z2p / models.py
galmetzer's picture
app
cd438c2
from networks import *
class PosADANet(nn.Module):
def encode(self, shp):
device = self.omega.device
B, _, H, W = shp
row = torch.arange(H).to(device) / H
enc_row1 = torch.sin(self.omega[None, :] * row[:, None])
enc_row2 = torch.cos(self.omega[None, :] * row[:, None])
rows = torch.cat([enc_row1.unsqueeze(1).repeat((1, W, 1)), enc_row2.unsqueeze(1).repeat((1, W, 1))], dim=-1)
col = torch.arange(W).to(device) / W
enc_col1 = torch.sin(self.omega[None, :] * col[:, None])
enc_col2 = torch.cos(self.omega[None, :] * col[:, None])
cols = torch.cat([enc_col1.unsqueeze(0).repeat((H, 1, 1)), enc_col2.unsqueeze(0).repeat((H, 1, 1))], dim=-1)
encoding = torch.cat([rows, cols], dim=-1)
encoding = encoding.permute(2, 0, 1).unsqueeze(0).repeat((B, 1, 1, 1))
return encoding
def get_encoding(self, x):
shp1 = x.shape
singelton = self.positional_encoding is not None\
and self.positional_encoding.shape[0] == shp1[0] and self.positional_encoding.shape[2:] == shp1[2:]
if singelton:
return self.positional_encoding
self.positional_encoding = self.encode(x.shape)
return self.positional_encoding
def __init__(self, input_channels, output_channels, n_style, bilinear=True, padding='zero', full_ada=True,
nfreq=20, magnitude=10):
super(PosADANet, self).__init__()
factor = 2 if bilinear else 1
self.omega = nn.Parameter(torch.rand(nfreq) * magnitude)
self.omega.requires_grad = False
self.positional_encoding = None
self.full_ada = full_ada
self.style_encoder = FullyConnected(n_style, W_SIZE, layers=6)
self.padding = padding
self.input_channels = input_channels + nfreq * 4
self.n_classes = output_channels
self.bilinear = bilinear
self.channels = [512 // factor, 256 // factor, 128 // factor]
self.inc = DoubleConv(self.input_channels, 64)
self.down1 = Down(64, 128, padding=padding, ada=self.full_ada)
self.down2 = Down(128, 256, padding=padding, ada=self.full_ada)
self.down3 = Down(256, 512, padding=padding, ada=self.full_ada)
self.down4 = Down(512, 1024 // factor, padding=padding, ada=self.full_ada)
self.up1 = Up(1024, 512 // factor, bilinear, ada=True, padding=padding)
self.up2 = Up(512, 256 // factor, bilinear, ada=True, padding=padding)
self.up3 = Up(256, 128 // factor, bilinear, ada=True, padding=padding)
self.up4 = Up(128, 64, bilinear, padding=padding, ada=True)
self.outc = OutConv(64, output_channels, padding=padding)
def forward(self, x, style):
w = self.style_encoder(style)
encoding = self.get_encoding(x)
x = torch.cat([x, encoding], dim=1)
x1 = self.inc(x)
if self.full_ada:
x2 = self.down1(x1, w=w)
x3 = self.down2(x2, w=w)
x4 = self.down3(x3, w=w)
x5 = self.down4(x4, w=w)
else:
x2 = self.down1(x1)
x3 = self.down2(x2)
x4 = self.down3(x3)
x5 = self.down4(x4)
x = self.up1(x5, x4, w=w)
x = self.up2(x, x3, w=w)
x = self.up3(x, x2, w=w)
x = self.up4(x, x1, w=w)
logits = self.outc(x)
return logits