MAE_ATMO/torch_MAE.ipynb

In [3]:

import torch
import torch.nn as nn
import torch.optim as optim
from torch.utils.data import DataLoader, TensorDataset, random_split
import os
import numpy as np
import matplotlib.pyplot as plt
import cv2

In [1]:

def load_data(pix, use_type='train'):
    datasets = list()
    file_list = [x for x in os.listdir(f"./out_mat/{pix}/{use_type}/") if x.endswith('.npy')]
    for file in file_list:
        file_img = np.load(f"./out_mat/{pix}/{use_type}/{file}")[:,:,:7]
        datasets.append(file_img)
    return np.asarray(datasets)

In [4]:

train_set = load_data(96, 'train')
val_set = load_data(96, 'valid')
test_set = load_data(96, 'test')

In [5]:

def load_mask(mask_rate):
    mask_files = os.listdir(f'./out_mat/96/mask/{mask_rate}')
    masks = list()
    for file in mask_files:
        d = cv2.imread(f'./out_mat/96/mask/{mask_rate}/{file}', cv2.IMREAD_GRAYSCALE)
        d = (d > 0) * 1
        masks.append(d)
    return np.asarray(masks)

In [6]:

masks = load_mask(20)

In [7]:

maxs = train_set.max(axis=0)
mins = train_set.min(axis=0)

In [8]:

len(train_set)

Out[8]:

26749

In [9]:

norm_train = (train_set - mins) / (maxs-mins)

In [10]:

del train_set

In [11]:

norm_valid = (val_set - mins) / (maxs-mins)

In [12]:

del val_set

In [13]:

norm_test = (test_set - mins) / (maxs-mins)

In [14]:

del test_set

In [15]:

norm_train.shape

Out[15]:

(26749, 96, 96)

In [ ]:

trans_train = np.transpose(norm_train, (0, 3, 1, 2))
trans_val = np.transpose(norm_valid, (0, 3, 1, 2))
trans_test = np.transpose(norm_test, (0, 3, 1, 2))

In [16]:

# 可视化特定特征的函数
def visualize_feature(input_feature,masked_feature, output_feature, title):
    plt.figure(figsize=(12, 6))
    plt.subplot(1, 3, 1)
    plt.imshow(input_feature[0].cpu().numpy())
    plt.title(title + " Input")
    plt.subplot(1, 3, 2)
    plt.imshow(masked_feature[0].cpu().numpy())
    plt.title(title + " Masked")
    plt.subplot(1, 3, 3)
    plt.imshow(output_feature[0].detach().cpu().numpy())
    plt.title(title + " Recovery")
    plt.show()

In [ ]:

# 设置随机种子以确保结果的可重复性
torch.manual_seed(0)
np.random.seed(0)

# 数据准备
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print(device)
# 将numpy数组转换为PyTorch张量
tensor_train = torch.tensor(trans_train.astype(np.float32), device=device)
tensor_valid = torch.tensor(trans_val.astype(np.float32), device=device)
tensor_test = torch.tensor(trans_test.astype(np.float32), device=device)

In [ ]:

# 创建一个数据集和数据加载器
train_set = TensorDataset(tensor_train, tensor_train)  # 输出和标签相同，因为我们是自编码器
val_set = TensorDataset(tensor_valid, tensor_valid)
test_set = TensorDataset(tensor_test, tensor_test)
batch_size = 64
train_loader = DataLoader(train_set, batch_size=batch_size, shuffle=True)
val_loader = DataLoader(val_set, batch_size=batch_size, shuffle=False)
test_loader = DataLoader(test_set, batch_size=batch_size, shuffle=False)

In [ ]:

def mask_data(data, device, masks):
    mask_inds = np.random.choice(masks.shape[0], data.shape[0])
    mask = torch.from_numpy(masks[mask_inds]).to(device)
    tmp_first_channel = data[:, 0, :, :] * mask
    masked_data = torch.clone(data)
    masked_data[:, 0, :, :] = tmp_first_channel
    return masked_data

In [ ]:

class SEBlock(nn.Module):
    def __init__(self, in_channels, reduced_dim):
        super(SEBlock, self).__init__()
        self.se = nn.Sequential(
            nn.AdaptiveAvgPool2d(1),  # 全局平均池化
            nn.Conv2d(in_channels, reduced_dim, kernel_size=1),
            nn.ReLU(),
            nn.Conv2d(reduced_dim, in_channels, kernel_size=1),
            nn.Sigmoid()  # 使用Sigmoid是因为我们要对通道进行权重归一化
        )

    def forward(self, x):
        return x * self.se(x)

In [ ]:

