In this assignment you are asked to implement 3 algorithms for few-shot learning and compare their performance between each other:
This assignment assumes that you are familiar with these algorithms and pytorch. A GPU card is not strictly required, but highly recommended. Otherwise, experimenting and debugging will take you a lot of time.
The unpretrained baseline is already implemented for you and you can launch it by calling
CUDA_VISIBLE_DEVICES=0 python run_experiment.py -m unpretrained_baselinePlease refer to the instructions file.
Method hyperparameters are supposed to be stored in utils/config.py.
The main entry-point is run_experiment.py which launches experiments.
The most "diffucult" file in the template is utils/data.py which loads data and generates training episodes.
This episode generation is quite cumbersome, but it shouldn't be necessary to dive too deep into it.
Each training episode consists on ds_train and ds_test which are used to train the model on a given task.
ds_test is used for outer loop update only, hence MAML is the only method that uses it.
While implementing MAML, you may face the problem when using nn.Module since we need to be able to perform a forward pass with parameter values that are different from what a module currently has.
That's why we provided models/pure_layers.py for you and PureLeNet class in models/lenet.py.
You may like to read this article to get the idea of pure functions in deep learning.
You are also encouraged to check open-source MAML implementations to get the idea of what the problem with nn.Module is and why we need to reinvent this wheel.
trainers/trainer.py is responsible for training/evaluating baselines.
Since there is a lot of code sharing, ProtoNet and MAML trainers inherit from it and override several things.
Other files should be self-explanatory.
You can search for TODO keyword to see what pieces are missing.
Create an issue or send a PR if you found a bug or want to propose an update.