All Tools
T
Dev ToolsFreeOpen Source
TORCHOPT
Differentiable PyTorch optimizers for meta-learning
Apache-2.0
ABOUT
Meta-learning needs gradients through an inner optimizer, but PyTorch step() is not a differentiable transform. TorchOpt exposes functional, implicit-diff optimizers so MAML-style loops nest updates without a homemade tape.
INSTALL
pip install torchoptINTEGRATION GUIDE
1. Differentiate through inner SGD steps in a MAML trainer
2. Replace torch.optim with a functional optimizer in a loop
3. Apply implicit differentiation to a bilevel PyTorch problem
4. Meta-learn hyperparameters by unrolling TorchOpt updates
TAGS
pythonpytorchoptimizationmeta-learningautodiffimplicit-differentiation