How to Binarize datasets in PyTorch?

1

Hey, I am binarizing a dataset in pytorch, and I am doing it in my test dataset, here is the code -

from google.colab import drive
drive.mount('/content/drive')
data = "/content/drive/My Drive/AMD_new"


import torch
import helper
from torch import nn
from torch import optim
import torch.nn.functional as F
from torchvision import datasets, transforms, models
import torchvision.models as models

from torchvision import datasets ,transforms


#Changning the transform of the data-
transform_train = transforms.Compose([transforms.RandomHorizontalFlip(),
                                     transforms.RandomResizedCrop(224),
                                     # transforms.CenterCrop(224),
                                     transforms.ToTensor(),
                                     transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
                                     ])

transform_test = transforms.Compose([transforms.RandomHorizontalFlip(),
                                     transforms.RandomResizedCrop(224),
                                     # transforms.CenterCrop(224),
                                     transforms.ToTensor(),
                                     transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
                                     ])

# choose the training and test datasets
train_data = datasets.ImageFolder(data+"/train", transform=transform_train)
test_data = datasets.ImageFolder(data+"/val", transform = transform_test)

dataloader_train = torch.utils.data.DataLoader(train_data, batch_size=32, shuffle=True, num_workers=2)
dataloader_test = torch.utils.data.DataLoader(test_data, batch_size=32, num_workers=2)

but when I am using these line at last, it is giving me error -

# Binarize the output
dataloader_test = label_binarize(dataloader_test, classes=[0, 1, 2, 3])
nb_classes = dataloader_test.shape[1]

Any lead on where this error comes from, as per I know it is not possible to use scikit with out binarize the data. Error is

TypeError: Singleton array array(<torch.utils.data.dataloader.DataLoader object at 0x7fc3048321d0>,
dtype=object) cannot be considered a valid collection.
It means my dataloader_test is a single object right, but how, it has four class and each class has some 10 photos. Thank you or your help.

2

Not sure what you mean, maybe you should try target_transform or write your own dataset

3

Sorry for the late reply. I am trying to get ROC curve using Scikit learn, I have fpr and tpr too, but to print the ROC, I need to binarize the dataoader_test. How can I do that, when I trying this, resulting this -

E: Package 'python-software-properties' has no installation candidate
··········
fuse: mountpoint is not empty
fuse: if you are sure this is safe, use the 'nonempty' mount option
Drive already mounted at /content/drive; to attempt to forcibly remount, call drive.mount("/content/drive", force_remount=True).
---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
<ipython-input-6-e1dbe1e9866d> in <module>()
     58 
     59 # Binarize the output
---> 60 dataloader_test = label_binarize(dataloader_test, classes=[0, 1, 2, 3])
     61 nb_classes = dataloader_test.shape[1]

/usr/local/lib/python3.6/dist-packages/sklearn/preprocessing/label.py in label_binarize(y, classes, neg_label, pos_label, sparse_output)
    579         # XXX Workaround that will be removed when list of list format is
    580         # dropped
--> 581         y = check_array(y, accept_sparse='csr', ensure_2d=False, dtype=None)
    582     else:
    583         if _num_samples(y) == 0:

/usr/local/lib/python3.6/dist-packages/sklearn/utils/validation.py in check_array(array, accept_sparse, accept_large_sparse, dtype, order, copy, force_all_finite, ensure_2d, allow_nd, ensure_min_samples, ensure_min_features, warn_on_dtype, estimator)
    575     shape_repr = _shape_repr(array.shape)
    576     if ensure_min_samples > 0:
--> 577         n_samples = _num_samples(array)
    578         if n_samples < ensure_min_samples:
    579             raise ValueError("Found array with %d sample(s) (shape=%s) while a"

/usr/local/lib/python3.6/dist-packages/sklearn/utils/validation.py in _num_samples(x)
    140         if len(x.shape) == 0:
    141             raise TypeError("Singleton array %r cannot be considered"
--> 142                             " a valid collection." % x)
    143         # Check that shape is returning an integer or default to len
    144         # Dask dataframes may not return numeric shape[0] value

TypeError: Singleton array array(<torch.utils.data.dataloader.DataLoader object at 0x7fc3048321d0>,
      dtype=object) cannot be considered a valid collection.

You know what need to be done? Follwing this tutorial - https://scikit-learn.org/stable/auto_examples/model_selection/plot_roc.html

4

You need to first get all data out by:

all_label = torch.zeros(N,K)
for i,(data, label) in enumerate(dataloader):
      all_label[i]=label
label_binarize(all_label)

I’ll advise you to do it in Dataset, it’s a very easy operation, no need to make it complicated.

5

So, basically as I am not very experienced, do I have to initialize N and K with some random integers? And I have already done this loop state in my later part like this -

for epoch in range(epochs):
  
  running_loss = 0
  model.train()
  for images, labels in dataloader_train:
    
    #steps += 1
    images, labels = images.to(device), labels.to(device)
    
    optimizer.zero_grad()
    
    output = model.forward(images)
    p = torch.nn.functional.softmax(output, dim=1)
    prediction = torch.argmax(p, dim=1)
    #loss = torch.nn.functional.nll_loss(torch.log(p), y)
    loss = criterion(output, labels)
    loss.backward()
    optimizer.step()
    
    running_loss += loss.item()

so here in the first cell, where I have written my dataloader code, I have to write like this -

all_label = torch.zeros(N,K)
for i,(image, labels) in enumerate(dataloader_test):
      all_label[i]=labels
label_binarize(all_label)

am I right? Thanks.