networks.py

"""
General networks for pytorch.

Algorithm-specific networks should go else-where.
"""
import torch
from torch import nn as nn
from torch.nn import functional as F

import utils.pytorch_util as ptu
from utils.core import eval_np


def identity(x):
    return x


class Mlp(nn.Module):
    def __init__(
            self,
            hidden_sizes,
            output_size,
            input_size,
            init_w=3e-3,
            hidden_activation=F.relu,
            output_activation=identity,
            hidden_init=ptu.fanin_init,
            b_init_value=0.1,
    ):
        super().__init__()

        self.input_size = input_size
        self.output_size = output_size
        self.hidden_activation = hidden_activation
        self.output_activation = output_activation
        self.fcs = []
        in_size = input_size

        for i, next_size in enumerate(hidden_sizes):
            fc = nn.Linear(in_size, next_size)
            in_size = next_size
            hidden_init(fc.weight)
            fc.bias.data.fill_(b_init_value)
            self.__setattr__("fc{}".format(i), fc)
            self.fcs.append(fc)

        self.last_fc = nn.Linear(in_size, output_size)
        self.last_fc.weight.data.uniform_(-init_w, init_w)
        self.last_fc.bias.data.uniform_(-init_w, init_w)

    def forward(self, input, return_preactivations=False):
        h = input
        for i, fc in enumerate(self.fcs):
            h = fc(h)
            h = self.hidden_activation(h)
        preactivation = self.last_fc(h)
        output = self.output_activation(preactivation)
        if return_preactivations:
            return output, preactivation
        else:
            return output


class FlattenMlp(Mlp):
    """
    Flatten inputs along dimension 1 and then pass through MLP.
    """

    def forward(self, *inputs, **kwargs):
        flat_inputs = torch.cat(inputs, dim=1)
        return super().forward(flat_inputs, **kwargs)