Using Direct Feedback Alignment for training Neural Ordinary Differential Equations

Requirements : torchdiffeq. (pip install torchdiffeq)

2 fichiers éxécutables pour 2 expériences + pdf explication et résulats.

Spiral dataset test

usage: python ode_dfa_spiral_test.py [-h] [--method {DFA,adjoint}] [--data_size DATA_SIZE] [--batch_time BATCH_TIME] [--batch_size BATCH_SIZE] [--niters NITERS] [--test_freq TEST_FREQ] [--viz] [--gpu GPU]

MNIST classification test

Problème de fonctionnement ici de DFA dans le code, refaire la partie backward DFA.

usage : python ode_dfa_MNIST_test.py [-h] [--DFA {True,False}] [--name_model NAME_MODEL] [--adjoint {True,False}] [--nepochs NEPOCHS] [--data_aug {True,False}] [--lr LR] [--w_decay W_DECAY] [--batch_size BATCH_SIZE] [--test_batch_size TEST_BATCH_SIZE] [--gpu GPU]

where :

• DFA : if True, we use DFA to do the back propagation in the training, default=True
• save_model : if True we save the model that we learned, default=True
• NAME_MODEL : the name of the model that we learned (if we save it),default='ode_dfa_MNIST.pt'
• adjoint : if True ,default=False
• nepochs : number of epochs in the training, default=1
• data_aug : if True we do data augmentation,default=True
• lr : learning rate,default=0.001
• w_decay : weight decay, default=0.1
• batch_size: size of a batch for the training, default=128
• test_batch_size : size of a batch for the test,default=1000
• gpu : number of the gpu that we will use, if gpu is not available we use cpu,default=1