_depth)
mesh = o3d.geometry.TriangleMesh()
mesh.vertices = o3d.utility.Vector3dVector(vertices)
mesh.triangles = o3d.utility.Vector3iVector(faces.astype(np.int32))
mesh.vertex_colors = o3d.utility.Vector3dVector(
np.reshape(
np.array(scaled_image, dtype=np.float32) / 255,
(scaled_width * scaled_height, 3)))
o3d.io.write_triangle_mesh(out_mesh_path, mesh)