Step 7: Connecting Deep Lake Datasets to ML Frameworks
Connecting Deep Lake Datasets to machine learning frameworks such as PyTorch and TensorFlow.
Deep Lake Datasets can be connected to popular ML frameworks such as PyTorch and TensorFlow using minimal boilerplate code. Our methods enable you to train models while streaming data from the cloud without bottlenecking the training process!
There are two syntaxes that can be used to train models in Pytorch using Deep Lake datasets:
- 1.Deep Lake Data Loaders are highly-optimized and unlock the fastest streaming and shuffling using Deep Lake's internal shuffling method. However, they do not support custom sampling or fully-random shuffling that is possible using PyTorch datasets + data loaders.
- 2.Pytorch Datasets + Data Loaders enable all the customizability supported by PyTorch. However, they have highly sub-optimal streaming using Deep Lake datasets and may result in 5X+ slower performance compared to using Deep Lake data loaders.
Best option for fast streaming!
The fastest streaming of data to GPUs using PyTorch is achieved using Deep Lake's built-in PyTorch dataloader
ds.pytorch() . If your model training is highly sensitive to the randomization of the input data, please pre-shuffle the data, or explore our writeup onShuffling in ds.pytorch().import deeplake
from torchvision import datasets, transforms, models
ds = deeplake.dataset('hub://activeloop/cifar100-train') # Deep Lake Dataset
The
transform parameter in ds.pytorch() is a dictionary where the key is the tensor name and the value is the transformation function for that tensor. If a tensor's data does not need to be returned, the tensor should be omitted from the keys. If a tensor's data does not need to be modified during preprocessing, the transformation function for the tensor is set as None.tform = transforms.Compose([
transforms.ToPILImage(), # Must convert to PIL image for subsequent operations to run
transforms.RandomRotation(20), # Image augmentation
transforms.ToTensor(), # Must convert to pytorch tensor for subsequent operations to run
transforms.Normalize([0.5, 0.5, 0.5], [0.5, 0.5, 0.5]),
])
#PyTorch Dataloader
dataloader= ds.pytorch(batch_size = 16, num_workers = 2,
transform = {'images': tform, 'labels': None}, shuffle = True)
Transform are sometimes more complex where the same transform might need to be applied to all tensors, or tensors need to be combined in a transform. In this case, you can use the syntax below to perform the exact same transform as above:
def transform(sample_in):
return {'images': tform(sample_in['images']), 'labels': sample_in['labels']}
#PyTorch Dataloader
dataloader= ds.pytorch(batch_size = 16, num_workers = 2,
transform = transform,
tensors = ['images', 'labels'],
shuffle = True)
Some datasets such as imagenet contain both grayscale and color images, which can cause errors when the transformed images are passed to the model. To convert only the grayscale images to color format, you can add this Torchvision transform to your pipeline:
transforms.Lambda(lambda x: x.repeat(int(3/x.shape[0]), 1, 1))Best option for full customizability.
Deep Lake datasets can be integrated in the PyTorch Dataset class by passing the
ds object to the PyTorch Dataset's constructor and pulling data in the __getitem__ method using self.ds.image[ids].numpy():from torch.utils.data import DataLoader, Dataset
class ClassificationDataset(Dataset):
def __init__(self, ds, transform = None):
self.ds = ds
self.transform = transform
def __len__(self):
return len(self.ds)
def __getitem__(self, idx):
image = self.ds.images[idx].numpy()
label = self.ds.labels[idx].numpy(fetch_chunks = True).astype(np.int32)
if self.transform is not None:
image = self.transform(image)
sample = {"images": image, "labels": label}
return sample
When loading data sequentially, or when randomly loading samples from a tensor that fits into the cache (such as
class_labels) it is recommended to set fetch_chunks = True. This increases the data loading speed by avoiding separate requests for each individual sample. This is not recommended when randomly loading large tensors, because the data is deleted from the cache before adjacent samples from a chunk are used.The PyTorch dataset + data loader is instantiated using the built-in PyTorch functions:
cifar100_pytorch = ClassificationDataset(ds_train, transform = tform)
dataloader_pytroch = DataLoader(dataset_pt, batch_size = 16, num_workers = 2, shuffle = True)
You can iterate through both data loaders above using the exact same syntax. Loading the first batch of data using the Deep Lake data loader may take up to 30 seconds because the shuffle buffer is filled before any data is returned.
for data in dataloader:
print(data)
break
# Training Loop
for data in dataloader_pytorch:
print(data)
break
# Training Loop
For more information on training, check out the tutorial on Training an Image Classification Model in PyTorch
Deep Lake Datasets can be converted to TensorFlow Datasets using
ds.tensorflow(). Downstream, functions from the tf.Data API such as map, shuffle, etc. can be applied to process the data before training.ds # Deep Lake Dataset object, to be used for training
ds_tf = ds.tensorflow() # A TensorFlow Dataset
Last modified 1mo ago