self.add(Flatten()) # (64,16,8,8) -> (64,16*8*8=1024)
self.add(Linear(in_features=1024, out_features=256)) #
self.add(BatchNorm2d(256))
self.add(Activation(afn))
self.add(Linear(in_features=256, out_features=10))
self.add(BatchNorm2d(256))
self.add(Activation('softmax'))
if __name__ == "__main__":
id = 'lenet5'
if id == 'lenet5':
dataset = DigitsDataset(norm=True, one_hot=True)
datas = dataset.datas
labels = dataset.labels
# 分隔数据集
train_x, test_x, train_y, test_y = train_test_split(dataset.datas,
dataset.labels,
test_size=0.3,
shuffle=True)
# test_x, val_x, test_y, val_y = train_test_split(test_x,
# test_y,
# test_size=0.3,
# shuffle=True)
train_x = train_x.reshape(-1, 1, 8, 8)
test_x = test_x.reshape(-1, 1, 8, 8)
# val_x = val_x.reshape(-1, 1, 8, 8)
import matplotlib.pyplot as plt
import numpy as np
from dataset.digits_dataset import DigitsDataset
from dataset.multi_class_dataset import MultiClassDataset
from dataset.nonlinear_dataset import NonlinearDataset
from sklearn.model_selection import train_test_split
from core.knn_lib import KNN
from core.kd_tree_lib import KdTree
if __name__ == "__main__":
source = 'multi'
if source == 'digits':
# get dataset
dataset = DigitsDataset(data_type='train')
# get model
knn = KNN(dataset.datas, dataset.labels, k=5)
# get sample
sample_id = 1507
sample, label = dataset[sample_id] # 用第2000个样本做测试
# test and show
pred = knn.predict_single(sample)
print("the sample label is %d, predict is %d" % (label, pred))
plt.figure()
plt.subplot(1, 1, 1)
plt.imshow(sample.reshape(8, 8), cmap='gray')
train_x, test_x, train_y, test_y = train_test_split(dataset.datas,
dataset.labels,
test_size=0.3)
# get model
logi = LogisticReg(train_feats,
train_labels,
lr=0.0001,
n_epoch=100,
batch_size=32)
logi.train()
logi.evaluation(train_feats, train_labels)
# evaluation
logi.evaluation(test_feats, test_labels)
if dataset == 'digits':
# 注意，对于图片等数据，至少需要norm之后再训练，否则loss会爆炸
dataset = DigitsDataset(binary=True, norm=True)
train_feats, test_feats, train_labels, test_labels = train_test_split(
dataset.datas, dataset.labels, test_size=0.3)
# get model
logi = LogisticReg(train_feats,
train_labels,
lr=0.001,
n_epoch=500,
batch_size=-1)
logi.train()
# evaluation
acc = logi.evaluation(test_feats, test_labels)
if dataset == 'breastcancer':
perc.vis_boundary()
if dataset == 'mnist': # 必须特征归一化，同时w必须初始化为0，否则会导致inf问题
# acc = [email protected]/batch8/w0/norm
dataset = MnistDataset(data_type='train', binary=True) # 采用mnist数据集
train_feats, test_feats, train_labels, test_labels = train_test_split(
dataset.datas, dataset.labels, test_size=0.3)
# get model
soft = Perceptron(train_feats, train_labels)
soft.train(lr=0.00001, n_epoch=10, batch_size=8)
# evaluation
acc = soft.evaluation(test_feats, test_labels)
print('acc = %f' % acc)
if dataset == 'digits':
# acc = [email protected]/batch64/w0/norm
dataset = DigitsDataset(data_type='binary')
train_feats, test_feats, train_labels, test_labels = train_test_split(
dataset.datas, dataset.labels, test_size=0.3)
# get model
soft = Perceptron(train_feats, train_labels)
soft.train(lr=0.0001, n_epoch=1000, batch_size=64)
# evaluation
acc = soft.evaluation(test_feats, test_labels)
print('acc = %f' % acc)