Looking at image data

Always look at your data before training. These functions take whatever you have: a Dataset, a DataLoader, an (x, y) batch, a tensor plus labels, or a list of PIL images. Class names are read from dataset.classes when it exists.

The examples use FashionMNIST:

import torch, torchvision
from torchvision import transforms
import colabutilscs4337 as cu

train_ds = torchvision.datasets.FashionMNIST("data", train=True, download=True,
                                             transform=transforms.ToTensor())
train_loader = torch.utils.data.DataLoader(train_ds, batch_size=64, shuffle=True)

One row per class

cu.viz.show_by_class(train_ds)                    # 8 random examples of each class
cu.viz.show_by_class(train_ds, n_per_class=12)

x, y = next(iter(train_loader))
labels_map = {0: "T-Shirt", 1: "Trouser", 2: "Pullover", 3: "Dress", 4: "Coat",
              5: "Sandal", 6: "Shirt", 7: "Sneaker", 8: "Bag", 9: "Ankle Boot"}
cu.viz.show_by_class(x, y, labels_map, n_per_class=5)

The class names go on the left. A class with no examples in the batch gets an empty row, which is a quick way to spot an unbalanced batch.

A grid of samples

cu.viz.show_images(train_ds)                       # 32 random images
cu.viz.show_images(train_loader, n=16)             # first batch of a loader
cu.viz.show_images(x, y, train_ds.classes, ncols=4)
cu.viz.show_images(train_ds, seed=0)               # the same sample every time

Predictions vs. labels

Pass preds= and every title becomes prediction / truth, green when correct and red when wrong:

model.eval()
with torch.no_grad():
    preds = model(x).argmax(dim=1)
cu.viz.show_images(x, y, train_ds.classes, preds=preds)

Images as a PIL strip

tensor_to_image() joins images into one PIL.Image, handy for saving or logging:

cu.viz.tensor_to_image(x[:8])                # side by side (a row)
cu.viz.tensor_to_image(x[:8], dim=1)         # top to bottom (a column)
cu.viz.tensor_to_image(x[:8], pad=2)         # with white gaps
cu.viz.tensor_to_image(img_a, img_b)         # several tensors at once
cu.viz.tensor_to_image(x[:8]).save(run / "samples.png")

Normalized images

If your transform includes Normalize (for example the ImageNet statistics used with pretrained models), images are rescaled to look right automatically. To undo the normalization exactly, pass the same statistics:

cu.viz.show_images(x, y, classes, mean=cu.viz.IMAGENET_MEAN, std=cu.viz.IMAGENET_STD)

fastai

With fastai DataLoaders, cu.viz.show_images(dls) simply calls fastai’s own dls.show_batch(). For a single image prepared the way the model sees it, use learner_input().