class GPR_Class:
def __init__(self):
self.logger = logging.getLogger('GPR Classfication')
self.gp_package = 'sklearn'
self.verbose = False
self.verbose_debug = False
self.gp = GaussianProcessRegressor()
self.data = []
self.x_min = []
self.x_max = []
self.y_min = []
self.y_max = []
self.X = []
self.Y = []
self.label_array = np.array([])
self.label_array_train = np.array([])
self.label_array_test = np.array([])
self.label_array_eval = np.array([])
self.images_array = np.array([])
self.images_array_train = np.array([])
self.images_array_test = np.array([])
self.images_array_eval = np.array([])
self.indices_distributed = np.array([])
self.train_indices = np.array([])
self.test_indices = np.array([])
self.evaluation_indices = np.array([])
self.features_train = np.array([])
self.features_test = np.array([])
self.features_eval = np.array([])
# Number of inducing point for gyptorch SGPR
self.n_inducing_points = 100
@staticmethod
def normalize(x: List[float], min_in: float, max_in: float, min_out: float,
max_out: float) -> list:
x_norm = ((x - min_in) / (max_in - min_in) * (max_out - min_out) +
min_out)
return x_norm
def create_GPR_model(self, X, Y):
self.x_min = np.amin(X, axis=0)
self.x_max = np.amax(X, axis=0)
self.y_min = np.amin(Y, axis=0)
self.y_max = np.amax(Y, axis=0)
assert np.min(self.y_min) < np.max(
self.y_max), 'y_min should be smaller than y_max'
for i in range(len(self.x_min)):
if self.x_min[i] == self.x_max[i]:
if abs(self.x_max[i]) > 0:
add_value = min(1e-1, abs(self.x_max[i] * 1e-2))
else:
add_value = 1e-1
self.x_max[i] += add_value
# self.logger.warning('Added ' + str(add_value) + ' to maximum of variable ' +
# str(i) + ' as min and max was equal')
assert np.min(self.x_min) < np.max(
self.x_max), 'x_min should be smaller than x_max'
self.X = X
self.Y = Y
t1 = time.time()
x_norm = self.normalize(X,
min_in=self.x_min,
max_in=self.x_max,
min_out=0,
max_out=1)
t2 = time.time()
if self.verbose_debug:
print('normalizing X took : ' + str(t2 - t1) + 's')
t1 = time.time()
y_norm = self.normalize(Y,
min_in=self.y_min,
max_in=self.y_max,
min_out=0,
max_out=1)
t2 = time.time()
if self.verbose_debug:
print('normalizing Y took : ' + str(t2 - t1) + 's')
kernel = 1.0 * Matern(length_scale=0.1, length_scale_bounds=(1e-5, 1e5), nu=2.5) \
+ WhiteKernel()
# kernel = 1.0 * RBF()
t1 = time.time()
if self.gp_package == 'sklearn':
self.gp = GaussianProcessRegressor(kernel=kernel,
alpha=0.0).fit(x_norm, y_norm)
elif self.gp_package == 'gp_flow':
self.gp = GaussianProcessRegressor(kernel=kernel,
alpha=0.0).fit(x_norm, y_norm)
self.gp = gpflow.models.GPR(
x_norm,
y_norm,
kern=gpflow.kernels.Matern52(input_dim=x_norm.shape[1]) +
gpflow.kernels.Linear(input_dim=x_norm.shape[1]) +
gpflow.kernels.White(input_dim=x_norm.shape[1]))
self.gp.compile()
opt = gpflow.train.ScipyOptimizer()
opt.minimize(self.gp)
# elif self.gp_package == 'gpytorch':
# num_tasks = len(y_norm[0])
# likelihood = gpytorch.likelihoods.GaussianLikelihood(num_tasks=num_tasks).cuda()
# train_x = torch.tensor(x_norm)
# # Add Extra dimension at position zero that represents the different inputs
# train_x = train_x.unsqueeze(0).repeat(num_tasks, 1, 1)
#
# train_x_cuda = train_x.cuda()
# train_y_cuda = torch.tensor(y_norm).transpose(dim0=-2, dim1=-1).cuda()
#
# self.gp = gp_torch_sgpr.gp_torch_sgpr(
# train_x=train_x_cuda,
# train_y=train_y_cuda,
# likelihood=likelihood,
# verbose=self.verbose,
# kernel='Matern_1_5',
# n_inducing_points=self.n_inducing_points
# ).cuda()
# self.gp.float()
# # self.gp.train_gp_model(train_x=train_x_cuda, train_y=train_y_cuda)
# self.gp.train_gp_model_adam(train_x=train_x_cuda, train_y=train_y_cuda)
else:
