Libraries
Open-source tools for efficient AI.
Two libraries, Apache 2.0 licensed. Drop them into any PyTorch project to compress and benchmark your models in a few lines of code.
01
FasterAI
Compress any PyTorch model.
A toolkit for shrinking and accelerating neural networks with research-backed techniques. Drop into existing training pipelines in a few lines of code.
$
pip install fasterai
02
FasterBench
Benchmark what matters.
Measure speed, size, memory, compute, and energy in a single function call. The complete picture of your model's efficiency, with no extra wiring.
$
pip install fasterbench
03
Quickstart
From pretrained to production.
Load a pretrained model, compress it, benchmark it. Four steps to a smaller, faster model.
from fasterai.prune.all import *
from fasterbench.benchmark import *
from torchvision.models import resnet18
# 1) Load a baseline model
model = resnet18()
# 2) Compress in seconds
pruner = Pruner(model, 50, 'global', large_final)
pruner.prune_model()
# 3) Benchmark compressed model
bench = benchmark(model, dummy)
# 4) Save compressed model
torch.save(model.state_dict(), "compressed-model.pth")