python - Convert Pandas dataframe to PyTorch tensor?

Question

I want to train a simple neural network with PyTorch on a pandas dataframe df.

One of the columns is named "Target", and it is the target variable of the network. How can I use this dataframe as input to the PyTorch network?

I tried this, but it doesn't work:

import pandas as pd
import torch.utils.data as data_utils

target = pd.DataFrame(df['Target'])
train = data_utils.TensorDataset(df, target)
train_loader = data_utils.DataLoader(train, batch_size=10, shuffle=True)

Answer 1

I'm referring to the question in the title as you haven't really specified anything else in the text, so just converting the DataFrame into a PyTorch tensor.

Without information about your data, I'm just taking float values as example targets here.

Convert Pandas dataframe to PyTorch tensor?

import pandas as pd
import torch
import random

# creating dummy targets (float values)
targets_data = [random.random() for i in range(10)]

# creating DataFrame from targets_data
targets_df = pd.DataFrame(data=targets_data)
targets_df.columns = ['targets']

# creating tensor from targets_df 
torch_tensor = torch.tensor(targets_df['targets'].values)

# printing out result
print(torch_tensor)

Output:

tensor([ 0.5827,  0.5881,  0.1543,  0.6815,  0.9400,  0.8683,  0.4289,
         0.5940,  0.6438,  0.7514], dtype=torch.float64)

Tested with Pytorch 0.4.0.

I hope this helps, if you have any further questions - just ask. :)

Answer 2

Maybe try this to see if it can fix your problem(based on your sample code)?

train_target = torch.tensor(train['Target'].values.astype(np.float32))
train = torch.tensor(train.drop('Target', axis = 1).values.astype(np.float32)) 
train_tensor = data_utils.TensorDataset(train, train_target) 
train_loader = data_utils.DataLoader(dataset = train_tensor, batch_size = batch_size, shuffle = True)

Answer 3

You can use below functions to convert any dataframe or pandas series to a pytorch tensor

import pandas as pd
import torch

# determine the supported device
def get_device():
    if torch.cuda.is_available():
        device = torch.device('cuda:0')
    else:
        device = torch.device('cpu') # don't have GPU 
    return device

# convert a df to tensor to be used in pytorch
def df_to_tensor(df):
    device = get_device()
    return torch.from_numpy(df.values).float().to(device)

df_tensor = df_to_tensor(df)
series_tensor = df_to_tensor(series)

Answer 4

Simply convert the pandas dataframe -> numpy array -> pytorch tensor. An example of this is described below:

import pandas as pd
import numpy as np
import torch

df = pd.read_csv('train.csv')
target = pd.DataFrame(df['target'])
del df['target']
train = data_utils.TensorDataset(torch.Tensor(np.array(df)), torch.Tensor(np.array(target)))
train_loader = data_utils.DataLoader(train, batch_size = 10, shuffle = True)

Hopefully, this will help you to create your own datasets using pytorch (Compatible with the latest version of pytorch).

Answer 5

You can pass the df.values attribute (a numpy array) to the Dataset constructor directly:

import torch.utils.data as data_utils

# Creating np arrays
target = df['Target'].values
features = df.drop('Target', axis=1).values

# Passing to DataLoader
train = data_utils.TensorDataset(features, target)
train_loader = data_utils.DataLoader(train, batch_size=10, shuffle=True)

---

Note: Your features (df) also contains the target variable (df['Target']) i.e. your network is 'cheating', since it can see the targets in the input. You need to remove this column from the set of features.

Answer 6

#This works for me

target = torch.tensor(df['Targets'].values)
features = torch.tensor(df.drop('Targets', axis = 1).values)

train = data_utils.TensorDataset(features, target)
train_loader = data_utils.DataLoader(train, batch_size=10, shuffle=True)