如何为多输出回归问题编写 Pytorch Dataset 类和 DataLoader？

Question

我正在处理基于事件的数据（包括时间戳、x 和 y 坐标以及极性）来解决神经形态问题。但是，我的问题是，当我尝试预测多个输出时，如何创建数据集和正确的数据加载器？具体来说，我试图预测速度矢量的 x 分量、y 分量和 z 分量。这就是我当前的自定义 Dataset 类的样子：

import os
import ast
import torch
import tonic
import torchvision
import numpy as np 
import pandas as pd
import tonic.transforms as transforms
from torch.utils.data import DataLoader

class SyntheticRecording(tonic.Dataset):
    """
        Synthetic event camera recordings dataset.
    """
    def __init__(self, csv_file):
        super(SyntheticRecording, self).__init__()
        self.csv_file = csv_file
        df = pd.read_csv(self.csv_file, index_col = False)
        self.events = df['Events'] # Select only last column of dataframe
        self.target = df[['Vel_x', 'Vel_y', 'Vel_z']] # Select every column except last column of dataframe
        assert(self.target.shape[0] == len(self.events))
        self.sensor_size = (1920, 1080, 2)
    
    """
        Retrieve the index i to get the ith sample from the dataset. Apply the appropriate transformations.
    """
    def __getitem__(self, index):
      list_ = ast.literal_eval(self.events[index])
      t = []
      x = []
      y = []
      p = []
      for e in list_:
        t.append(e[0] * 1e6) # Convert to microseconds
        x.append(e[1])
        y.append(e[2])
        p.append(e[3])
      events = tonic.io.make_structured_array(x, y, t, p) # Ordering is xytp now
      
      # Denoise removes isolated, one-off events
      frame_transform = transforms.Compose([transforms.Denoise(filter_time = 10000), 
                                            transforms.ToFrame(sensor_size = self.sensor_size, 
                                                         time_window = 1000)
                                            ])
      transformed_frames = frame_transform(events)
      vel_x = np.array(self.target.loc[index][0]).astype('float')
      vel_y = np.array(self.target.loc[index][1]).astype('float')
      vel_z = np.array(self.target.loc[index][2]).astype('float')
      
      sample = {'frames': transformed_frames,
                'vel_x': vel_x,
                'vel_y': vel_y,
                'vel_z': vel_z}

      return sample

这就是我尝试创建 DataLoader 的方式：

batch_size = 16
trainloader = DataLoader(sr, batch_size = batch_size, collate_fn = tonic.collation.PadTensors(), shuffle = True, drop_last = True)

每当我尝试迭代框架和目标值（3 个值）时，我都会收到以下错误：

for frames, targets in trainloader:
  print(frames.shape)
  print(targets.shape)

frames, targets = next(iter(trainloader))

结果为：ValueError ：太多值无法解压（预期为 2）

创建可以处理多重预测回归问题的 DataLoader 的正确方法是什么？

编辑：

我使用它作为我的资源来尝试处理多输出部分： https://medium.com/jdsc-tech-blog/multioutput-cnn-in-pytorch-c5f702d4915f

Answer 1

也许试试这个：

for sample in trainloader:
   print(sample) 

通过这种方式，您可以检查打印的输出中返回了多少个值。

Answer 2

有点晚了，但我也偶然发现了同样的问题，这是我解决它的方法：

问题不在于迭代数据加载器，而是来自您传递给 collat​​e_fn >DataLoader 因为它假设您的目标只有 1 个值。您得到的错误来自 tonic 的 PadTensor()< /a> 函数迭代新批次：

for sample, target in batch:
    ...

它假设您没有字典，但有一个可以解包的元组。后来它假设目标也只是一个值。

如何修复

1. 修改整理函数：

创建一个新类，在其中删除目标仅为一个值的假设，如下所示：

class MultiTargetPadTensors:
    def __init__(self, batch_first: bool = True):
        self.batch_first = batch_first

    def __call__(self, batch):
        samples_output = []
        targets_output = []

        max_length = max([sample.shape[0] for sample, target in batch])

        for sample, target in batch:
            if not isinstance(sample, torch.Tensor):
                sample = torch.tensor(sample)

            sample = torch.cat(
                (
                    sample,
                    torch.zeros(
                        max_length - sample.shape[0],
                        *sample.shape[1:],
                        device=sample.device
                    ),
                )
            )
            samples_output.append(sample)
            targets_output.append(target)
        
        samples_output = torch.stack(samples_output, 0 if self.batch_first else 1)
        targets_output = ??? # TODO: Stack the targets according to your use case
        return (samples_output, targets_output)

DataLoader 现在应如下所示：DataLoader(sr, batch_size = batch_size, collat​​e_fn = MultiTargetPadTensors(), shuffle = True, drop_last = True)

2. 调整数据集：

返回一个元组，例如 这：

    ...
    sample = (transformed_frames, [vel_x, vel_y, vel_z])
    return sample