File size: 6,729 Bytes
d4c980e |
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 |
# coding=utf-8
# Copyright 2020 The Google Research Authors.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# pylint: skip-file
"""Layers for defining NCSN++.
"""
from . import layers
import score_models.layers.up_or_downsampling2d as up_or_down_sampling
import torch.nn as nn
import torch
import torch.nn.functional as F
import numpy as np
conv1x1 = layers.ddpm_conv1x1
conv3x3 = layers.ddpm_conv3x3
NIN = layers.NIN
default_init = layers.default_init
class GaussianFourierProjection(nn.Module):
"""Gaussian Fourier embeddings for noise levels."""
def __init__(self, embedding_size=256, scale=1.0):
super().__init__()
self.W = nn.Parameter(torch.randn(embedding_size) * scale, requires_grad=False)
def forward(self, x):
x_proj = x[:, None] * self.W[None, :] * 2 * np.pi
return torch.cat([torch.sin(x_proj), torch.cos(x_proj)], dim=-1)
class Combine(nn.Module):
"""Combine information from skip connections."""
def __init__(self, dim1, dim2, method='cat'):
super().__init__()
self.Conv_0 = conv1x1(dim1, dim2)
self.method = method
def forward(self, x, y):
h = self.Conv_0(x)
if self.method == 'cat':
return torch.cat([h, y], dim=1)
elif self.method == 'sum':
return h + y
else:
raise ValueError(f'Method {self.method} not recognized.')
class AttnBlockpp(nn.Module):
"""Channel-wise self-attention block. Modified from DDPM."""
def __init__(self, channels, skip_rescale=False, init_scale=0.):
super().__init__()
self.GroupNorm_0 = nn.GroupNorm(num_groups=min(channels // 4, 32), num_channels=channels,
eps=1e-6)
self.NIN_0 = NIN(channels, channels)
self.NIN_1 = NIN(channels, channels)
self.NIN_2 = NIN(channels, channels)
self.NIN_3 = NIN(channels, channels, init_scale=init_scale)
self.skip_rescale = skip_rescale
def forward(self, x):
B, C, H, W = x.shape
h = self.GroupNorm_0(x)
q = self.NIN_0(h)
k = self.NIN_1(h)
v = self.NIN_2(h)
w = torch.einsum('bchw,bcij->bhwij', q, k) * (int(C) ** (-0.5))
w = torch.reshape(w, (B, H, W, H * W))
w = F.softmax(w, dim=-1)
w = torch.reshape(w, (B, H, W, H, W))
h = torch.einsum('bhwij,bcij->bchw', w, v)
h = self.NIN_3(h)
if not self.skip_rescale:
return x + h
else:
return (x + h) / np.sqrt(2.)
class ResnetBlockDDPMpp(nn.Module):
"""ResBlock adapted from DDPM."""
def __init__(self, act, in_ch, out_ch=None, temb_dim=None, conv_shortcut=False,
dropout=0.1, skip_rescale=False, init_scale=0.):
super().__init__()
out_ch = out_ch if out_ch else in_ch
self.GroupNorm_0 = nn.GroupNorm(num_groups=min(in_ch // 4, 32), num_channels=in_ch, eps=1e-6)
self.Conv_0 = conv3x3(in_ch, out_ch)
if temb_dim is not None:
self.Dense_0 = nn.Linear(temb_dim, out_ch)
self.Dense_0.weight.data = default_init()(self.Dense_0.weight.data.shape)
nn.init.zeros_(self.Dense_0.bias)
self.GroupNorm_1 = nn.GroupNorm(num_groups=min(out_ch // 4, 32), num_channels=out_ch, eps=1e-6)
self.Dropout_0 = nn.Dropout(dropout)
self.Conv_1 = conv3x3(out_ch, out_ch, init_scale=init_scale)
if in_ch != out_ch:
if conv_shortcut:
self.Conv_2 = conv3x3(in_ch, out_ch)
else:
self.NIN_0 = NIN(in_ch, out_ch)
self.skip_rescale = skip_rescale
self.act = act
self.out_ch = out_ch
self.conv_shortcut = conv_shortcut
def forward(self, x, temb=None):
h = self.act(self.GroupNorm_0(x))
h = self.Conv_0(h)
if temb is not None:
h += self.Dense_0(self.act(temb))[:, :, None, None]
h = self.act(self.GroupNorm_1(h))
h = self.Dropout_0(h)
h = self.Conv_1(h)
if x.shape[1] != self.out_ch:
if self.conv_shortcut:
x = self.Conv_2(x)
else:
x = self.NIN_0(x)
if not self.skip_rescale:
return x + h
else:
return (x + h) / np.sqrt(2.)
class ResnetBlockBigGANpp(nn.Module):
def __init__(self, act, in_ch, out_ch=None, temb_dim=None, up=False, down=False,
dropout=0.1, fir=False, fir_kernel=(1, 3, 3, 1),
skip_rescale=True, init_scale=0.):
super().__init__()
out_ch = out_ch if out_ch else in_ch
self.GroupNorm_0 = nn.GroupNorm(num_groups=min(in_ch // 4, 32), num_channels=in_ch, eps=1e-6)
self.up = up
self.down = down
self.fir = fir
self.fir_kernel = fir_kernel
self.Conv_0 = conv3x3(in_ch, out_ch)
if temb_dim is not None:
self.Dense_0 = nn.Linear(temb_dim, out_ch)
self.Dense_0.weight.data = default_init()(self.Dense_0.weight.shape)
nn.init.zeros_(self.Dense_0.bias)
self.GroupNorm_1 = nn.GroupNorm(num_groups=min(out_ch // 4, 32), num_channels=out_ch, eps=1e-6)
self.Dropout_0 = nn.Dropout(dropout)
self.Conv_1 = conv3x3(out_ch, out_ch, init_scale=init_scale)
if in_ch != out_ch or up or down:
self.Conv_2 = conv1x1(in_ch, out_ch)
self.skip_rescale = skip_rescale
self.act = act
self.in_ch = in_ch
self.out_ch = out_ch
def forward(self, x, temb=None):
h = self.act(self.GroupNorm_0(x))
if self.up:
if self.fir:
h = up_or_down_sampling.upsample_2d(h, self.fir_kernel, factor=2)
x = up_or_down_sampling.upsample_2d(x, self.fir_kernel, factor=2)
else:
h = up_or_down_sampling.naive_upsample_2d(h, factor=2)
x = up_or_down_sampling.naive_upsample_2d(x, factor=2)
elif self.down:
if self.fir:
h = up_or_down_sampling.downsample_2d(h, self.fir_kernel, factor=2)
x = up_or_down_sampling.downsample_2d(x, self.fir_kernel, factor=2)
else:
h = up_or_down_sampling.naive_downsample_2d(h, factor=2)
x = up_or_down_sampling.naive_downsample_2d(x, factor=2)
h = self.Conv_0(h)
# Add bias to each feature map conditioned on the time embedding
if temb is not None:
h += self.Dense_0(self.act(temb))[:, :, None, None]
h = self.act(self.GroupNorm_1(h))
h = self.Dropout_0(h)
h = self.Conv_1(h)
if self.in_ch != self.out_ch or self.up or self.down:
x = self.Conv_2(x)
if not self.skip_rescale:
return x + h
else:
return (x + h) / np.sqrt(2.)
|