amatorch is a Python implementation of Accuracy Maximization Analysis (AMA),
a supervised feature learning method based on maximizing the accuracy
of a probabilistic decoder for a given task. AMA has been primarily
used to obtain image-computable ideal observers for various visual tasks
to study human vision.
amatorch currently implements the AMA-Gauss variant of the AMA model
(more variants are in development). It also includes functions to
train and test the model.
AMA models have two main components: an encoding stage with learnable filters, and a decoding stage that uses filter response distributions to estimate the stimulus class.
An AMA model can be initialized with a random set of filters and
trained with amatorch as follows:
import amatorch.optim as optim
from amatorch.datasets import disparity_data
from amatorch.models import AMAGauss
data = disparity_data()
ama = AMAGauss(
stimuli=data["stimuli"],
labels=data["labels"],
n_filters=2,
response_noise=0.05,
c50=1.0,
)
# Fit model
loss, training_time = optim.fit(
model=ama,
stimuli=data["stimuli"],
labels=data["labels"],
epochs=20,
batch_size=512,
learning_rate=0.1,
decay_step=4,
decay_rate=0.5,
)The resulting model can be used to obtain posterior probabilities of the classes for a given stimulus and the estimated class.
posterior = ama.get_posteriors(data["stimuli"])
estimated_class = ama.get_estimates(data["stimuli"])See the tutorials for more details on the model structure, its variants and its usage.
To install the package, clone the repository, go to the downloaded directory and install using pip. In the command line, this can be done as follows:
git clone [email protected]:dherrera1911/amatorch.git
cd amatorch
pip install -e .We recommend installing the package in a virtual
environment (e.g. using conda). For more detailed instructions, see the
installation section of the tutorials.