raise AssertionError('unknown GP Method: ' + self.gp_package)
t2 = time.time()
if self.verbose_debug:
print('training GPR : ' + str(t2 - t1) + 's')
return self.gp
def predict(self, x_test, do_normlization: bool = False):
x_interpolation = x_test.copy()
if len(x_test.shape) == 1:
x_interpolation = x_interpolation[np.newaxis, :]
n_test_points = x_interpolation.shape[0]
for i in range(n_test_points):
x_interpolation[i] = self.normalize(x_interpolation[i],
min_in=self.x_min,
max_in=self.x_max,
min_out=0,
max_out=1)
if self.gp_package == 'sklearn':
y_mean_interpol, y_std_norm = self.gp.predict(x_interpolation,
return_std=True)
elif self.gp_package == 'gp_flow':
y_mean_interpol, y_var_norm = self.gp.predict_y(x_interpolation)
y_std_norm = y_var_norm**0.5
# elif self.gp_package == 'gpytorch':
# x_interpolation_cuda = torch.tensor(x_interpolation).cuda()
# y_mean_interpol, y_std_norm = self.gp.make_prediction(x_interpolation_cuda)
else:
raise AssertionError('unknown GP Method: ' + self.gp_package)
y_mean = y_mean_interpol
y_std = y_std_norm
for i in range(n_test_points):
y_mean[i] = self.normalize(y_mean[i],
min_in=0,
max_in=1,
min_out=self.y_min,
max_out=self.y_max)
if isinstance(self.y_max, np.float):
y_std[i] = y_std_norm[i] * ((self.y_max - self.y_min)**2)
else:
y_std[i] = y_std_norm[i] * (
(max(self.y_max) - min(self.y_min))**2)
return y_mean, y_std
# def plot_1D_interpolation(self):
# x_test = np.linspace(0, 1, 1000)
# x_interpolation = self.normalize(x_test,
# min_in=0, max_in=1,
# min_out=self.x_min, max_out=self.x_max)
# # Rescale and perform GPR prediction
# y_mean, y_cov = self.predict(x_interpolation)
#
# plt.figure()
# plt.plot(x_interpolation, y_mean, 'k', lw=3, zorder=9)
# plt.fill_between(x_interpolation, (y_mean - np.sqrt(np.diag(y_cov))).transpose(),
# (y_mean + np.sqrt(np.diag(y_cov))).transpose(),
# alpha=0.5, color='k')
# plt.scatter(self.X[:, 0], self.Y,
# c='r', s=1, zorder=10, edgecolors=(0, 0, 0))
# plt.tight_layout()
# plt.show()
@staticmethod
def read_label(filename, label_choices, output_type: str = 'binary'):
basename = os.path.basename(filename)
labels: bytes = basename.split('_')[0]
indices = [label_choices.index(l) for l in labels]
if output_type == 'binary':
data = [[0 for i in range(len(label_choices))]
for j in range(len(labels))]
for iL in range(0, len(labels)):
data[iL][indices[iL]] = 1
else:
data = indices
return data
def read_labels(self,
dir_name,
ext='.png',
label_choices='0',
output_type: str = 'binary'):
fd = [
os.path.join(dir_name, fn) for fn in os.listdir(dir_name)
if fn.endswith(ext)
]
label_array = np.array(
[self.read_label(ifd, label_choices, output_type) for ifd in fd])
n_examples = label_array.shape[0]
n_labels = label_array.shape[1]
if output_type == 'binary':
n_label_choices = label_array.shape[2]
else:
n_label_choices = 1
label_array = label_array.reshape(
[n_examples, n_label_choices * n_labels])
return label_array
@staticmethod
def read_images(dir_name, ext='.png', is_single_image=False):
if is_single_image:
im_raw = [Image.open(dir_name).convert('L')]
else:
fd = [
os.path.join(dir_name, fn) for fn in os.listdir(dir_name)
if fn.endswith(ext)
]
im_raw = [Image.open(iFile).convert('L') for iFile in fd]
image_data = [np.asarray(i) for i in im_raw]
image_data = [i / 255 for i in image_data]
