深度学习实验

实验一 numpy创建全连接神经网络

python 复制代码

import os
import numpy as np
import tensorflow as tf
from tensorflow import keras
from tensorflow.keras import layers, datasets, optimizers

os.environ["TF_CPP_MIN_LOG_LEVEL"] = "2"

# 准备数据
def mnist_dataset():
    (x, y), (x_test, y_test) = datasets.mnist.load_data()
    x = x / 255.0
    x_test = x_test / 255.0
    return (x, y), (x_test, y_test)

# 定义矩阵乘法类
class Matmul():
    def __init__(self):
        self.mem = {}

    def forward(self, x, W):
        h = np.matmul(x, W)
        self.mem = {"x": x, "W": W}
        return h

    def backward(self, grad_y):
        x = self.mem["x"]
        W = self.mem["W"]
        grad_x = np.matmul(grad_y, W.T)
        grad_W = np.matmul(x.T, grad_y)
        return grad_x, grad_W

# 定义Relu类
class Relu():
    def __init__(self):
        self.mem = {}

    def forward(self, x):
        self.mem["x"] = x
        return np.where(x > 0, x, np.zeros_like(x))

    def backward(self, grad_y):
        x = self.mem["x"]
        return (x > 0).astype(np.float32) * grad_y

# 定义Softmax类
class Softmax():
    def __init__(self):
        self.mem = {}
        self.epsilon = 1e-12

    def forward(self, x):
        x_exp = np.exp(x)
        denominator = np.sum(x_exp, axis=1, keepdims=True)
        out = x_exp / (denominator + self.epsilon)
        self.mem["out"] = out
        self.mem["x_exp"] = x_exp
        return out

    def backward(self, grad_y):
        s = self.mem["out"]
        sisj = np.matmul(np.expand_dims(s, axis=2), np.expand_dims(s, axis=1))
        g_y_exp = np.expand_dims(grad_y, axis=1)
        tmp = np.matmul(g_y_exp, sisj)
        tmp = np.squeeze(tmp, axis=1)
        softmax_grad = -tmp + grad_y * s
        return softmax_grad

# 定义交叉熵类
class Cross_entropy():
    def __init__(self):
        self.epsilon = 1e-12
        self.mem = {}

    def forward(self, x, labels):
        log_prob = np.log(x + self.epsilon)
        out = np.mean(np.sum(-log_prob * labels, axis=1))
        self.mem["x"] = x
        return out

    def backward(self, labels):
        x = self.mem["x"]
        return -1 / (x + self.epsilon) * labels

# 建立模型
class myModel():
    def __init__(self):
        self.W1 = np.random.normal(size=[28*28+1, 100])
        self.W2 = np.random.normal(size=[100, 10])
        self.mul_h1 = Matmul()
        self.relu = Relu()
        self.mul_h2 = Matmul()
        self.softmax = Softmax()
        self.cross_en = Cross_entropy()

    def forward(self, x, labels):
        x = x.reshape(-1, 28*28)
        bias = np.ones(shape=[x.shape[0], 1])
        x = np.concatenate([x, bias], axis=1)
        self.h1 = self.mul_h1.forward(x, self.W1)
        self.h1_relu = self.relu.forward(self.h1)
        self.h2 = self.mul_h2.forward(self.h1_relu, self.W2)
        self.h2_soft = self.softmax.forward(self.h2)
        self.loss = self.cross_en.forward(self.h2_soft, labels)

    def backward(self, labels):
        self.loss_grad = self.cross_en.backward(labels)
        self.h2_soft_grad = self.softmax.backward(self.loss_grad)
        self.h2_grad, self.W2_grad = self.mul_h2.backward(self.h2_soft_grad)
        self.h1_relu_grad = self.relu.backward(self.h2_grad)
        self.h1_grad, self.W1_grad = self.mul_h1.backward(self.h1_relu_grad)

# 计算准确率
def compute_accuracy(prob, labels):
    predictions = np.argmax(prob, axis=1)
    truth = np.argmax(labels, axis=1)
    return np.mean(predictions == truth)

# 迭代一个epoch
def train_one_step(model, x, y):
    model.forward(x, y)
    model.backward(y)
    model.W1 -= 1e-5 * model.W1_grad
    model.W2 -= 1e-5 * model.W2_grad
    loss = model.loss
    accuracy = compute_accuracy(model.h2_soft, y)
    return loss, accuracy

# 计算测试集上的loss和准确率
def test(model, x, y):
    model.forward(x, y)
    loss = model.loss
    accuracy = compute_accuracy(model.h2_soft, y)
    return loss, accuracy

