Dataset&Dataloader

1

Hi this is my dataset and codes.
In training code has tensor error and after transform has to pilimage error.
this is my code:

import pandas as pd
import torchvision.transforms as trs
from torch.utils.data import Dataset, DataLoader
from ReadData import mean, std

#Dataset
class welldata(Dataset):
def init(self, press_file=‘press.csv’, label_file=‘label.csv’):
self.press = pd.read_csv(press_file)
self.label = pd.read_csv(label_file)
self.tr = trs.Compose([
trs.Normalize((mean), (std))
trs.ToTensor()
])

def __len__(self):
    return len(self.press)

def __getitem__(self, idx):
    press = self.tr(self.press.iloc[idx])
    label = self.tr(self.label.iloc[idx])
    return press, label

well_dataloader = DataLoader(dataset= welldata(), batch_size=5, shuffle=True)

#Creat model
class model(nn.Module):
def init(self):
super(model, self).init()
self.lazy1 = nn.LazyLinear(1)
self.rnn1 = nn.RNN(2, 2, 20)
self.relu1 = nn.ReLU()
self.bifc1 = nn.Bilinear(2, 5, 5)
self.relu2 = nn.Hardtanh()
self.lstm1 = nn.LSTM(5, 5)
self.relu3 = nn.ReLU()

def forward(self, z1, z2):
    y = self.lazy1(z1)
    y = torch.reshape(y, (z1.size(0), 2))
    y, _ = self.rnn1(y)
    y = self.relu1(y)
    y = self.bifc1(y, z2)
    y = self.relu2(y)
    y, _ = self.lstm1(y)
    y = self.relu3(y)
    return y

Optimizer

model = model()
optimizer = torch.optim.AdamW(model.parameters(), lr=0.1)

Loss

loss = nn.SoftMarginLoss()

#Training
for x1, x2, yt in well_dataloader:
optimizer.zero_grad()
yp = model(x1, x2)
loss_value = loss(yp, yt)
loss_value.backward()
optimizer.step()

and this is my data
after debuging I’ll created data

column 0 and 1 is list and has many num float
column 0 and 1 are input data in lazylinear
and other column are input in RNN
and they are input in bilinear Network
this data are label and network must be answer

Data
Data1920×2016 358 KB

what’s the problem?

2

It’s unclear from the code and screenshots what the actual error is (at least to me). Could you describe the issue in more detail and, if possible, post a minimal, executable, and formatted code snippet to reproduce the issue?

3 
import torch
import torch.nn as nn
import pandas as pd
import torchvision.transforms as trs
from torch.utils.data import Dataset, DataLoader
from ReadData import mean, std

class welldata(Dataset):
    def __init__(self, press_file='press.csv', label_file='label.csv'):
        self.press = pd.read_csv(press_file)
        self.label = pd.read_csv(label_file)
        self.tr = trs.Compose([
                               trs.ToTensor(),
                               trs.Normalize((mean), (std))
                              ])

    def __len__(self):
        return len(self.press)

    def __getitem__(self, idx):
        press = self.press.iloc[idx]
        # press = self.tr(self.press.iloc[idx])
        label = self.label.iloc[idx]
        # label = self.tr(self.label.iloc[idx])
        return press, label

well_dataloader = DataLoader(dataset= welldata(), batch_size=5, shuffle=True)

class model(nn.Module):
    def __init__(self):
        super(model, self).__init__()
        self.lazy1 = nn.LazyLinear(1)
        self.rnn1 = nn.RNN(2, 2, 20)
        self.relu1 = nn.ReLU()
        self.bifc1 = nn.Bilinear(2, 5, 5)
        self.relu2 = nn.Hardtanh()
        self.lstm1 = nn.LSTM(5, 5)
        self.relu3 = nn.ReLU()

    def forward(self, z1, z2):
        y = self.lazy1(z1)
        y = torch.reshape(y, (z1.size(0), 2))
        y, _ = self.rnn1(y)
        y = self.relu1(y)
        y = self.bifc1(y, z2)
        y = self.relu2(y)
        y, _ = self.lstm1(y)
        y = self.relu3(y)
        return y

# Optimizer
model = model()
optimizer = torch.optim.AdamW(model.parameters(), lr=0.1)

# Loss
loss = nn.SoftMarginLoss()

for x1, x2, yt in well_dataloader:
    optimizer.zero_grad()
    yp = model(x1, x2)
    loss_value = loss(yp, yt)
    loss_value.backward()
    optimizer.step()

raise TypeError(f"pic should be PIL Image or ndarray. Got {type(pic)}")
TypeError: pic should be PIL Image or ndarray. Got <class ‘pandas.core.series.Series’>

this is error after using transforms
and my data type is not image

raise TypeError(default_collate_err_msg_format.format(elem_type))
TypeError: default_collate: batch must contain tensors, numpy arrays, numbers, dicts or lists; found <class ‘pandas.core.series.Series’>

and this error without using transforms

4

You would need to transform the pandas.Series to an image before trying to apply the transformation on it.

5

Is there a problem with my data set?
I feel that I have not written anything in the data set

6

I don’t know if your dataset has any other issues, but right now the code is failing since a pandas.Series object is not supported in the used transformations as it does not represent an image type.
Assuming the internal numpy array represents an image, you could create a PIL.Image via:

image = PIL.Image.fromarray(series.iloc[idx].to_numpy())

but would of course need to check the shape and dtype making sure it is indeed a valid image type.