Model-Agnostic Meta-Learning with transformers

Model-Agnostic Meta-Learning applied on layers of transformers.

Dependencies

This code requires the following:

• cuda 9
• Python >= 3.5
• Tensorflow >= 1.10

Data

ARSC Amazon Customer Reviews Sentiment Classification
The dataset comprises reviews for 23 types of products on Amazon, and for each domain, there are 3 different binary classification tasks. The three tasks in each domain are combined so that there are enough data for each class, so in total there are 23 different tasks/domains. 18 domains in data/meta_train_tasks.list are used for meta-testing and 4 domains in data/meta_test_tasks.list are used for meta-training.

Train

download bert
Run train.py.