Torchvision 0.2.2 [best] — Bonus Inside

Because 0.2.2 is the last version without complex C++/CUDA extensions (pure Python), package managers like

Torchvision 0.2.2 was a solid foundation—it democratized image classification, detection, and augmentation for thousands of developers. But like all software, it aged. Today, it serves as a time capsule: a snapshot of computer vision at the cusp of the deep learning explosion. torchvision 0.2.2

conda create -n tv022 python=3.6 conda activate tv022 conda install pytorch=1.0.1 torchvision=0.2.2 -c pytorch Because 0

Models expected torch.FloatTensor for images. If you passed a DoubleTensor (default from some numpy conversions), you’d get cryptic RuntimeError: Expected object of scalar type Float but got scalar type Double . conda create -n tv022 python=3

Unlike modern versions which use weights=ResNet18_Weights.DEFAULT , version 0.2.2 used the boolean argument pretrained=True . If you are migrating code, this is one of the first depreciation warnings you will encounter.