Learning to Learn with Contrastive Meta-Objective

Code implementation for paper Learning to Learn with Contrastive Meta-Objective (NeurIPS 2025 Oral) (https://arxiv.org/abs/2410.05975)

Few-Shot Image Classification:

optimization-based/metric-based:

• install learn2learn (https://learn2learn.net/);
• run "ml/mamlc_miniimagenet.py", "ml/fomamlc_miniimagenet.py", "ml/mamlc_tieredimagenet.py", "ml/protonetc_tieredimagenet.py", "ml/reptilec_miniimagenet.py", "ml/reptilec_miniimagenet.py" following the instructions in "README.md" in https://github.com/learnables/learn2learn/tree/master/examples/vision.

amortization-based:

• get public code "simple-cnaps" (https://github.com/peymanbateni/simple-cnaps);
• copy "ml/runc_simple_cnaps_mt.py" to "simple-cnaps/simple-cnaps-src/", and run following the corresponding instructions in "simple-cnaps".

In-Context Learning:

• get public code "in-context-learning" (https://github.com/dtsip/in-context-learning)
• copy "icl/ctrain.py" and "icl/ceval.py" to "in-context-learning/src/", and copy "icl/ceval.py" to "in-context-learning/src/conf/";
• refer to the corresponding instructions in "in-context-learning" and run python ctrain.py --config conf/ctoy.yaml for training ConML-ICL, and python ceval.py for testing.