73from typing import Optional, Tuple
74
75import torch
76from torch import nn
77
78from labml_helpers.module import Module
79from labml_nn.lstm import LSTMCell

82class HyperLSTMCell(Module):

90    def __init__(self, input_size: int, hidden_size: int, hyper_size: int, n_z: int):

108        super().__init__()

121        self.hyper = LSTMCell(hidden_size + input_size, hyper_size, layer_norm=True)

127        self.z_h = nn.Linear(hyper_size, 4 * n_z)

129        self.z_x = nn.Linear(hyper_size, 4 * n_z)

131        self.z_b = nn.Linear(hyper_size, 4 * n_z, bias=False)

134        d_h = [nn.Linear(n_z, hidden_size, bias=False) for _ in range(4)]
135        self.d_h = nn.ModuleList(d_h)

137        d_x = [nn.Linear(n_z, hidden_size, bias=False) for _ in range(4)]
138        self.d_x = nn.ModuleList(d_x)

140        d_b = [nn.Linear(n_z, hidden_size) for _ in range(4)]
141        self.d_b = nn.ModuleList(d_b)

144        self.w_h = nn.ParameterList([nn.Parameter(torch.zeros(hidden_size, hidden_size)) for _ in range(4)])

146        self.w_x = nn.ParameterList([nn.Parameter(torch.zeros(hidden_size, input_size)) for _ in range(4)])

149        self.layer_norm = nn.ModuleList([nn.LayerNorm(hidden_size) for _ in range(4)])
150        self.layer_norm_c = nn.LayerNorm(hidden_size)

152    def forward(self, x: torch.Tensor,
153                h: torch.Tensor, c: torch.Tensor,
154                h_hat: torch.Tensor, c_hat: torch.Tensor):

161        x_hat = torch.cat((h, x), dim=-1)

163        h_hat, c_hat = self.hyper(x_hat, h_hat, c_hat)

166        z_h = self.z_h(h_hat).chunk(4, dim=-1)

168        z_x = self.z_x(h_hat).chunk(4, dim=-1)

170        z_b = self.z_b(h_hat).chunk(4, dim=-1)

173        ifgo = []
174        for i in range(4):

176            d_h = self.d_h[i](z_h[i])

178            d_x = self.d_x[i](z_x[i])

185            y = d_h * torch.einsum('ij,bj->bi', self.w_h[i], h) + \
186                d_x * torch.einsum('ij,bj->bi', self.w_x[i], x) + \
187                self.d_b[i](z_b[i])
188
189            ifgo.append(self.layer_norm[i](y))

192        i, f, g, o = ifgo

195        c_next = torch.sigmoid(f) * c + torch.sigmoid(i) * torch.tanh(g)

198        h_next = torch.sigmoid(o) * torch.tanh(self.layer_norm_c(c_next))
199
200        return h_next, c_next, h_hat, c_hat

203class HyperLSTM(Module):

208    def __init__(self, input_size: int, hidden_size: int, hyper_size: int, n_z: int, n_layers: int):

213        super().__init__()

216        self.n_layers = n_layers
217        self.hidden_size = hidden_size
218        self.hyper_size = hyper_size

222        self.cells = nn.ModuleList([HyperLSTMCell(input_size, hidden_size, hyper_size, n_z)] +
223                                   [HyperLSTMCell(hidden_size, hidden_size, hyper_size, n_z) for _ in
224                                    range(n_layers - 1)])

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

234        n_steps, batch_size = x.shape[:2]

237        if state is None:
238            h = [x.new_zeros(batch_size, self.hidden_size) for _ in range(self.n_layers)]
239            c = [x.new_zeros(batch_size, self.hidden_size) for _ in range(self.n_layers)]
240            h_hat = [x.new_zeros(batch_size, self.hyper_size) for _ in range(self.n_layers)]
241            c_hat = [x.new_zeros(batch_size, self.hyper_size) for _ in range(self.n_layers)]

243        else:
244            (h, c, h_hat, c_hat) = state

248            h, c = list(torch.unbind(h)), list(torch.unbind(c))
249            h_hat, c_hat = list(torch.unbind(h_hat)), list(torch.unbind(c_hat))

252        out = []
253        for t in range(n_steps):

255            inp = x[t]

257            for layer in range(self.n_layers):

259                h[layer], c[layer], h_hat[layer], c_hat[layer] = \
260                    self.cells[layer](inp, h[layer], c[layer], h_hat[layer], c_hat[layer])

262                inp = h[layer]

264            out.append(h[-1])

267        out = torch.stack(out)
268        h = torch.stack(h)
269        c = torch.stack(c)
270        h_hat = torch.stack(h_hat)
271        c_hat = torch.stack(c_hat)

274        return out, (h, c, h_hat, c_hat)