git clone git@github.com:V0XNIHILI/torch-mate.git
cd torch-mate
# Make sure you have pip 23 or higher
pip install -e .TorchMate consists of a few different modules:
torch_mate/data: A collection of datasets, samplers, transforms and utilities for special use cases (i.e. few-shot learning, pre-loading datasets, siamese approaches).torch_mate/models: A collection of models that can be used to quickly prototype and test different architectures.torch_mate/utils: A collection of utilities that can be used to simplify common tasks when working with CUDA devices and PyTorch.
This project is licensed under the MIT License - see the LICENSE file for details.
--
meta-learning unsupervised learning pruning quantization knowledge distillation
Always one training method but maybe multiple validation methods
THINK ABOUT LOGGING via {"loss": loss} (returning from training_step for example) how to make specialized pytorch lightning data modules how to make forward pass easily changable how to indicate which algorithm is used in which step log number of parameters look at all TODOs (# TODO) Should generic train step and test step be configurable via hparams?
Different layers for different passes Different algo in forward
Solutions Special forward
P>M>F F = copy model for each task and finetune it