User Manual

The user Manual describes the usage of the user-facing functionalities of the project.

Training and testing a single setup on the MNIST dataset

The training takes the configurations from the specified json file and trains a single model on the MNIST dataset. The in-between results are saved for further testing

The testing takes the saved configuration and gets the loss and classification ratio on the MNIST. Note: If the dataset is not present then the code tries download it from the database behind PyTorch.

Command for training:

python -m pseudo_backprop.experiments.train_mnist --params model_config.json

Command for testing:

python -m pseudo_backprop.experiments.test_mnist --params model_config.json

An example for the model_config.json file is:

{
    "batch_size": 20,
    "layers": [784, 1000, 1000, 10],
    "epochs": 120,
    "model_folder": "model_gen_pseudo",
    "model_type": "gen_pseudo",
    "learning_rate": 0.00005,
    "momentum": 0.0,
    "weight_decay": 0.0,
    "random_seed": 456732,
    "pinverse_recalc": 500,
    "gen_samples": 60000,
    "dataset": "mnist",
    "dataset_path": "../../../../dataset/mnist"
}

Entry	Type	Meaning
batch_size	number	size of minibatch; number of examples used per iteration
layers	list of integers	Number of neurons in the layers. The first layer is the input.
epochs	integer	Length of training. One epoch is one training on the whole training set
model_folder	str, relative path	folder to create the output to
model_type	string	The used model. See explanation below.
learning_rate	float	the learning rate
momentum	float	the momentum term in the pytorch sgd grad descent optimization
weight_decay	float	weight decay term with L2 regularization
random_seed	int	Random number seed to be set
pinverse_recalc	int	After how many iterations will be the (generalized) pseudoinverse recalculated
gen_samples	integer	how many examples are used for calculating the generalized pseudo-inverse
dataset	string	Dataset to be used: mnist or cifar10
dataset_path	string, relative path	path to the dataset, if not found it will be downloaded

The potential entries for the model_type are:

• backprop: the standard backpropagation algorithm.
• fa : feedback alignment following to Lillicrap et al. (2016). The backward weights are drawn from the same probability distribution as the forward weights.
• pseudo_backprop : Pseudobackprop algorithm. The backward weights are set as the Moore-Penrose Pseudoinverse.
• gen_pseudo : Generalized pseudobackprop algorithm. The backward weights are set as the generalized pseudoinverse of the forward weights.

The pinverse_recalc is only applicable for the pseudo backrpop and the generalized pseudobackprop algorithm, otherwise they are ignored. The gen_samples is only applicable for the generalized pseudoinverse algorithm, otherwise it is ignored.

Settings for the repeated experiments

An example for the master/exp_control.json is:

{
    "random_seed": 229892,
    "seed_start": 452391,
    "repetitions": 10,
    "simulator": "css",
    "test_per_images": 60000
}

The parameters are:

Entry	Type	Meaning
random_seed	number	random seed for generating the random seeds of the single repetitions
seed_start	number	Starting point for the seed of the single repetitions. Random numbers are added to this to generate the further seeds.
repetitions	integer	Length of training. One epoch is one training on the whole training set
simulator	string	name of the simulator. See below
test_per_images	number	Sets how often the tests will be executed after the training has finished. E.g. here after every 60000 images

The potential simulator parameters are:

• direct: The simulation is started in the same environment as a subprocess as the outer command
• css: The simulations is started using the master/slurm_script.sh as a starting slurm script