如何在 Pytorch Lightning 中使用 numpy 数据集

Question

我想使用NumPy 制作一个数据集，然后想训练和测试一个简单的模型，例如“线性或逻辑”。

我正在努力学习Pytorch Lightning。我找到了一个tutorial，我们可以使用 NumPy 数据集并且可以使用均匀分布here。作为一个新人，我没有完全理解，我该怎么做！

我的代码如下

import numpy as np 
import pytorch_lightning as pl 
from torch.utils.data import random_split, DataLoader, TensorDataset
import torch
from torch.autograd import Variable
from torchvision import transforms

np.random.seed(42)

device = 'cuda' if torch.cuda.is_available() else 'cpu'

class DataModuleClass(pl.LightningDataModule):
    def __init__(self):
        super().__init__()
        self.constant = 2
        self.batch_size = 10
        self.transform = transforms.Compose([
            transforms.ToTensor()
        ])
        
    def prepare_data(self):
        a = np.random.uniform(0, 500, 500)
        b = np.random.normal(0, self.constant, len(x))

        c = a + b
        X = np.transpose(np.array([a, b]))

        idx = np.arange(500)
        np.random.shuffle(idx)
        
        # Uses foirst 400 random indices for training
        train_idx = idx[:400]
        # Uses the remaining indices for validation
        val_idx = idx[400:]
        
        # Generate train and validation dataset
        x_train, y_train = X[train_idx], y[train_idx]
        x_val, y_val = X[val_idx], y[val_idx]
        
        # Converting numpy array to Tensor
        self.x_train_tensor = torch.from_numpy(x_train).float().to(device)
        self.y_train_tensor = torch.from_numpy(y_train).float().to(device)
        
        self.x_val_tensor = torch.from_numpy(x_val).float().to(device)
        self.y_val_tensor = torch.from_numpy(y_val).float().to(device)
        
        training_dataset = TensorDataset(self.x_train_tensor, self.y_train_tensor)
        
        validation_dataset = TensorDataset(self.x_val_tensor, self.y_val_tensor)

        return training_dataset, validation_dataset
        
    def train_dataloader(self):
        training_dataloader = prepare_data() # Most probably this is wrong way!!!
        return DataLoader(self.training_dataloader)

    def val_dataloader(self):
        validation_dataloader = prepare_data() # Most probably this is wrong way!!!
        return DataLoader(self.validation_dataloader)
        
    # def test_dataloader(self):
        
obj = DataModuleClass()
print(obj.prepare_data())  

这部分是根据给出的答案完成的[这里，我想把a and b as features和c当作label or target variable。]

现在，如何将数据集传递给`训练和验证方法？

Answer 1

您可以使用以下代码从prepare_data() 或setup() 获取数据。

def prepare_data(self):
    a = np.random.uniform(0, 500, 500)
    b = np.random.normal(0, self.constant, len(a))

    c = a + b
    X = np.transpose(np.array([a, b]))

    # Converting numpy array to Tensor
    self.x_train_tensor = torch.from_numpy(X).float().to(device)
    self.y_train_tensor = torch.from_numpy(c).float().to(device)

    training_dataset = TensorDataset(self.x_train_tensor, self.y_train_tensor)

    self.training_dataset = training_dataset

def setup(self):
    data = self.training_dataset
    self.train_data, self.val_data = random_split(data, [400, 100])

def train_dataloader(self):
    return DataLoader(self.train_data)

def val_dataloader(self):
    return DataLoader(self.val_data)

您可以使用random_split() 拆分数据集。

Answer 2

此代码将返回标签为 y 和 a,b 作为 500 个随机示例合并到 X 中的 2 个特征。

import torch
from torch.autograd import Variable

def prepare_data(self):
    a = np.random.uniform(0, 500, 500) # random feature 1 x 500
    b = np.random.normal(0, 2, len(a)) # random feature 2 x 500
    X = np.transpose(np.array([a,b]))  # Merging feature 1 and 2 x 500
    y = np.random.randint(0,2,len(a))  #  random Labels as 0 and 1
    X = Variable(torch.from_numpy(X).float())            # Converting numpy array X to Torch tensor with auto_grad enabled
    y = Variable(torch.from_numpy(y).float())            # Converting numpy array y to Torch tensor with auto_grad enabled
    return X,y

Answer 3

只是你必须返回 torch 张量

import numpy as np 
import pytorch_lightning as pl 
from torch.utils.data import random_split, DataLoader

class DataModuleClass(pl.LightningDataModule):
    def __init__(self):
        super().__init__()
        self.constant = 2
        self.batch_size = 20
    self.transform = transforms.Compose([
        transforms.ToTensor()
    ])
        
    def prepare_data(self):
        a = np.random.uniform(0, 500, 500)
        b = np.random.normal(0, self.constant, len(a))
        c = a + b
        
        return torch.from_numpy(a).float(), torch.from_numpy(b).float(), torch.from_numpy(c).float()