data_path = '~/data/decompress_mnist'
if __name__ == '__main__':
x_train, x_test, y_train, y_test = mnist.load_mnist(data_path, dst_size=(32, 32), shuffle=True)
x_train = x_train / 255 - 0.5
x_test = x_test / 255 - 0.5
data = {
'X_train': x_train,
'y_train': y_train,
'X_val': x_test,
'y_val': y_test
}
model = models.LeNet5(in_channels=1, out_channels=10, dropout=0.5)
criterion = nn.CrossEntropyLoss()
optimizer = optim.SGD(model.params, lr=1e-3, momentum=0.9, nesterov=True)
stepLR = optim.StepLR(optimizer, 5, gamma=0.1)
solver = pynet.Solver(model, data, criterion, optimizer,
lr_scheduler=stepLR, batch_size=128, num_epochs=10, print_every=1)
solver.train()
plt = Draw()
plt(solver.loss_history)
plt.multi_plot((solver.train_acc_history, solver.val_acc_history), ('train', 'val'),
title='准确率', xlabel='迭代/次', ylabel='准确率', save_path='acc.png')
print('best_train_acc: %f; best_val_acc: %f' % (solver.best_train_acc, solver.best_val_acc))
def nin_train():
x_train, x_test, y_train, y_test = cifar.load_cifar10(data_path,
shuffle=True)
# 标准化
x_train = x_train / 255.0 - 0.5
x_test = x_test / 255.0 - 0.5
net = models.nin(in_channels=3, p_h=p_h)
criterion = nn.CrossEntropyLoss()
accuracy = vision.Accuracy()
loss_list = []
train_list = []
test_list = []
best_train_accuracy = 0.995
best_test_accuracy = 0.995
range_list = np.arange(0, x_train.shape[0] - batch_size, step=batch_size)
for i in range(epochs):
total_loss = 0
num = 0
start = time.time()
for j in range_list:
data = x_train[j:j + batch_size]
labels = y_train[j:j + batch_size]
scores = net(data)
loss = criterion(scores, labels)
total_loss += loss
num += 1
grad_out = criterion.backward()
net.backward(grad_out)
net.update(lr=learning_rate, reg=reg)
end = time.time()
print('one epoch need time: %.3f' % (end - start))
print('epoch: %d loss: %f' % (i + 1, total_loss / num))
loss_list.append(total_loss / num)
if (i % 20) == 19:
# # 每隔20次降低学习率
# learning_rate *= 0.5
train_accuracy = accuracy.compute_v2(x_train,
y_train,
net,
batch_size=batch_size)
test_accuracy = accuracy.compute_v2(x_test,
y_test,
net,
batch_size=batch_size)
train_list.append(train_accuracy)
test_list.append(test_accuracy)
print(loss_list)
print(train_list)
print(test_list)
if train_accuracy > best_train_accuracy and test_accuracy > best_test_accuracy:
path = 'nin-epochs-%d.pkl' % (i + 1)
utils.save_params(net.get_params(), path=path)
break
draw = vision.Draw()
draw(loss_list, xlabel='迭代/20次')
draw.multi_plot((train_list, test_list), ('训练集', '测试集'),
title='精度图',
xlabel='迭代/20次',
ylabel='精度值',
save_path='acc.png')