Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

DRAFT: 54 dataset #59

Draft
wants to merge 4 commits into
base: main
Choose a base branch
from
Draft

DRAFT: 54 dataset #59

Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
efbc317
Add labels to dataset object
gpitteri Jun 25, 2024
bb46c75
Merge branch 'main' into 54_dataset
giorgiapitteri Jul 2, 2024
5ddd683
device set to cpu by default
giorgiapitteri Jul 2, 2024
38f962d
minor changes
giorgiapitteri Jul 12, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
73 changes: 49 additions & 24 deletions src/explib/datasets.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -21,8 +21,9 @@ class DequantizedDataset(torch.utils.data.Dataset):
def __init__(
self,
dataset: T.Union[os.PathLike, torch.utils.data.Dataset, np.ndarray],
Copy link
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Should we also specify here that dataset can be a torch.Tensor as well?

Copy link
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Would it be better to have a dataset object like a dictionary with images/labels? Otherwise I would renamed this "images" instead of dataset since we now have the labels object as well.

labels: T.Union[np.ndarray, torch.Tensor],
num_bits: int = 8,
device: torch.device = None,
device: torch.device = "cpu",
):
if isinstance(dataset, torch.utils.data.Dataset) or isinstance(
dataset, np.ndarray
Expand All @@ -31,8 +32,15 @@ def __init__(
else:
self.dataset = pd.read_csv(dataset).values

#
if not isinstance(self.dataset, torch.Tensor):
self.dataset = torch.tensor(self.dataset)

self.dataset = self.dataset.to(device)

if not isinstance(labels, torch.Tensor):
Copy link
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

If dataset is a torch.utils.data.Dataset we don't need labels right? It automatically has a getitem method that returns the data and the label. So I would need to add a check here.

Copy link
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I think The baseclass getitem method is going to fail if the dataset is only a np array or a torch.Tensor. I understand there is the correct implementation in the derived class but maybe here we would need to put a check or raise an error?

labels = torch.Tensor(labels)
self.labels = labels.to(device)

self.num_bits = num_bits
self.num_levels = 2**num_bits
self.transform = transforms.Compose(
Expand All @@ -43,9 +51,9 @@ def __init__(
)

def __getitem__(self, index: int):
x, y = self.dataset[index]
x = self.dataset[index]
x = Tensor(self.transform(x))
return x, y
return x, self.labels[index]

def __len__(self):
return len(self.dataset)
Expand Down Expand Up @@ -241,7 +249,8 @@ def __init__(
dataloc: os.PathLike = None,
train: bool = True,
label: T.Optional[int] = None,
scale: bool = False
scale: bool = False,
device: torch.device = "cpu"
):
rel_path = (
"FashionMNIST/raw/train-images-idx3-ubyte"
Expand All @@ -256,21 +265,28 @@ def __init__(
if scale:
dataset = dataset[:, ::3, ::3]
dataset = dataset.reshape(dataset.shape[0], -1)
if label is not None:
rel_path = (

rel_path = (
"FashionMNIST/raw/train-labels-idx1-ubyte"
if train
else "FashionMNIST/raw/t10k-labels-idx1-ubyte"
)
path = os.path.join(dataloc, rel_path)
labels = idx2numpy.convert_from_file(path)
path = os.path.join(dataloc, rel_path)
labels = idx2numpy.convert_from_file(path)

if label is not None:
dataset = dataset[labels == label]
super().__init__(dataset, num_bits=8)
labels = labels[labels == label]

super().__init__(dataset, torch.Tensor(labels), num_bits=8, device=device)

def __getitem__(self, index: int):
x = Tensor(self.dataset[index].copy())
if not isinstance(self.dataset, torch.Tensor):
x = Tensor(self.dataset[index].copy())
else:
x = self.dataset[index]
x = self.transform(x)
return x, 0
return x, self.labels[index]


class FashionMnistSplit(DataSplit):
Expand All @@ -279,19 +295,21 @@ def __init__(
dataloc: os.PathLike = None,
val_split: float = 0.1,
label: T.Optional[int] = None,
device: torch.device = "cpu"
):
self.label = label
if dataloc is None:
dataloc = os.path.join(os.getcwd(), "data")
self.dataloc = dataloc
self.train = FashionMnistDequantized(self.dataloc, train=True, label=label)
self.train = FashionMnistDequantized(self.dataloc, train=True, label=label, device=device)
shuffle = torch.randperm(len(self.train))
self.val = torch.utils.data.Subset(
self.train, shuffle[: int(len(self.train) * val_split)]
)
self.train = torch.utils.data.Subset(
self.train, shuffle[int(len(self.train) * val_split) :]
)
self.test = FashionMnistDequantized(self.dataloc, train=False, label=label)
self.test = FashionMnistDequantized(self.dataloc, train=False, label=label, device=device)

def get_train(self) -> torch.utils.data.Dataset:
return self.train
Expand All @@ -312,7 +330,7 @@ def __init__(
digit: T.Optional[int] = None,
flatten=True,
scale: bool = False,
device: torch.device = None
device: torch.device = "cpu"
):
if train:
rel_path = "MNIST/raw/train-images-idx3-ubyte"
Expand All @@ -323,27 +341,32 @@ def __init__(
MNIST(dataloc, train=train, download=True)

dataset = idx2numpy.convert_from_file(path)

if scale:
dataset = dataset[:, ::3, ::3]
if flatten:
dataset = dataset.reshape(dataset.shape[0], -1)

if train:
rel_path = "MNIST/raw/train-labels-idx1-ubyte"
else:
rel_path = "MNIST/raw/t10k-labels-idx1-ubyte"
path = os.path.join(dataloc, rel_path)
labels = idx2numpy.convert_from_file(path)

if digit is not None:
if train:
rel_path = "MNIST/raw/train-labels-idx1-ubyte"
else:
rel_path = "MNIST/raw/t10k-labels-idx1-ubyte"
path = os.path.join(dataloc, rel_path)
labels = idx2numpy.convert_from_file(path)
dataset = dataset[labels == digit]
super().__init__(torch.Tensor(dataset), num_bits=8, device=device)
labels = labels[labels == digit]

super().__init__(torch.Tensor(dataset), torch.Tensor(labels), num_bits=8, device=device)

def __getitem__(self, index: int):
if not isinstance(self.dataset, torch.Tensor):
x = Tensor(self.dataset[index].copy())
else:
x = self.dataset[index]
x = self.transform(x)
return x, 0
return x, self.labels[index]

class MnistSplit(DataSplit):
def __init__(
Expand All @@ -352,8 +375,9 @@ def __init__(
val_split: float = 0.1,
digit: T.Optional[int] = None,
scale: bool = False,
device: torch.device = None
device: torch.device = "cpu"
):
self.digit = digit
if dataloc is None:
dataloc = os.path.join(os.getcwd(), "data")
self.dataloc = dataloc
Expand Down Expand Up @@ -393,3 +417,4 @@ def __init__(
if not os.path.exists(path):
CIFAR10(dataloc, train=train, download=True)


Loading