ICNN.py

'''
Copyright 2020 Amanpreet Singh,
               Martin Bauer,
               Sarang Joshi

Redistribution and use in source and binary forms, with or without modification, are permitted provided that the following conditions are met:

1. Redistributions of source code must retain the above copyright notice, this list of conditions and the following disclaimer.

2. Redistributions in binary form must reproduce the above copyright notice, this list of conditions and the following disclaimer in the documentation and/or other materials provided with the distribution.

THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.

'''


from torch import nn
import torch

'''
Custom Activation function used in the CINNs
Right now this is :     0  for x <= 0
                        x^(2.5) for x > 0
'''
class customActivation(nn.Module):
    def __init(self):
        super.__init__()

    def forward(self, x):
        x[x <= 0] = 0

        return x**2.5

'''
Custom Activation function used in the CINNs
Right now this is :    Softplus ^ (1.1)
'''
class softplus_power(nn.Module):
    def __init(self):
        super.__init__()

    def forward(self, x):
        m = nn.Softplus()
        x = m(x)
        return x**1.1


'''
Simple Fully connected network with some special functions as defined below.
Parameters :        n_inputs --- number of inputs
                    n_outputs -- number of outputs
                    n_layers --- number of hidden layers
                    n_npl ------ number of nodes per hidden layer
'''
class ICNN(nn.Module):
    def __init__(self, n_inputs, n_outputs, n_layers, n_npl):
        super().__init__()
        self.n_inps = n_inputs
        if n_layers == 0:
            self.layers = nn.ModuleList()
            self.activations = nn.ModuleList()
            self.layers.append(nn.Linear(n_inputs, n_outputs))
        else:
            self.layers = nn.ModuleList()
            self.activations = nn.ModuleList()
            self.layers.append(nn.Linear(n_inputs, n_npl))
            self.activations.append(softplus_power())
            for x in range(n_layers-1):
                self.layers.append(nn.Linear(n_npl + n_inputs, n_npl))
                self.activations.append(softplus_power())
            self.layers.append(nn.Linear(n_npl + n_inputs, n_outputs))

    '''
    Forward pass - Simple fully connected network with given activations
    '''
    def forward(self, x):
        inps = x
        out = x
        for i in range(len(self.layers)-1):
            if i > 0:
                out = torch.cat((out, inps), dim=1)
            out = self.layers[i](out)
            out = self.activations[i](out)

        if len(self.layers) != 1:
            out = torch.cat((out, inps), dim=1)

        out = self.layers[-1](out)
        return out

    '''
    Makes the all the weights positive. 
    Refer to : https://arxiv.org/pdf/1609.07152.pdf
    '''
    def make_convex(self):
        with torch.no_grad():
            for i in range(1, len(self.layers)):
                if i == 0:
                    self.layers[i].weight[:, :][self.layers[i].weight[:, :] < 0] = \
                        torch.abs(self.layers[i].weight[:, :][self.layers[i].weight[:, :] < 0])
                self.layers[i].weight[:, :-self.n_inps][self.layers[i].weight[:, :-self.n_inps] < 0] = \
                    torch.abs(self.layers[i].weight[:, :-self.n_inps][self.layers[i].weight[:, :-self.n_inps] < 0])

    '''
    Sets all negative weights (W^z) to 0
    '''
    def project(self):
        with torch.no_grad():
            for i in range(1, len(self.layers)):
                if i == 0:
                    self.layers[i].weight[:, :][self.layers[i].weight[:, :] < 0] = 0
                self.layers[i].weight[:, :-self.n_inps][self.layers[i].weight[:, :-self.n_inps] < 0] = 0

    '''
    Saves the model in the specified path and name with .pth extension
    '''
    def save(self, path, model_name):
        torch.save(self.state_dict(), path + model_name + '.pth')

    '''
    Loads the model given its path.
    Different Architecture or activation function (with parameters) in the saved model will throw an error
    '''
    def load(self, path):
        self.load_state_dict(torch.load(path))