def main():
num_node = 28 * 28
bm = BoltzmanMachine(num_node, num_hidden=100, lr=0.001)
images = load_mnist(normalize=False)["train_images"][:10]
# images = (images > 0.5).astype(np.int)
num_iter = 1000
for i in range(num_iter):
print(i)
target = np.random.choice(len(images))
x_sampled = images[target]
bm.update(x_sampled)
#noised = noise(images, 0.1)
noised = images.copy()
fig = plt.figure()
H = np.floor(np.sqrt(len(noised)))
W = 2 * np.ceil(np.sqrt(len(noised)))
for i in range(len(noised)):
ax = fig.add_subplot(H, W, 2 * i + 1)
ax.imshow(noised[i].reshape(28, 28))
mean = bm.gibss_xhx(noised[i], 1000)
ax = fig.add_subplot(H, W, 2 * i + 2)
ax.imshow(mean.reshape(28, 28))
plt.show()
10分类 采用OvR的方法来训练
fisher new
"""
import torch as tc
import loadmnist as lm
import linage_classify
#设置cuda device
if tc.cuda.is_available():
device = tc.device('cuda', 0)
else:
device = tc.device('cpu')
#载入MNIST数据集
img, label = lm.load_mnist("./data", "train", 'tc')
test, tl = lm.load_mnist("./data", "t10k", 'tc')
#数据转为cuda对象
img = img.to(device)
#把MNIST数据集的标签分类10个二分类器,0为正类,1为负类
filter_vec = tc.tensor(range(10))
filter_mat = filter_vec.repeat(label.size()[0], 1).byte()
label_mat = tc.clamp(label.repeat(10, 1) - filter_mat.t(), 0, 1).cuda()
#初始化线性分类器,fisher分类器
lcm = linage_classify.FisherClassify()
##lcm.cuda()##使用cuda
#训练所有的10个分类器并将结果存入y_hat中
import tensorflow as tf
import loadmnist as lm
tx,ty = lm.load_mnist('C:\\Users\\Zy\\Desktop\\MNIST','t10k')
saver = tf.train.import_meta_graph('saves/classify3-40000.meta')
graph = tf.get_default_graph()
def Test(input_x,input_y):
global graph
if len(input_x) == len(input_y) and len(input_x) > 0:
x = graph.get_tensor_by_name('x:0')
y = graph.get_tensor_by_name('y:0')
keep_prob = graph.get_tensor_by_name('keep_prob:0')
accuracy = graph.get_tensor_by_name('accuracy:0')
pred_pos = graph.get_tensor_by_name('pred_pos:0')
with tf.Session() as sess:
saver.restore(sess, tf.train.latest_checkpoint('saves/'))
if len(input_x) == 1:
rs = sess.run(pred_pos, feed_dict={x: input_x, y: input_y, keep_prob: 1})
print(input_y)
print('result : %d' % (rs[0]+1))
else:
rs = sess.run(accuracy, feed_dict={x: input_x, y: input_y, keep_prob: 1})
print('data(%d) : ac is %f'%(len(input_x),rs*100))
else:
print('wrong format you input')