# binary_image_array = [(iData > 125.5) * 1.0 for iData in image_data]
# image_data/255
return image_data
@staticmethod
def generate_patch_aggregation(images_array, scaling_factor=[2, 2]):
features_list = \
[
[
np.sum(iPatch)
for iPatch
in image.extract_patches_2d(image=iImage,
patch_size=(round(iImage.shape[0] / scaling_factor[0]),
round(iImage.shape[1] / scaling_factor[1])))
]
for iImage in images_array
]
return np.array(features_list)
@staticmethod
def convolution_step_edges_max_pooling(im_array,
windows_size: tuple,
weights: np.array = np.array([]),
weight_shape: list = []):
if len(weights) > 0:
if len(weight_shape) == 0:
raise AssertionError('weight_shape is empty')
weight_matrix = weights.reshape(weight_shape)
else:
weight_matrix = [[1, -1], [-1, 1]]
im_array_convolve = scipy.signal.convolve(in1=im_array,
in2=weight_matrix,
mode="valid")
im_array_convolve_max = im_array_convolve * (
im_array_convolve == scipy.ndimage.filters.maximum_filter(
input=im_array_convolve, size=windows_size))
im_array_convolve_max_reduce = measure.block_reduce(
im_array_convolve_max, windows_size, np.max)
return im_array_convolve_max_reduce
@staticmethod
def generate_indices_train_test_evaluation(label_array,
ratio: tuple = (0.8, 0.1, 0.1)):
assert (max(ratio) <= 1) & (
min(ratio) >= 0), 'Each entry of ratio must be between 0.0 and 1.0'
unique_label = len(label_array[0])
indices = [None] * unique_label
n_data = len(label_array)
for i in range(n_data):
idx = label_array[i].argmax()
if indices[idx] is None:
indices[idx] = []
indices[idx] += [i]
np.random.seed(0)
indices_distributed = [None] * unique_label
train_indices = []
test_indices = []
evaluation_indices = []
for i in range(unique_label):
n_data_partial = len(indices[i])
indices_rand = np.random.permutation(n_data_partial)
n_train = round(ratio[0] * n_data_partial)
n_test = round((n_data_partial - n_train) / 2)
n_evaluation = n_data_partial - n_train - n_test
# 0:X does not include X
n_test_actual = round(n_test * ratio[1])
n_evaluation_actual = round(n_evaluation * ratio[2])
indices_distributed[i] = [
itemgetter(*indices_rand[0:n_train])(indices[i]),
itemgetter(*indices_rand[n_train:n_train + n_test_actual])(
indices[i]),
itemgetter(*indices_rand[n_train + n_test:n_train + n_test +
n_evaluation_actual])(indices[i])
]
# if (len(indices_distributed[i][0]) + len(indices_distributed[i][1]) + len(indices_distributed[i][2])) \
# != n_data_partial:
# print('stop')
train_indices += indices_distributed[i][0]
test_indices += indices_distributed[i][1]
evaluation_indices += indices_distributed[i][2]
return indices_distributed, train_indices, test_indices, evaluation_indices
def generate_train_test_eval_indices(self, ratio_train_test_eval: tuple):
self.indices_distributed, self.train_indices, self.test_indices, self.evaluation_indices = \
self.generate_indices_train_test_evaluation(self.label_array, ratio=ratio_train_test_eval)
def load_data(self, dir_name: str, label_choices: str,
ratio_train_test_eval: tuple):
self.label_array = self.read_labels(dir_name=dir_name,
label_choices=label_choices,
output_type='binary')
self.images_array = self.read_images(dir_name=dir_name)
self.generate_train_test_eval_indices(ratio_train_test_eval)
self.images_array_train = itemgetter(*self.train_indices)(
self.images_array)
self.images_array_test = itemgetter(*self.test_indices)(
self.images_array)
self.images_array_eval = itemgetter(*self.evaluation_indices)(
self.images_array)
