File size: 6,495 Bytes
338c3e4 | 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 | from typing import Optional
import torch
from torch import nn, Tensor
from .base_model import AutoCfdModel
class ResidualBlock(nn.Module):
def __init__(
self,
in_chan: int,
out_chan: int,
hidden_chan: int,
kernel_size: int,
stride: int = 1,
padding: int = 0,
dropout_rate: float = 0.2,
bias: bool = True,
use_1x1conv: bool = False,
):
super().__init__()
if in_chan != out_chan:
assert use_1x1conv
self.in_chan = in_chan
self.out_chan = out_chan
self.hidden_chan = hidden_chan
self.kernel_size = kernel_size
self.stride = stride
self.padding = padding
self.bias = bias
# Sub-modules
self.conv1 = nn.Conv2d(
in_chan,
hidden_chan,
kernel_size,
stride,
padding,
bias=bias,
padding_mode="replicate",
)
self.bn1 = nn.BatchNorm2d(hidden_chan)
self.dropout = nn.Dropout(p=dropout_rate)
self.act = nn.GELU()
self.conv2 = nn.Conv2d(
hidden_chan,
out_chan,
kernel_size,
stride,
padding,
bias=bias,
padding_mode="replicate",
)
self.bn2 = nn.BatchNorm2d(out_chan)
if use_1x1conv:
self.res_conv = nn.Conv2d(
in_chan,
out_chan,
kernel_size=1,
stride=stride,
padding=0,
bias=bias,
)
else:
self.res_conv = None
def forward(self, x: Tensor) -> Tensor:
if self.res_conv is not None:
residual = self.res_conv(x)
else:
residual = x
x = self.conv1(x)
x = self.bn1(x)
x = self.dropout(x)
x = self.act(x)
x = self.conv2(x)
x = self.bn2(x)
x += residual
return x
class ResNet(AutoCfdModel):
def __init__(
self,
in_chan: int,
out_chan: int,
n_case_params: int,
loss_fn: nn.Module,
hidden_chan: int = 32,
num_blocks: int = 4,
kernel_size: int = 7,
padding: int = 3,
stride: int = 1,
):
super().__init__(loss_fn)
assert in_chan == out_chan
self.in_chan = in_chan
self.out_chan = out_chan
self.n_case_params = n_case_params
self.hidden_chan = hidden_chan
self.num_blocks = num_blocks
self.kernel_size = kernel_size
self.padding = padding
self.stride = stride
# Create sequence of residual blocks
blocks = [
ResidualBlock(
in_chan + 1 + n_case_params, # + 1 for mask
hidden_chan,
64,
kernel_size,
stride,
padding,
use_1x1conv=True,
),
]
for _ in range(num_blocks):
blocks.append(
ResidualBlock(
hidden_chan,
hidden_chan,
64,
kernel_size,
stride,
padding,
use_1x1conv=False,
)
)
blocks.append(
ResidualBlock(
hidden_chan,
out_chan,
64,
kernel_size,
stride,
padding,
use_1x1conv=True,
)
)
self.blocks = nn.Sequential(*blocks)
def forward(
self,
inputs: Tensor,
case_params: Tensor,
mask: Optional[Tensor] = None,
label: Optional[Tensor] = None,
) -> dict:
"""
Args:
- x: (B, in_chan, h, w)
- case_params: (B, 3)
- mask: (h, w) or (B, h, w). 1 for interior, 0 for boundaries.
- label: (B, out_chan, h, w)
Returns:
(B, out_chan, h, w) or (B, out_chan, h, w), loss
"""
residual = inputs[:, : self.out_chan]
batch_size, n_chan, height, width = inputs.shape
if mask is None:
mask = torch.ones((batch_size, height, width)).to(inputs.device)
else:
if mask.dim() == 3:
mask = mask.unsqueeze(1) # (B, 1, h, w)
inputs = torch.cat([inputs, mask], dim=1) # (B, c + 1, h, w)
# Add case params as additional channels
case_params = case_params.unsqueeze(-1).unsqueeze(-1) # (B, c, 1, 1)
# (B, n_params, h, w)
case_params = case_params.expand(
-1, -1, inputs.shape[-2], inputs.shape[-1]
)
inputs = torch.cat(
[inputs, case_params], dim=1
) # (B, c + n_params, h, w)
inputs = self.blocks(inputs) # (B, c, h, w)
preds = inputs + residual
if mask is not None:
# Mask out predictions.
preds = preds * mask
if label is not None:
if mask is not None:
# Mask out labels
label = label * mask
loss = self.loss_fn(preds=preds, labels=label)
return dict(
preds=preds,
loss=loss,
)
else:
return dict(preds=preds)
def generate(
self,
inputs: Tensor,
case_params: Tensor,
mask: Optional[Tensor] = None,
):
outputs = self.forward(inputs, case_params=case_params, mask=mask)
preds = outputs["preds"]
return preds
def generate_many(
self,
inputs: Tensor,
case_params: Tensor,
steps: int,
mask: Tensor,
):
"""
x: (c, h, w)
mask: (h, w). 1 for interior, 0 for boundaries.
Returns:
(steps, c, h, w)
"""
if inputs.dim() == 3:
inputs = inputs.unsqueeze(0) # (1, c, h, w)
case_params = case_params.unsqueeze(0) # (1, p)
mask = mask.unsqueeze(0) # (1, h, w)
cur_frame = inputs # (1, c, h, w)
frames = [cur_frame]
# boundaries = (1 - mask) * cur_frame # (1, c, h, w)
for _ in range(steps):
cur_frame = self.generate(
cur_frame, case_params=case_params, mask=mask
)
frames.append(cur_frame)
return frames
|