# 定义Masked Autoencoder模型
class MaskedAutoencoder(nn.Module):
    def __init__(self):
        super(MaskedAutoencoder, self).__init__()
        self.encoder = nn.Sequential(
            nn.Conv2d(7, 32, kernel_size=3, stride=2, padding=1),
            nn.ReLU(),
            nn.Conv2d(32, 64, kernel_size=3, stride=2, padding=1),
            nn.ReLU(),
            nn.Conv2d(64, 128, kernel_size=3, stride=2, padding=1),
            nn.ReLU(),
            SEBlock(128, 128)
        )
        self.decoder = nn.Sequential(
            nn.ConvTranspose2d(128, 32, kernel_size=3, stride=2, padding=1, output_padding=1),
            nn.ReLU(),
            nn.ConvTranspose2d(32, 16, kernel_size=3, stride=2, padding=1, output_padding=1),
            nn.ReLU(),
            nn.ConvTranspose2d(16, 7, kernel_size=3, stride=2, padding=1, output_padding=1),
            nn.Sigmoid()  # 使用Sigmoid是因为输入数据是0-1之间的
        )

    def forward(self, x):
        encoded = self.encoder(x)
        decoded = self.decoder(encoded)
        return decoded

# 实例化模型、损失函数和优化器
model = MaskedAutoencoder()
criterion = nn.MSELoss()
optimizer = torch.optim.Adam(model.parameters(), lr=0.001)

In [ ]:

# 训练函数
def train_epoch(model, device, data_loader, criterion, optimizer):
    model.train()
    running_loss = 0.0
    for batch_idx, (data, _) in enumerate(data_loader):
        masked_data = mask_data(data, device, masks)
        optimizer.zero_grad()
        reconstructed = model(masked_data)
        loss = criterion(reconstructed, data)
        loss.backward()
        optimizer.step()
        running_loss += loss.item()
    return running_loss / (batch_idx + 1)

In [ ]:

# 评估函数
def evaluate(model, device, data_loader, criterion):
    model.eval()
    running_loss = 0.0
    with torch.no_grad():
        for batch_idx, (data, _) in enumerate(data_loader):
            data = data.to(device)
            masked_data = mask_data(data, device, masks)
            reconstructed = model(masked_data)
            if batch_idx == 8:
                rand_ind = np.random.randint(0, len(data))
                visualize_feature(data[rand_ind], masked_data[rand_ind], reconstructed[rand_ind], title='NO_2')
            loss = criterion(reconstructed, data)
            running_loss += loss.item()
    return running_loss / (batch_idx + 1)

In [ ]:

# 测试函数
def test(model, device, data_loader):
    model.eval()
    with torch.no_grad():
        for batch_idx, (data, _) in enumerate(data_loader):
            data = data.to(device)
            masked_data = mask_data(data, device, masks)
            masked_ind = np.argwhere(masked_data[0][0]==0)
            reconstructed = model(masked_data)
            recon_no2 = reconstructed[0][0]
            ori_no2 = data[0][0]
    return

In [ ]:

model = model.to(device)

num_epochs = 100
train_losses = list()
val_losses = list()
for epoch in range(num_epochs):
    train_loss = train_epoch(model, device, train_loader, criterion, optimizer)
    train_losses.append(train_loss)
    val_loss = evaluate(model, device, val_loader, criterion)
    val_losses.append(val_loss)
    print(f'Epoch {epoch+1}, Train Loss: {train_loss}, Val Loss: {val_loss}')

# 测试模型
test_loss = evaluate(model, device, test_loader, criterion)
print(f'Test Loss: {test_loss}')

In [ ]:

tr_ind = list(range(len(train_losses)))
val_ind = list(range(len(val_losses)))
plt.plot(train_losses, label='train_loss')
plt.plot(val_losses, label='val_loss')
plt.legend(loc='best')

In [ ]:

with torch.no_grad():
    device = 'cpu'
    for batch_idx, (data, _) in enumerate(test_loader):
        model = model.to(device)
        data = data.to(device)
        masked_data = mask_data(data, device, masks)
        reconstructed = model(masked_data)
        tr_maxs = np.transpose(maxs, (2, 0, 1))
        tr_mins = np.transpose(mins, (2, 0, 1))
        rev_data = data * (tr_maxs - tr_mins) + tr_mins
        rev_recon = reconstructed * (tr_maxs - tr_mins) + tr_mins
        data_label = ((rev_data!=0) * (masked_data==0) * rev_data)[:, 0]
        recon_no2 = ((rev_data!=0) * (masked_data==0) * rev_recon)[:, 0]
        break

In [ ]:

real = data_label.flatten()
pred = recon_no2.flatten()

In [ ]:

from sklearn.metrics import mean_squared_error, r2_score, mean_absolute_percentage_error, mean_absolute_error

In [ ]:

mean_squared_error(real, pred)

In [ ]:

mean_absolute_percentage_error(real, pred)

In [ ]:

r2_score(real, pred)

In [ ]:

mean_absolute_error(real, pred)

In [ ]:

visualize_feature(data[5], masked_data[5], reconstructed[5], 'NO2')

In [3]:

len('The total $R^2$ for under 40\% missing data test set was 0.88.')

Out[3]:

62

In [ ]: