File size: 7,569 Bytes
fa128ec |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 |
import numpy as np
import torch
import torch.nn as nn
from torch.nn.init import xavier_uniform
class KAE(nn.Module):
def __init__(self, w=32, h=32, c=1, nb_hidden=300, nb_active=16):
super().__init__()
self.nb_hidden = nb_hidden
self.nb_active = nb_active
self.encode = nn.Sequential(
nn.Linear(w*h*c, nb_hidden, bias=False)
)
self.bias = nn.Parameter(torch.zeros(w*h*c))
self.params = nn.ParameterList([self.bias])
self.apply(_weights_init)
def forward(self, X):
size = X.size()
X = X.view(X.size(0), -1)
h = self.encode(X)
Xr, _ = self.decode(h)
Xr = Xr.view(size)
return Xr
def decode(self, h):
thetas, _ = torch.sort(h, dim=1, descending=True)
thetas = thetas[:, self.nb_active:self.nb_active+1]
h = h * (h > thetas).float()
Xr = torch.matmul(h, self.encode[0].weight) + self.bias
Xr = nn.Sigmoid()(Xr)
return Xr, h
class ZAE(nn.Module):
def __init__(self, w=32, h=32, c=1, nb_hidden=300, theta=1):
super().__init__()
self.nb_hidden = nb_hidden
self.theta = theta
self.encode = nn.Sequential(
nn.Linear(w*h*c, nb_hidden, bias=False)
)
self.bias = nn.Parameter(torch.zeros(w*h*c))
self.params = nn.ParameterList([self.bias])
self.apply(_weights_init)
def forward(self, X):
size = X.size()
X = X.view(X.size(0), -1)
h = self.encode(X)
Xr, _ = self.decode(h)
Xr = Xr.view(size)
return Xr
def decode(self, h):
h = h * (h > self.theta).float()
Xr = torch.matmul(h, self.encode[0].weight) + self.bias
Xr = nn.Sigmoid()(Xr)
return Xr, h
class DenseAE(nn.Module):
def __init__(self, w=32, h=32, c=1, encode_hidden=(300,), decode_hidden=(300,), ksparse=True, nb_active=10, denoise=None):
super().__init__()
self.encode_hidden = encode_hidden
self.decode_hidden = decode_hidden
self.ksparse = ksparse
self.nb_active = nb_active
self.denoise = denoise
# encode layers
layers = []
hid_prev = w * h * c
for hid in encode_hidden:
layers.extend([
nn.Linear(hid_prev, hid),
nn.ReLU(True)
])
hid_prev = hid
self.encode = nn.Sequential(*layers)
# decode layers
layers = []
for hid in decode_hidden:
layers.extend([
nn.Linear(hid_prev, hid),
nn.ReLU(True)
])
hid_prev = hid
layers.extend([
nn.Linear(hid_prev, w * h * c),
nn.Sigmoid()
])
self.decode = nn.Sequential(*layers)
self.apply(_weights_init)
def forward(self, X):
size = X.size()
if self.denoise is not None:
X = X * ((torch.rand(X.size()) <= self.denoise).float()).to(X.device)
X = X.view(X.size(0), -1)
h = self.encode(X)
if self.ksparse:
h = ksparse(h, nb_active=self.nb_active)
Xr = self.decode(h)
Xr = Xr.view(size)
return Xr
def ksparse(x, nb_active=10):
mask = torch.ones(x.size())
for i, xi in enumerate(x.data.tolist()):
inds = np.argsort(xi)
inds = inds[::-1]
inds = inds[nb_active:]
if len(inds):
inds = np.array(inds)
inds = torch.from_numpy(inds).long()
mask[i][inds] = 0
return x * (mask).float().to(x.device)
class ConvAE(nn.Module):
def __init__(self, w=32, h=32, c=1, nb_filters=64, spatial=True, channel=True, channel_stride=4):
super().__init__()
self.spatial = spatial
self.channel = channel
self.channel_stride = channel_stride
self.encode = nn.Sequential(
nn.Conv2d(c, nb_filters, 5, 1, 0),
nn.ReLU(True),
nn.Conv2d(nb_filters, nb_filters, 5, 1, 0),
nn.ReLU(True),
nn.Conv2d(nb_filters, nb_filters, 5, 1, 0),
)
self.decode = nn.Sequential(
nn.ConvTranspose2d(nb_filters, c, 13, 1, 0),
nn.Sigmoid()
)
self.apply(_weights_init)
def forward(self, X):
size = X.size()
h = self.encode(X)
h = self.sparsify(h)
Xr = self.decode(h)
return Xr
def sparsify(self, h):
if self.spatial:
h = spatial_sparsity(h)
if self.channel:
h = strided_channel_sparsity(h, stride=self.channel_stride)
return h
class SimpleConvAE(nn.Module):
def __init__(self, w=32, h=32, c=1, nb_filters=64, spatial=True, channel=True, channel_stride=4):
super().__init__()
self.spatial = spatial
self.channel = channel
self.channel_stride = channel_stride
self.encode = nn.Sequential(
nn.Conv2d(c, nb_filters, 13, 1, 0),
nn.ReLU(True),
)
self.decode = nn.Sequential(
nn.ConvTranspose2d(nb_filters, c, 13, 1, 0),
nn.Sigmoid()
)
self.apply(_weights_init)
def forward(self, X):
size = X.size()
h = self.encode(X)
h = self.sparsify(h)
Xr = self.decode(h)
return Xr
def sparsify(self, h):
if self.spatial:
h = spatial_sparsity(h)
if self.channel:
h = strided_channel_sparsity(h, stride=self.channel_stride)
return h
class DeepConvAE(nn.Module):
def __init__(self, w=32, h=32, c=1, nb_filters=64, nb_layers=3, spatial=True, channel=True, channel_stride=4):
super().__init__()
self.spatial = spatial
self.channel = channel
self.channel_stride = channel_stride
layers = [
nn.Conv2d(c, nb_filters, 5, 1, 0),
nn.ReLU(True),
]
for _ in range(nb_layers - 1):
layers.extend([
nn.Conv2d(nb_filters, nb_filters, 5, 1, 0),
nn.ReLU(True),
])
self.encode = nn.Sequential(*layers)
layers = []
for _ in range(nb_layers - 1):
layers.extend([
nn.ConvTranspose2d(nb_filters, nb_filters, 5, 1, 0),
nn.ReLU(True),
])
layers.extend([
nn.ConvTranspose2d(nb_filters, c, 5, 1, 0),
nn.Sigmoid()
])
self.decode = nn.Sequential(*layers)
self.apply(_weights_init)
def forward(self, X):
size = X.size()
h = self.encode(X)
h = self.sparsify(h)
Xr = self.decode(h)
return Xr
def sparsify(self, h):
if self.spatial:
h = spatial_sparsity(h)
if self.channel:
h = strided_channel_sparsity(h, stride=self.channel_stride)
return h
def spatial_sparsity(x):
maxes = x.amax(dim=(2,3), keepdims=True)
return x * equals(x, maxes)
def equals(x, y, eps=1e-8):
return torch.abs(x-y) <= eps
def strided_channel_sparsity(x, stride=1):
B, F = x.shape[0:2]
h, w = x.shape[2:]
x_ = x.view(B, F, h // stride, stride, w // stride, stride)
mask = equals(x_, x_.amax(axis=(1, 3, 5), keepdims=True))
mask = mask.view(x.shape).float()
return x * mask
def _weights_init(m):
if hasattr(m, 'weight'):
xavier_uniform(m.weight.data)
if m.bias is not None:
m.bias.data.fill_(0)
|