YouTipPyTorch Data Transforms |
\n\nIn PyTorch, data transformation is a mechanism for processing data during loading, converting raw data into a format suitable for model training. It is primarily accomplished using tools provided by torchvision.transforms.
Data transforms can not only perform basic data preprocessing (such as normalization, resizing, etc.), but also help with data augmentation (such as random cropping, flipping, etc.), thereby improving the model's generalization ability.
\n\nWhy are Data Transforms Needed?
\n\nData Preprocessing:
\n- \n
- Adjust data format, size, and range to make them suitable for model input. \n
- For example, images need to be resized to a fixed size, converted to tensor format, and normalized to [0, 1]. \n
Data Augmentation:
\n- \n
- Apply transformations to data during training to increase diversity. \n
- For example, increase data sample variations through random rotation, flipping, and cropping to avoid overfitting. \n
Flexibility:
\n- \n
- By defining a series of transform operations, data can be processed dynamically, simplifying the complexity of data loading. \n
In PyTorch, the torchvision.transforms module provides various transformation operations for image processing.
Basic Transform Operations
\n\n| Transform Function Name | \nDescription | \nExample | \n
|---|---|---|
transforms.ToTensor() | \nConverts PIL images or NumPy arrays to PyTorch tensors, automatically normalizing pixel values to [0, 1]. | \ntransform = transforms.ToTensor() | \n
transforms.Normalize(mean, std) | \nNormalizes images so that data follows a zero mean and unit variance. | \ntransform = transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) | \n
transforms.Resize(size) | \nResizes images to ensure consistent input sizes for the network. | \ntransform = transforms.Resize((256, 256)) | \n
transforms.CenterCrop(size) | \nCrops a region of the specified size from the center of the image. | \ntransform = transforms.CenterCrop(224) | \n
1. ToTensor
\nConverts PIL images or NumPy arrays to PyTorch tensors.
\nIt also normalizes pixel values from [0, 255] to [0, 1].
\nfrom torchvision import transforms transform = transforms.ToTensor()2. Normalize
\nStandardizes data to conform to specific mean and standard deviation.
\nCommonly used for image data to normalize pixel values to zero mean and unit variance.
\ntransform = transforms.Normalize(mean=[0.5], std=[0.5]) # Normalize to [-1, 1]3. Resize
\nAdjusts the size of the image.
\ntransform = transforms.Resize((128, 128)) # SetResize image to 128x1284. CenterCrop
\nCrops a region of the specified size from the center of the image.
\ntransform = transforms.CenterCrop(128) # Crop a 128x128 regionData Augmentation Operations
\n\n| Transform Function Name | \nDescription | \nExample | \n
|---|---|---|
transforms.RandomHorizontalFlip(p) | \nRandomly flips the image horizontally. | \ntransform = transforms.RandomHorizontalFlip(p=0.5) | \n
transforms.RandomRotation(degrees) | \nRandomly rotates the image. | \ntransform = transforms.RandomRotation(degrees=45) | \n
transforms.ColorJitter(brightness, contrast, saturation, hue) | \nAdjusts the brightness, contrast, saturation, and hue of the image. | \ntransform = transforms.ColorJitter(brightness=0.2, contrast=0.2, saturation=0.2, hue=0.1) | \n
transforms.RandomCrop(size) | \nRandomly crops a region of the specified size. | \ntransform = transforms.RandomCrop(224) | \n
transforms.RandomResizedCrop(size) | \nRandomly crops the image and resizes it to the specified size. | \ntransform = transforms.RandomResizedCrop(224) | \n
1. RandomCrop
\nRandomly crops a specified size from the image.
\ntransform = transforms.RandomCrop(128)2. RandomHorizontalFlip
\nFlips the image horizontally with a certain probability.
\ntransform = transforms.RandomHorizontalFlip(p=0.5) # 50% Random flip with probability3. RandomRotation
\nRandomly rotates by a certain angle.
\ntransform = transforms.RandomRotation(degrees=30) # Random rotation -30 to +30 degrees4. ColorJitter
\nRandomly changes the brightness, contrast, saturation, or hue of the image.
\ntransform = transforms.ColorJitter(brightness=0.5, contrast=0.5)Composing Transforms
\n\n| Transform Function Name | \nDescription | \nExample | \n
|---|---|---|
transforms.Compose() | \nCombines multiple transforms together, applying them sequentially in order. | \ntransform = transforms.Compose([transforms.ToTensor(), transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]), transforms.Resize((256, 256))]) | \n
Combine multiple transforms using transforms.Compose.
transform = transforms.Compose([ transforms.Resize((128, 128)), transforms.RandomHorizontalFlip(p=0.5), transforms.ToTensor(), transforms.Normalize(mean=[0.5], std=[0.5])])Custom Transforms
\nIf the functionality provided by transforms does not meet your needs, you can implement custom classes or functions.
Examples
\nclass CustomTransform:\n\ndef __call__ (self, x):\n\n# Custom transform logic can be defined here\n\nreturn x * 2\n\ntransform = CustomTransform()\n\n
Examples
\nApplying Transforms to an Image Dataset
\nLoad the MNIST dataset and apply transforms.
\n\nExample
\nfrom torchvision import datasets, transforms\n\nfrom torch.utils.data import DataLoader\n\n# Define transforms\n\n transform = transforms.Compose([\n\n transforms.Resize((128,128)),\n\n transforms.ToTensor(),\n\n transforms.Normalize(mean=[0.5], std=[0.5])\n\n])\n\n# Load dataset\n\n train_dataset = datasets.MNIST(root='./data', train=True, transform=transform, download=True)\n\n# Use DataLoader\n\n train_loader = DataLoader(dataset=train_dataset, batch_size=32, shuffle=True)\n\n# View transformed data\n\nfor images, labels in train_loader:\n\nprint("Image tensor size:", images.size())# [batch_size, 1, 128, 128]\n\nbreakThe output result is:
\nImage tensor size: torch.Size([32, 1, 128, 128])Visualizing Transform Effects
\nThe following code demonstrates a comparison between original and transformed data.
\n\nExample
\nimport matplotlib.pyplot as plt\n\nfrom torchvision import datasets\n\nfrom torchvision import datasets, transforms\n\n# Visualize original and augmented images\n\n transform_augment = transforms.Compose([\n\n transforms.RandomHorizontalFlip(),\n\n transforms.RandomRotation(30),\n\n transforms.ToTensor()\n\n])\n\n# Load dataset\n\n dataset = datasets.MNIST(root='./data', train=True, download=True, transform=transform_augment)\n\n# Display Image\n\ndef show_images(dataset):\n\n fig, axs = plt.subplots(1,5, figsize=(15,5))\n\nfor i in range(5):\n\n image, label = dataset\n\n axs.imshow(image.squeeze(0), cmap='gray')# Set (1, H, W) Convert to (H, W)\n\n axs.set_title(f"Label: {label}")\n\n axs.axis('off')\n\n plt.show()\n\nshow_images(dataset)Displayed as shown below:
\n