File size: 7,217 Bytes
2b21abc |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 |
import torch
import torch.nn as nn
class CausalLSTMCell(nn.Module):
def __init__(self, in_channel, num_hidden, height, width, filter_size, stride, layer_norm):
super(CausalLSTMCell, self).__init__()
self.num_hidden = num_hidden
self.padding = filter_size // 2
self._forget_bias = 1.0
if layer_norm:
self.conv_x = nn.Sequential(
nn.Conv2d(in_channel, num_hidden * 7, kernel_size=filter_size,
stride=stride, padding=self.padding, bias=False),
nn.LayerNorm([num_hidden * 7, height, width])
)
self.conv_h = nn.Sequential(
nn.Conv2d(num_hidden, num_hidden * 4, kernel_size=filter_size,
stride=stride, padding=self.padding, bias=False),
nn.LayerNorm([num_hidden * 4, height, width])
)
self.conv_c = nn.Sequential(
nn.Conv2d(num_hidden, num_hidden * 3, kernel_size=filter_size,
stride=stride, padding=self.padding, bias=False),
nn.LayerNorm([num_hidden * 3, height, width])
)
self.conv_m = nn.Sequential(
nn.Conv2d(num_hidden, num_hidden * 3, kernel_size=filter_size,
stride=stride, padding=self.padding, bias=False),
nn.LayerNorm([num_hidden * 3, height, width])
)
self.conv_o = nn.Sequential(
nn.Conv2d(num_hidden * 2, num_hidden, kernel_size=filter_size,
stride=stride, padding=self.padding, bias=False),
nn.LayerNorm([num_hidden, height, width])
)
self.conv_c2m = nn.Sequential(
nn.Conv2d(num_hidden, num_hidden * 4, kernel_size=filter_size,
stride=stride, padding=self.padding, bias=False),
nn.LayerNorm([num_hidden * 4, height, width])
)
self.conv_om = nn.Sequential(
nn.Conv2d(num_hidden, num_hidden, kernel_size=filter_size,
stride=stride, padding=self.padding, bias=False),
nn.LayerNorm([num_hidden, height, width])
)
else:
self.conv_x = nn.Sequential(
nn.Conv2d(in_channel, num_hidden * 7, kernel_size=filter_size,
stride=stride, padding=self.padding, bias=False),
)
self.conv_h = nn.Sequential(
nn.Conv2d(num_hidden, num_hidden * 4, kernel_size=filter_size,
stride=stride, padding=self.padding, bias=False),
)
self.conv_c = nn.Sequential(
nn.Conv2d(num_hidden, num_hidden * 3, kernel_size=filter_size,
stride=stride, padding=self.padding, bias=False),
)
self.conv_m = nn.Sequential(
nn.Conv2d(num_hidden, num_hidden * 3, kernel_size=filter_size,
stride=stride, padding=self.padding, bias=False),
)
self.conv_o = nn.Sequential(
nn.Conv2d(num_hidden * 2, num_hidden, kernel_size=filter_size,
stride=stride, padding=self.padding, bias=False),
)
self.conv_c2m = nn.Sequential(
nn.Conv2d(num_hidden, num_hidden * 4, kernel_size=filter_size,
stride=stride, padding=self.padding, bias=False),
)
self.conv_om = nn.Sequential(
nn.Conv2d(num_hidden, num_hidden, kernel_size=filter_size,
stride=stride, padding=self.padding, bias=False),
)
self.conv_last = nn.Conv2d(num_hidden * 2, num_hidden, kernel_size=1,
stride=1, padding=0, bias=False)
def forward(self, x_t, h_t, c_t, m_t):
x_concat = self.conv_x(x_t)
h_concat = self.conv_h(h_t)
c_concat = self.conv_c(c_t)
m_concat = self.conv_m(m_t)
i_x, f_x, g_x, i_x_prime, f_x_prime, g_x_prime, o_x = \
torch.split(x_concat, self.num_hidden, dim=1)
i_h, f_h, g_h, o_h = torch.split(h_concat, self.num_hidden, dim=1)
i_m, f_m, m_m = torch.split(m_concat, self.num_hidden, dim=1)
i_c, f_c, g_c = torch.split(c_concat, self.num_hidden, dim=1)
i_t = torch.sigmoid(i_x + i_h + i_c)
f_t = torch.sigmoid(f_x + f_h + f_c + self._forget_bias)
g_t = torch.tanh(g_x + g_h + g_c)
c_new = f_t * c_t + i_t * g_t
c2m = self.conv_c2m(c_new)
i_c, g_c, f_c, o_c = torch.split(c2m, self.num_hidden, dim=1)
i_t_prime = torch.sigmoid(i_x_prime + i_m + i_c)
f_t_prime = torch.sigmoid(f_x_prime + f_m + f_c + self._forget_bias)
g_t_prime = torch.tanh(g_x_prime + g_c)
m_new = f_t_prime * torch.tanh(m_m) + i_t_prime * g_t_prime
o_m = self.conv_om(m_new)
o_t = torch.tanh(o_x + o_h + o_c + o_m)
mem = torch.cat((c_new, m_new), 1)
h_new = o_t * torch.tanh(self.conv_last(mem))
return h_new, c_new, m_new
class GHU(nn.Module):
def __init__(self, in_channel, num_hidden, height, width, filter_size,
stride, layer_norm, initializer=0.001):
super(GHU, self).__init__()
self.filter_size = filter_size
self.padding = filter_size // 2
self.num_hidden = num_hidden
self.layer_norm = layer_norm
if layer_norm:
self.z_concat = nn.Sequential(
nn.Conv2d(in_channel, num_hidden * 2, kernel_size=filter_size,
stride=stride, padding=self.padding, bias=False),
nn.LayerNorm([num_hidden, height, width])
)
self.x_concat = nn.Sequential(
nn.Conv2d(in_channel, num_hidden * 2, kernel_size=filter_size,
stride=stride, padding=self.padding, bias=False),
nn.LayerNorm([num_hidden, height, width])
)
else:
self.z_concat = nn.Sequential(
nn.Conv2d(in_channel, num_hidden * 2, kernel_size=filter_size,
stride=stride, padding=self.padding, bias=False),
)
self.x_concat = nn.Sequential(
nn.Conv2d(in_channel, num_hidden * 2, kernel_size=filter_size,
stride=stride, padding=self.padding, bias=False),
)
if initializer != -1:
self.initializer = initializer
self.apply(self._init_weights)
def _init_weights(self, m):
if isinstance(m, (nn.Conv2d)):
nn.init.uniform_(m.weight, -self.initializer, self.initializer)
def _init_state(self, inputs):
return torch.zeros_like(inputs)
def forward(self, x, z):
if z is None:
z = self._init_state(x)
z_concat = self.z_concat(z)
x_concat = self.x_concat(x)
gates = x_concat + z_concat
p, u = torch.split(gates, self.num_hidden, dim=1)
p = torch.tanh(p)
u = torch.sigmoid(u)
z_new = u * p + (1-u) * z
return z_new
|