#

Recurrent Highway Networks

This is a PyTorch implementation of Recurrent Highway Networks.

11from typing import Optional
12
13import torch
14from torch import nn
15
16from labml_helpers.module import Module

#

Recurrent Highway Network Cell

This implements equations $(6) - (9)$ .

$s_{d}^{t} = h_{d}^{t} ⊙ g_{d}^{t} + s_{d - 1}^{t} ⊙ c_{d}^{t}$

where

h_{0}^{t} g_{0}^{t} c_{0}^{t} = tanh (l i n_{h x} (x) + l i n_{h s} (s_{D}^{t - 1})) = σ (l i n_{g x} (x) + l i n_{g s}^{1} (s_{D}^{t - 1})) = σ (l i n_{c x} (x) + l i n_{cs}^{1} (s_{D}^{t - 1}))

and for $0 < d < D$

h_{d}^{t} g_{d}^{t} c_{d}^{t} = tanh (l i n_{h s}^{d} (s_{d}^{t})) = σ (l i n_{g s}^{d} (s_{d}^{t})) = σ (l i n_{cs}^{d} (s_{d}^{t}))

$⊙$ stands for element-wise multiplication.

Here we have made a couple of changes to notations from the paper. To avoid confusion with time, gate is represented with $g$ , which was $t$ in the paper. To avoid confusion with multiple layers we use $d$ for depth and $D$ for total depth instead of $l$ and $L$ from the paper.

We have also replaced the weight matrices and bias vectors from the equations with linear transforms, because that's how the implementation is going to look like.

We implement weight tying, as described in paper, $c_{d}^{t} = 1 - g_{d}^{t}$ .

19class RHNCell(Module):

#

input_size is the feature length of the input and hidden_size is the feature length of the cell. depth is $D$ .

57    def __init__(self, input_size: int, hidden_size: int, depth: int):

#

63        super().__init__()
64
65        self.hidden_size = hidden_size
66        self.depth = depth

#

We combine $l i n_{h s}$ and $l i n_{g s}$ , with a single linear layer. We can then split the results to get the $l i n_{h s}$ and $l i n_{g s}$ components. This is the $l i n_{h s}^{d}$ and $l i n_{g s}^{d}$ for $0 \leq d < D$ .

70        self.hidden_lin = nn.ModuleList([nn.Linear(hidden_size, 2 * hidden_size) for _ in range(depth)])

#

Similarly we combine $l i n_{h x}$ and $l i n_{g x}$ .

73        self.input_lin = nn.Linear(input_size, 2 * hidden_size, bias=False)

#

x has shape [batch_size, input_size] and s has shape [batch_size, hidden_size] .

75    def forward(self, x: torch.Tensor, s: torch.Tensor):

#

Iterate $0 \leq d < D$

82        for d in range(self.depth):

#

We calculate the concatenation of linear transforms for $h$ and $g$

84            if d == 0:

#

The input is used only when $d$ is $0$ .

86                hg = self.input_lin(x) + self.hidden_lin[d](s)
87            else:
88                hg = self.hidden_lin[d](s)

#

Use the first half of hg to get $h_{d}^{t}$

h_{0}^{t} h_{d}^{t} = tanh (l i n_{h x} (x) + l i n_{h s} (s_{D}^{t - 1})) = tanh (l i n_{h s}^{d} (s_{d}^{t}))

96            h = torch.tanh(hg[:, :self.hidden_size])

#

Use the second half of hg to get $g_{d}^{t}$

g_{0}^{t} g_{d}^{t} = σ (l i n_{g x} (x) + l i n_{g s}^{1} (s_{D}^{t - 1})) = σ (l i n_{g s}^{d} (s_{d}^{t}))

103            g = torch.sigmoid(hg[:, self.hidden_size:])
104
105            s = h * g + s * (1 - g)
106
107        return s

#

Multilayer Recurrent Highway Network

110class RHN(Module):

#

Create a network of n_layers of recurrent highway network layers, each with depth depth , $D$ .

115    def __init__(self, input_size: int, hidden_size: int, depth: int, n_layers: int):

#

120        super().__init__()
121        self.n_layers = n_layers
122        self.hidden_size = hidden_size

#

Create cells for each layer. Note that only the first layer gets the input directly. Rest of the layers get the input from the layer below

125        self.cells = nn.ModuleList([RHNCell(input_size, hidden_size, depth)] +
126                                   [RHNCell(hidden_size, hidden_size, depth) for _ in range(n_layers - 1)])

#

x has shape [seq_len, batch_size, input_size] and state has shape [batch_size, hidden_size] .

128    def forward(self, x: torch.Tensor, state: Optional[torch.Tensor] = None):

#

133        time_steps, batch_size = x.shape[:2]

#

Initialize the state if None

136        if state is None:
137            s = [x.new_zeros(batch_size, self.hidden_size) for _ in range(self.n_layers)]
138        else:

#

Reverse stack the state to get the state of each layer

📝 You can just work with the tensor itself but this is easier to debug

142            s = torch.unbind(state)

#

Array to collect the outputs of the final layer at each time step.

145        out = []

#

Run through the network for each time step

148        for t in range(time_steps):

#

Input to the first layer is the input itself

150            inp = x[t]

#

Loop through the layers

152            for layer in range(self.n_layers):

#

Get the state of the layer

154                s[layer] = self.cells[layer](inp, s[layer])

#

Input to the next layer is the state of this layer

156                inp = s[layer]

#

Collect the output of the final layer

158            out.append(s[-1])

#

Stack the outputs and states

161        out = torch.stack(out)
162        s = torch.stack(s)
163
164        return out, s