Deep Lake offers two approaches for splitting dataset for training and validation:
Fully random splitting by row number (index)
Pseudo-random splitting using Deep Lake's internal method that is optimized for fast streaming
Setting up the Environment
import deeplake
from PIL import Image
import numpy as np
import os, time
import random
import torch
from torchvision import transforms
import getpass
First, let's set up our environment and copy the into your organization. This dataset is an image classification dataset that categorizes images by clothing type (trouser, shirt, etc.). Copying the dataset into your organization enables you to make edits.
In order to achieve reproducibility, you may save the views and use them in the future. Each saved view is assigned a id for reference. Saved views are pointers to data, and they do not duplicate data in storage.
train_view.save_view()
val_view.save_view()
views_listĀ =Ā ds.get_views()print(views_list)
We can also load a view using:
train_view = ds.load_view(views_list[0].id)
val_view = ds.load_view(views_list[1].id)
print(f"Length of train_view is {len(train_view)}")
print(f"Length of val_view is {len(val_view)}")
train_view = ds.load_view(views_list[0].id, optimize = True, num_workers = 2)
val_view = ds.load_view(views_list[1].id, optimize = True, num_workers = 2)
print(f"Length of train_view is {len(train_view)}")
print(f"Length of val_view is {len(val_view)}")
Pseudo-random Deep Lake splitting that is optimized for performance
If high performance is required without duplicating data, we recommend using Deep Lake's internal random_split method, which splits the dataset pseudo-randomly in order to maintain fast streaming.
train_view, val_view = ds.random_split([0.8, 0.2])
print(f"Length of train_view is {len(train_view)}")
print(f"Length of val_view is {len(val_view)}")
Training a Model Using Views
tform = transforms.Compose([
transforms.RandomRotation(20), # Image augmentation
transforms.ToTensor(), # Must convert to pytorch tensor for subsequent operations to run
transforms.Normalize([0.5], [0.5]),
])
batch_size = 32
# Since torchvision transforms expect PIL images, we use the 'pil' decode_method for the 'images' tensor. This is much faster than running ToPILImage inside the transform
train_loader = train_view.pytorch(num_workers = 0, shuffle = True, transform = {'images': tform, 'labels': None}, batch_size = batch_size, decode_method = {'images': 'pil'})
val_loader = val_view.pytorch(num_workers = 0, transform = {'images': tform, 'labels': None}, batch_size = batch_size, decode_method = {'images': 'pil'})
for train_batch in train_loader:
print(train_batch['images'].shape)
break
for val_batch in val_loader:
print(val_batch['images'].shape)
break
Congrats! You successfully created dataloaders from Deep Lake views! š
When loading or saving a view, we can specify the flag optimize = True, which the data for optimal streaming performance. Note that this is a computationally intensive and it will duplicate the data from the view at the storage location.
Views and datasets can be used interchangeably for training models. In this tutorial, we show how to create and iterate over dataloaders for the training and validation views, and a .