# 实际训练
train_data, test_data = mnist_dataset()
train_label = np.zeros(shape=[train_data[0].shape[0], 10])
test_label = np.zeros(shape=[test_data[0].shape[0], 10])
train_label[np.arange(train_data[0].shape[0]), np.array(train_data[1])] = 1
test_label[np.arange(test_data[0].shape[0]), np.array(test_data[1])] = 1

model = myModel()

for epoch in range(50):
    loss, accuracy = train_one_step(model, train_data[0], train_label)
    print(f'epoch {epoch} : loss {loss} ; accuracy {accuracy}')

# 测试
loss, accuracy = test(model, test_data[0], test_label)
print(f'test loss {loss} ; accuracy {accuracy}')

实验二 Pytorch 的CNN

pip install torch==1.12.1+cu102 torchvision==0.13.1+cu102 torchaudio==0.12.1 --extra-index-url https://download.pytorch.org/whl/cu102

python 复制代码

import torch
from torch import nn
import torchvision
from torchvision import datasets, transforms
from tqdm import tqdm

# Hyper parameters
BATCH_SIZE = 100
EPOCHS = 10
LEARNING_RATE = 1e-4
KEEP_PROB_RATE = 0.7

# Set device to use
device = "cuda:0" if torch.cuda.is_available() else "cpu"

# Data transformation
transform = transforms.Compose([
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.5], std=[0.5])
])

# Download and load dataset
path = './data/'
train_data = datasets.MNIST(path, train=True, transform=transform, download=True)
test_data = datasets.MNIST(path, train=False, transform=transform)

# Create DataLoader
train_loader = torch.utils.data.DataLoader(dataset=train_data, batch_size=BATCH_SIZE, shuffle=True)
test_loader = torch.utils.data.DataLoader(dataset=test_data, batch_size=BATCH_SIZE)

# Define the CNN model
class Net(nn.Module):
    def __init__(self):
        super(Net, self).__init__()
        self.model = nn.Sequential(
            nn.Conv2d(in_channels=1, out_channels=32, kernel_size=7, padding=3, stride=1),
            nn.ReLU(),
            nn.MaxPool2d(kernel_size=2, stride=2),
            
            nn.Conv2d(in_channels=32, out_channels=64, kernel_size=5, stride=1, padding=2),
            nn.ReLU(),
            nn.MaxPool2d(kernel_size=2, stride=2),
            
            nn.Flatten(),
            nn.Linear(in_features=7*7*64, out_features=1024),
            nn.ReLU(),
            nn.Dropout(1 - KEEP_PROB_RATE),
            nn.Linear(in_features=1024, out_features=10),
            nn.Softmax(dim=1)
        )

    def forward(self, input):
        output = self.model(input)
        return output

net = Net()
net.to(device)
print(net)

# Define loss function and optimizer
loss_fn = nn.CrossEntropyLoss()
optimizer = torch.optim.Adam(params=net.parameters(), lr=LEARNING_RATE)

# Training and testing process
history = {'Test Loss': [], 'Test Accuracy': []}

for epoch in range(1, EPOCHS + 1):
    process_bar = tqdm(train_loader, unit='step')
    net.train(True)
    
    for step, (train_imgs, labels) in enumerate(process_bar):
        train_imgs = train_imgs.to(device)
        labels = labels.to(device)
        
        # Forward pass
        outputs = net(train_imgs)
        loss = loss_fn(outputs, labels)
        
        # Backward pass and optimization
        net.zero_grad()
        loss.backward()
        optimizer.step()
        
        # Compute accuracy
        predictions = torch.argmax(outputs, dim=1)
        accuracy = torch.sum(predictions == labels) / labels.shape[0]
        
        # Update progress bar
        process_bar.set_description(f"[{epoch}/{EPOCHS}] Loss: {loss.item():.4f}, Acc: {accuracy.item():.4f}")
    
    # Evaluate on test set
    net.train(False)
    correct = 0
    total_loss = 0
    
    with torch.no_grad():
        for test_imgs, labels in test_loader:
            test_imgs = test_imgs.to(device)
            labels = labels.to(device)
            
            outputs = net(test_imgs)
            loss = loss_fn(outputs, labels)
            total_loss += loss
            
            predictions = torch.argmax(outputs, dim=1)
            correct += torch.sum(predictions == labels)
    
    test_accuracy = correct / (BATCH_SIZE * len(test_loader))
    test_loss = total_loss / len(test_loader)
    
    history['Test Loss'].append(test_loss.item())
    history['Test Accuracy'].append(test_accuracy.item())
    
    process_bar.set_description(
        f"[{epoch}/{EPOCHS}] Loss: {loss.item():.4f}, Acc: {accuracy.item():.4f}, Test Loss: {test_loss.item():.4f}, Test Acc: {test_accuracy.item():.4f}"
    )
    process_bar.close()