Code to accompany the papers titled "When are equilibrium networks scoring algorithms?" and "Deep equilibrium models as estimators for continuous latent variables"

We roughly follow an sklearn API. For example, from input dimensionality 50 to latent dimensionality 2

from model import DeepPED
ped = DeepPED([50, 2])
latents = ped.fit_transform(data)

Here data is a Pytorch torch.utils.data.DataLoader.

All hyperparameters are set in the fit or fit_transform method, except for the layer widths, which are set at initialisation.

Available keyword hyperparameters are

• lamb strictly positive float, which is the lambda in the paper. For implementation, all lambda in each layer is equal.
• dist a string representing the combination of A and R. Currently implemented choices when R is the identity are 'bernoulli', 'binomial', 'cts-bernoulli', 'gauss' and 'poisson'. Also available is 'relu', which is a Gaussian likelihood and ReLU R.
• weight_decay a nonnegative float representing a factor for the amount of weight decay to include. This is equivalent to L2 regularisation i.e. a (truncated) Gaussian prior over the weights. The coefficient of the L2 regulariser in layer l is weight_decay * layer_width[l-1], where l is the index of the layer (starting at zero). So for example for a [50, 30, 15, 2] network, the amount of L2 regularisation in the last latent layer parameters is 50*weight_decay.
• num_epochs integer number of training epochs to run Adam optimiser for
• lr strictly positive learning rate for Adam optimiser
• data_loader_test a Pytorch torch.utils.data.DataLoader to project into latent space. If None, use the training data.
• layer_out the latent layer to output, with negative integer index. So -1 is the last layer, -2 is the second last layer, and so on.
• plot_freq and plot_bool are for development purposes. During training, plot some debugging every plot_freq epochs if plot_bool is True.

We provide a script called script_synthetic.py, which includes all the code required to reproduce the results reported in the paper. Each time this script is run, it completes one run of one of the rows in table 4 (of which 100 are reported in the paper). In order to reproduce results, modify the following global variables as desired

DIST_TRUE = ... #  'poisson' or 'gauss' or 'bernoulli'
DIMS_TRUE = ... #[50, 2]

There are some other globals there to play with as well, if you wish. You will need to run each setting 100 times, the execution of which is system dependent (e.g. with a job manager) and left up to the user. An example for a system using SLURM is provided, see synthetic_run_all.sh and synthetic_run.sh.