SimCLR in PyTorch, with numbers and no Apex

What it does

This is a readable PyTorch reimplementation of SimCLR, the Google framework that learns vision representations by teaching a model which augmented views of an image belong together. After pretraining on unlabeled data, it freezes the backbone and trains a linear classifier on top, giving you concrete top-1 accuracy on STL-10 and CIFAR-10 rather than just loss curves.

The interesting bit

The author skipped NVIDIA Apex in favor of PyTorch’s native AMP for 16-bit training, removing one common dependency headache. The bundled evaluation protocol is the quietly valuable part: most SimCLR repos stop at pretraining, but this one ships a Colab notebook that runs the full linear-probe test.

Key highlights

• Native PyTorch AMP for 16-bit training; no Apex required.
• Linear evaluation via an included Colab notebook, not just pretraining code.
• Reported top-1 accuracy: 74.45% on STL-10 and 69.82% on CIFAR-10 using ResNet-18.
• CPU debug mode available for quick smoke tests.
• Companion blog post breaks down the architecture in detail.

Caveats

• The provided checkpoints mix architectures and training durations—ResNet-18 runs for 100 epochs while ResNet-50 stops at 50—so you cannot directly compare model sizes from the table alone.
• Results are limited to small-scale datasets; ImageNet benchmarks are not included.

Verdict

Use this if you want a teaching-friendly SimCLR baseline with evaluation baked in. Look elsewhere if you need ImageNet-scale weights or a distributed training framework.