self.label_array_train = itemgetter(*self.train_indices)(
self.label_array)
self.label_array_test = itemgetter(*self.test_indices)(
self.label_array)
self.label_array_eval = itemgetter(*self.evaluation_indices)(
self.label_array)
def generate_test_train_eval_data(self,
windows_size: list,
weights: np.ndarray = np.ndarray([]),
weight_shape: list = [[]]):
def generate_data(data_set_type: str = 'train'):
if data_set_type == 'train':
data_set = self.images_array_train
elif data_set_type == 'test':
data_set = self.images_array_test
elif data_set_type == 'eval':
data_set = self.images_array_eval
else:
raise ValueError('Unknown data_set_type: ' + data_set_type)
data_out = np.array([
list(
itertools.chain(*[
self.convolution_step_edges_max_pooling(
i,
windows_size=windows_size[iW],
weights=weights[iW],
weight_shape=weight_shape[iW]).flatten().tolist()
for iW in range(len(windows_size))
])) for i in data_set
])
# try:
# data_out = data_out.reshape(data_out.shape[0:2])
# except:
# pass
return data_out
self.features_train = generate_data('train')
self.features_test = generate_data('test')
self.features_eval = generate_data('eval')
def train_and_test(self,
windows_size: list,
weights=np.array([]),
weight_shape: list = []):
assert len(weights.shape) == 2, 'weights need to be a 2d-Array'
assert type(weight_shape) == list, 'weight_shape need to be a list'
# reset to integer
windows_size = [
tuple([int(round(i)) for i in iW]) for iW in windows_size
]
t1 = time.time()
self.generate_test_train_eval_data(windows_size=windows_size,
weights=weights,
weight_shape=weight_shape)
t2 = time.time()
if self.verbose_debug:
print('generating data took : ' + str(t2 - t1) + 's')
# print('windows_size: ' + str(windows_size) + '- weights: ', weights)
t1 = time.time()
self.create_GPR_model(self.features_train, self.label_array_train)
t2 = time.time()
if self.verbose_debug:
print('generating GPR model took : ' + str(t2 - t1) + 's')
t1 = time.time()
pred_vec = self.predict(np.array(self.features_test))
predicted_output = [
pred_vec[0][i].argmax() for i in range(len(self.features_test))
]
t2 = time.time()
if self.verbose_debug:
print('test prediction took : ' + str(t2 - t1) + 's')
t1 = time.time()
test_accuracy = len([
predicted_output[i] for i in range(len(predicted_output))
if predicted_output[i] == self.label_array_test[i].argmax()
]) / len(predicted_output)
t2 = time.time()
if self.verbose_debug:
print('test accuracy took : ' + str(t2 - t1) + 's')
# test_accuracy = len([predicted_output[i] for i in range(len(predicted_output)) if
# predicted_output[i] == (itemgetter(*self.test_indices)(self.label_array))[i].argmax()
# ])/len(predicted_output)
if self.verbose:
pred_vec = self.predict(np.array(self.features_train))
predicted_output = [
pred_vec[0][i].argmax()
for i in range(len(self.features_train))
]
train_accuracy = len([
predicted_output[i] for i in range(len(predicted_output))
if predicted_output[i] == self.label_array_train[i].argmax()
]) / len(predicted_output)
pred_vec = self.predict(np.array(self.features_eval))
predicted_output = [
pred_vec[0][i].argmax() for i in range(len(self.features_eval))
]
eval_accuracy = len([
predicted_output[i] for i in range(len(predicted_output))
if predicted_output[i] == self.label_array_eval[i].argmax()
]) / len(predicted_output)
print('Train accuracy: ' + str(train_accuracy) +
' - Test accuracy: ' + str(test_accuracy) +
' - Evaluation Accuracy: ' + str(eval_accuracy))
return test_accuracy
