ICL

This is the code for ICL.

Requirements

• Python 3.6
• numpy==1.19.2
• sklearn==0.23.2
• torch==1.6.0
• torch-geometric==1.6.3
• torchmeta==1.7.0
• torchvision==0.7.0

Data

The data should be placed in the path ./data/ which should be the same with the code in dataset.py. For example, ./data/cv/ImageNet-2/.

ICL

We provide four variants of ICL:

• ICL: train_ICL_MAML_RL.py
• ICL (w/o LRL): train_ICL_MAML.py
• ICL (w/o M): train_ICL_RL.py
• ICL (w/o M+LRL): train_ICL.py

Usage

Commands for training and testing the method on CV and GRL datasets.

Train a resnet18 on ImageNet-2:

python -u main.py -m cv -d ImageNet-2 -en resnet18 -a 0.5

Train a gcn on PROTEINS_full:

python -u main.py -m graph -d PROTEINS_full -en gcn -a 0.5

More detailed parameter information please refer to main.py.
We provide some key parameter description:

Name	Description
mode	The mode of ICL: cv or graph (defaults to cv).
dataset	The dataset name (defaults to ImageNet-2).
encoder	The encoder: resnet18 or gcn (defaults to resnet18)
alpha	The growth rate of the data $\alpha$ (defaults to 0.5).

Output

We train the different variants on the same encoder trained on the old data, and report the training time, converge epoch, and the accuracy. For instance:

Begin!
...
Inference!
Epoch 0    , Loss 2.4097, time 0.8747, total time 0.8747
Epoch 10   , Loss 2.3572, time 0.6263, total time 7.0705
...
Eval!
Final Epoch 31   , Loss 1.7602, time 0.6178, total time 19.9385, acc 0.7187, acc-old 0.6961, acc-new 0.7108
ICL-MAML-RL!
Epoch 0    , Loss 2.9381, time 1.8081, total time 1.8081
Epoch 10   , Loss 2.8510, time 1.7075, total time 16.2273
...
Done!