Image sizes for training and prediction

Often, images that you use for training and inference have different heights and widths and different aspect ratios. That fact brings two challenges to a deep learning pipeline:

• PyTorch requires all images in a batch to have the same height and width.
• If a neural network is not fully convolutional, you have to use the same width and height for all images during training and inference. Fully convolutional architectures, such as UNet, can work with images of any size.

There are three common ways to deal with those challenges:

1. Resize all images and masks to a fixed size (e.g., 256x256 pixels) during training. After a model predicts a mask with that fixed size during inference, resize the mask to the original image size. This approach is simple, but it has a few drawbacks:

• The predicted mask is smaller than the image, and the mask may lose some context and important details of the original image.
• This approach may be problematic if images in your dataset have different aspect ratios. For example, suppose you are resizing an image with the size 1024x512 pixels (so an image with an aspect ratio of 2:1) to 256x256 pixels (1:1 aspect ratio). In that case, this transformation will distort the image and may also affect the quality of predictions.

2. If you use a fully convolutional neural network, you can train a model with image crops, but use original images for inference. This option usually provides the best tradeoff between quality, speed of training, and hardware requirements.
3. Do not alter the sizes of images and use source images both for training and inference. With this approach, you won't lose any information. However, original images could be quite large, so they may require a lot of GPU memory. Also, this approach requires more training time to obtain good results.

Some architectures, such as UNet, require that an image's size must be divisible by a downsampling factor of a network (usually 32), so you may also need to pad an image with borders. Albumentations provides a particular transformation for that case.