search

baiksung / MeTAL

Official PyTorch implementation of "Meta-Learning with Task-Adaptive Loss Function for Few-Shot Learning" (ICCV2021 Oral)
62 stars 13 forks source link

could your code run with maml++ and some config in your code i want to know its usage. #9

Open LinGinQiu opened 1 year ago

LinGinQiu commented 1 year ago

HI baiksung could your code run with maml++, if so, could your tell the config i should set? I would really like to know how long it takes to run this code. Because when I run your code as a comparison model, , it takes so long, even more than original maml(one-order). If it caused by the set the "second_order": true, or the "meta_loss": true. AND i also want to know what does the set learnable_per_layer_per_step_inner_loop_learning_rate to do ?

Best regards LGQ

LinGinQiu commented 1 year ago

by the way, i use single 2080ti, and i am sure GPU is running

baiksung commented 1 year ago

Dear LinGinQiu,

Since the default setting is set for second-order training, it will be much slower than first-order MAML. Compared to MAML (second-order), MeTAL is ~17% slower but it converges much faster. learnable_per_layer_per_step_inner_loop_learning_rate is not used in MeTAL but it's a legacy code coming from MAML++ since this code is mostly based on MAML++ public code. For more details on that, I suggest taking a look at MAML++. Although this version does not provide a json file for training with MAML++, you can modify MeTAL+ALFA to make it trainable with MAML++.

Best, Sungyong