I am new to Pytorch and I am working with 3D dataset. In the custom class I want to read a 3D image and return it as one 3D image and not as 2D slices. Now in dataloader I want to make sure that if batch size is set to two, then two 3D images are being passed. How can I do that with custom dataset class.
I am able to read the 3D images using nibabel. But I am not sure that, how to enusre that one 3D image should be considered as one entity.
Loading the volumes and transforming them into a tensor should work as the DataLoader shouldn’t care about the dimensions of your input as long as it can stack the samples in the batch dimension:
class MyDataset(Dataset):
def __init__(self):
self.data = torch.randn(100, 3, 128, 128, 128) # [nb_samples, channels, depth, width, height]
def __getitem__(self, index):
x = self.data[index]
return x
def __len__(self):
return len(self.data)
dataset = MyDataset()
loader = DataLoader(dataset, batch_size=2)
x = next(iter(loader))
print(x.shape)
# torch.Size([2, 3, 128, 128, 128]) # [batch_size, channels, depth, height, width]
@ptrblck , below is the part of my dataset class code:
for i, f in enumerate(files):
print("file_name", f)
nib_file = nib.load(os.path.join(images_path, f))
img = nib_file.get_fdata('unchanged', dtype=np.float32)
lbl = nib.load(os.path.join(data_path, 'Silver-standard', self.folder, f[:-7] + '_ss.nii.gz')).get_fdata(
'unchanged', dtype=np.float32)
if self.scale:
transformed = scaler.fit_transform(np.reshape(img, (-1, 1)))
self.img = np.reshape(transformed, img.shape)
if not self.sagittal:
self.img = np.moveaxis(img, -1, 0)
if self.rotate:
self.img = np.rot90(img, axes=(1, 2))
if img.shape[1] != img.shape[2]:
self.img = self.pad_image(img)
if not self.sagittal:
self.lbl = np.moveaxis(lbl, -1, 0)
if self.rotate:
self.lbl = np.rot90(lbl, axes=(1, 2))
if lbl.shape[1] != lbl.shape[2]:
self.lbl = self.pad_image(lbl)
spacing = [nib_file.header.get_zooms()] * img.shape[0]
self.voxel_dim = spacing
if i ==1: # done only to understand the shapes
break
def __getitem__(self, idx):
data = torch.from_numpy(self.img[idx])
labels = torch.from_numpy(self.lbl[idx])
return data, labels
Now in order to check the data is loaded as desired, I am trying to check the dimesion of 3D image on the first index of train_data as follows:
train_data = cc359_volume(config)
print("shape", train_data[0][1])
I expected it to give 200, 256,256, as nib_file_dim was (200, 256, 256) but it give the shape:
shape torch.Size([256, 256])
That means, the dataset class is not returning 3D volume as one entity. Rather, it is giving back one slice as indifvidual entity. Also I checked the dimesnion in loader as follows:
train_features, train_labels = next(iter(train_loader))
print(f"Feature batch shape: {train_features.size()}")
print(f"Labels batch shape: {train_labels.size()}")
It gives the following:
Feature batch shape: torch.Size([2, 256, 256])
Labels batch shape: torch.Size([2, 256, 256])
How can I get the datset to return the image as whole.
Negative strides are not supported and e.g. the np.rot90 transformation would create such negative strides as seen here:
img = np.random.randn(3, 224, 224, 224)
img = np.moveaxis(img, -1, 0)
print(img.strides)
# (8, 89915392, 401408, 1792)
# works
x = torch.from_numpy(img)
img = np.rot90(img, axes=(1, 2))
print(img.strides)
# (8, -401408, 89915392, 1792)
# breaks
x = torch.from_numpy(img)
# ValueError: At least one stride in the given numpy array is negative, and tensors with negative strides are not currently supported. (You can probably work around this by making a copy of your array with array.copy().)
# works again
x = torch.from_numpy(img.copy())
As the error message suggests, copy() the numpy array before trying to transform it into a PyTorch tensor.
In _getitem__(self, idx), data = torch.from_numpy(self.img[idx]) is retutning 256,256 (one slice dimesion). Why it not returning 200, 256, 256.
I don’t know and oyu would need to check the shape of img after loading it via nib_file.get_fdata and make sure no transformation is changing its shape in an unwanted way.
Transformations were not changing the shape. Previously in get_item()_, I was indexing on image it self like this:
def __getitem__(self, idx):
data = self.data[idx] #self.data has one image
That was changing the dimension and was taking the slices and givng [256,256]. I am now indexing on the file list in the folder itself as follows:
def __len__(self):
return len(self.files)
def __getitem__(self, idx):
nib_file = nib.load(os.path.join(self.images_path, self.files[idx]))
which gives the dimension [200,256,256]. Is my understanding correct?
Yes, loading the entire volume in the __getitem__ method is the right approach and I assumed self.data contained all volumes.