def main():
x_train, y_train = load_mnist()
x_test, y_test = load_mnist(kind='t10k')
# normalize
x_train = (x_train / 255).astype(np.float32)
x_test = (x_test / 255).astype(np.float32)
# hyper parameters
model_prefix = 'weights/LDA'
if not os.path.isdir(model_prefix):
os.makedirs(model_prefix)
weights = []
biases = []
# train each single digit
for j in range(10):
print('Training for digit %d' % j)
weight_file = os.path.join(model_prefix, '%d.npy' % j)
# deal with soft label
target_digit = j
label_test = (y_test == target_digit).astype(np.uint8)
weight = cal_LDA_weight(target_digit, x_train, y_train)
bias = cal_LDA_bias(target_digit, x_train, y_train, weight, lamb=0.6)
accuracy, recall = test(x_test, label_test, weight, bias)
# print('acc', accuracy)
print('rec', recall)
weights.append(weight)
biases.append(bias)
np.save(weight_file, weight)
final_accuracy = multi_test(x_test, y_test, weights, biases)
print('Final accuracy:', final_accuracy)
def main():
x_train, y_train = load_mnist()
x_test, y_test = load_mnist(kind='t10k')
# normalize
x_train = (x_train / 255).astype(np.float32)
x_test = (x_test / 255).astype(np.float32)
# hyper parameters
lr = 0.0001
regular = 0.5
epoch = 50
print_freq = 10
batch_size = 60000
model_prefix = 'weights/logistic_regression_ridge_loss'
if not os.path.isdir(model_prefix):
os.makedirs(model_prefix)
weights = []
# train each single digit
for j in range(10):
print('Training for digit %d' % j)
weight_file = os.path.join(model_prefix, '%d.npy' % j)
weight = np.zeros(PIXELS)
best_accuracy = 0
best_weight = np.zeros(PIXELS)
# deal with soft label, label is +1 or -1
target_digit = j
label_train = (y_train == target_digit).astype(np.int32)
label_test = (y_test == target_digit).astype(np.int32)
label_train = label_train * 2 - 1
label_test = label_test * 2 - 1
for i in range(epoch):
for batch in range(TRAIN_SAMPLES // batch_size):
weight = train(x_train[batch * batch_size:],
label_train[batch * batch_size:],
weight,
regular,
lr,
i,
num_samples=batch_size)
accuracy = test(x_test, label_test, weight, i)
if accuracy > best_accuracy:
best_accuracy = accuracy
best_weight = weight
if i > 0 and i % print_freq == 0:
print('Epoch[%d/%d]' % (i, epoch),
'current accuracy: %.4f' % accuracy)
weights.append(best_weight)
np.save(weight_file, best_weight)
print('Finish training for digit %d, best accuracy %.4f\n' %
(j, best_accuracy))
final_accuracy = multi_test(x_test, y_test, weights)
print('Final accuracy:', final_accuracy)
def main():
x_train, y_train = load_mnist()
x_test, y_test = load_mnist(kind='t10k')
# normalize
x_train = (x_train / 255).astype(np.float32)
x_test = (x_test / 255).astype(np.float32)
# hyper parameters
lr = 0.0001
epoch = 50
print_freq = 10
model_prefix = 'weights/logistic_regression'
if not os.path.isdir(model_prefix):
os.makedirs(model_prefix)
weights = []
# train each single digit
for j in range(10):
print('Training for digit %d' % j)
weight_file = os.path.join(model_prefix, '%d.npy' % j)
weight = np.zeros(PIXELS)
best_accuracy = 0
best_weight = np.zeros(PIXELS)
# deal with soft label
target_digit = j
label_train = (y_train == target_digit).astype(np.uint8)
label_test = (y_test == target_digit).astype(np.uint8)
for i in range(epoch):
weight = train(x_train, label_train, weight, lr, i)
accuracy = test(x_test, label_test, weight, i)
if accuracy > best_accuracy:
best_accuracy = accuracy
best_weight = weight
if i > 0 and i % print_freq == 0:
print('Epoch[%d/%d]' % (i, epoch),
'current accuracy: %.4f' % accuracy)
weights.append(best_weight)
np.save(weight_file, best_weight)
print('Finish training for digit %d, best accuracy %.4f\n' %
(j, best_accuracy))
final_accuracy = multi_test(x_test, y_test, weights)
print('Final accuracy:', final_accuracy)
def __init__(self, cfg, phase):
self.num_category = 10
self.cfg = cfg
self.phase = phase # train or test
if phase == 'test':
phase = 't10k'
self.images, self.labels = load_mnist(kind=phase)
self.resolution = cfg.resolution
self.transforms = transforms.ToTensor()
self.db = self._get_db()
logger.info('=> Loading {} images from {}'.format(
self.phase, self.cfg.root))
logger.info('=> num_images: {}'.format(len(self.db['x'])))
if not category_id == target_id:
continue
raw_image = self.input[i]
for j in range(self.fmaps.shape[1]):
fmap = self.fmaps[i][j]
grad = self.grads[category_id][j]
gcam += fmap * grad
gcam -= gcam.min()
if (gcam.max() != 0):
gcam /= gcam.max()
gcam = cv2.resize(gcam,
(self.init_image_size, self.init_image_size))
self.save(gcam, raw_image, i, category_id)
print('Group %d finish!' % self.group)
if __name__ == "__main__":
x, y = load_mnist(kind='t10k')
x_4 = x[y == 4]
y_4 = y[y == 4]
x_9 = x[y == 9]
y_9 = y[y == 9]
x_group = np.concatenate((x_4, x_9), axis=0)
y_group = np.concatenate((y_4, y_9), axis=0)
# print(y_group)
tSNE(x_group, y_group.astype(np.uint8), [4, 9])
import numpy as np
from MnistData import load_mnist
import matplotlib.pyplot as plt
import random
import copy
imgs, _ = load_mnist()
imgs = imgs[:15]
fig, ax = plt.subplots(
nrows=3,
ncols=5,
sharex=True,
sharey=True,
)
ax = ax.flatten()
for k in range(15):
img = imgs[k].reshape(28, 28, 1)
buffer = []
block_num = 4
block_size = 28 // block_num
for i in range(block_num):
for j in range(block_num):
block = copy.deepcopy(img[i * block_size:(i + 1) * block_size,
j * block_size:(j + 1) * block_size])
buffer.append(block)
random.shuffle(buffer)
# img = np.zeros((28,28))
for i in range(block_num):
for j in range(block_num):
from sklearn import svm
import numpy as np
from MnistData import load_mnist
x_train, y_train = load_mnist()
x_test, y_test = load_mnist(kind='t10k')
train_num = 6000
test_num = 1000
predictor = svm.SVC(gamma='scale',
C=1.0,
decision_function_shape='ovr',
kernel='sigmoid')
predictor.fit(x_train[:train_num], y_train[:train_num])
result = predictor.predict(x_test[:test_num])
accurancy = np.sum(np.equal(result, y_test[:test_num])) / test_num
print(accurancy)