File size: 9,558 Bytes
9e7a39a |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 |
from typing import List, Optional, Union
import torch
import torch.nn as nn
import torch.nn.functional as F
from omegaconf import ListConfig
from torchvision.utils import save_image
from ...util import append_dims, instantiate_from_config
class StandardDiffusionLoss(nn.Module):
def __init__(
self,
sigma_sampler_config,
type="l2",
offset_noise_level=0.0,
batch2model_keys: Optional[Union[str, List[str], ListConfig]] = None,
):
super().__init__()
assert type in ["l2", "l1"]
self.sigma_sampler = instantiate_from_config(sigma_sampler_config)
self.type = type
self.offset_noise_level = offset_noise_level
if not batch2model_keys:
batch2model_keys = []
if isinstance(batch2model_keys, str):
batch2model_keys = [batch2model_keys]
self.batch2model_keys = set(batch2model_keys)
def __call__(self, network, denoiser, conditioner, input, batch, *args, **kwarg):
cond = conditioner(batch)
additional_model_inputs = {
key: batch[key] for key in self.batch2model_keys.intersection(batch)
}
sigmas = self.sigma_sampler(input.shape[0]).to(input.device)
noise = torch.randn_like(input)
if self.offset_noise_level > 0.0:
noise = noise + self.offset_noise_level * append_dims(
torch.randn(input.shape[0], device=input.device), input.ndim
)
noised_input = input + noise * append_dims(sigmas, input.ndim)
model_output = denoiser(
network, noised_input, sigmas, cond, **additional_model_inputs
)
w = append_dims(denoiser.w(sigmas), input.ndim)
loss = self.get_diff_loss(model_output, input, w)
loss = loss.mean()
loss_dict = {"loss": loss}
return loss, loss_dict
def get_diff_loss(self, model_output, target, w):
if self.type == "l2":
return torch.mean(
(w * (model_output - target) ** 2).reshape(target.shape[0], -1), 1
)
elif self.type == "l1":
return torch.mean(
(w * (model_output - target).abs()).reshape(target.shape[0], -1), 1
)
class FullLoss(StandardDiffusionLoss):
def __init__(
self,
seq_len=12,
kernel_size=3,
gaussian_sigma=0.5,
min_attn_size=16,
lambda_local_loss=0.0,
lambda_ocr_loss=0.0,
ocr_enabled = False,
predictor_config = None,
*args, **kwarg
):
super().__init__(*args, **kwarg)
self.gaussian_kernel_size = kernel_size
gaussian_kernel = self.get_gaussian_kernel(kernel_size=self.gaussian_kernel_size, sigma=gaussian_sigma, out_channels=seq_len)
self.register_buffer("g_kernel", gaussian_kernel.requires_grad_(False))
self.min_attn_size = min_attn_size
self.lambda_local_loss = lambda_local_loss
self.lambda_ocr_loss = lambda_ocr_loss
self.ocr_enabled = ocr_enabled
if ocr_enabled:
self.predictor = instantiate_from_config(predictor_config)
def get_gaussian_kernel(self, kernel_size=3, sigma=1, out_channels=3):
# Create a x, y coordinate grid of shape (kernel_size, kernel_size, 2)
x_coord = torch.arange(kernel_size)
x_grid = x_coord.repeat(kernel_size).view(kernel_size, kernel_size)
y_grid = x_grid.t()
xy_grid = torch.stack([x_grid, y_grid], dim=-1).float()
mean = (kernel_size - 1)/2.
variance = sigma**2.
# Calculate the 2-dimensional gaussian kernel which is
# the product of two gaussian distributions for two different
# variables (in this case called x and y)
gaussian_kernel = (1./(2.*torch.pi*variance)) *\
torch.exp(
-torch.sum((xy_grid - mean)**2., dim=-1) /\
(2*variance)
)
# Make sure sum of values in gaussian kernel equals 1.
gaussian_kernel = gaussian_kernel / torch.sum(gaussian_kernel)
# Reshape to 2d depthwise convolutional weight
gaussian_kernel = gaussian_kernel.view(1, 1, kernel_size, kernel_size)
gaussian_kernel = gaussian_kernel.tile(out_channels, 1, 1, 1)
return gaussian_kernel
def __call__(self, network, denoiser, conditioner, input, batch, first_stage_model, scaler):
cond = conditioner(batch)
sigmas = self.sigma_sampler(input.shape[0]).to(input.device)
noise = torch.randn_like(input)
if self.offset_noise_level > 0.0:
noise = noise + self.offset_noise_level * append_dims(
torch.randn(input.shape[0], device=input.device), input.ndim
)
noised_input = input + noise * append_dims(sigmas, input.ndim)
model_output = denoiser(network, noised_input, sigmas, cond)
w = append_dims(denoiser.w(sigmas), input.ndim)
diff_loss = self.get_diff_loss(model_output, input, w)
local_loss = self.get_local_loss(network.diffusion_model.attn_map_cache, batch["seg"], batch["seg_mask"])
diff_loss = diff_loss.mean()
local_loss = local_loss.mean()
if self.ocr_enabled:
ocr_loss = self.get_ocr_loss(model_output, batch["r_bbox"], batch["label"], first_stage_model, scaler)
ocr_loss = ocr_loss.mean()
loss = diff_loss + self.lambda_local_loss * local_loss
if self.ocr_enabled:
loss += self.lambda_ocr_loss * ocr_loss
loss_dict = {
"loss/diff_loss": diff_loss,
"loss/local_loss": local_loss,
"loss/full_loss": loss
}
if self.ocr_enabled:
loss_dict["loss/ocr_loss"] = ocr_loss
return loss, loss_dict
def get_ocr_loss(self, model_output, r_bbox, label, first_stage_model, scaler):
model_output = 1 / scaler * model_output
model_output_decoded = first_stage_model.decode(model_output)
model_output_crops = []
for i, bbox in enumerate(r_bbox):
m_top, m_bottom, m_left, m_right = bbox
model_output_crops.append(model_output_decoded[i, :, m_top:m_bottom, m_left:m_right])
loss = self.predictor.calc_loss(model_output_crops, label)
return loss
def get_min_local_loss(self, attn_map_cache, mask, seg_mask):
loss = 0
count = 0
for item in attn_map_cache:
heads = item["heads"]
size = item["size"]
attn_map = item["attn_map"]
if size < self.min_attn_size: continue
seg_l = seg_mask.shape[1]
bh, n, l = attn_map.shape # bh: batch size * heads / n : pixel length(h*w) / l: token length
attn_map = attn_map.reshape((-1, heads, n, l)) # b, h, n, l
assert seg_l <= l
attn_map = attn_map[..., :seg_l]
attn_map = attn_map.permute(0, 1, 3, 2) # b, h, l, n
attn_map = attn_map.mean(dim = 1) # b, l, n
attn_map = attn_map.reshape((-1, seg_l, size, size)) # b, l, s, s
attn_map = F.conv2d(attn_map, self.g_kernel, padding = self.gaussian_kernel_size//2, groups=seg_l) # gaussian blur on each channel
attn_map = attn_map.reshape((-1, seg_l, n)) # b, l, n
mask_map = F.interpolate(mask, (size, size))
mask_map = mask_map.tile((1, seg_l, 1, 1))
mask_map = mask_map.reshape((-1, seg_l, n)) # b, l, n
p_loss = (mask_map * attn_map).max(dim = -1)[0] # b, l
p_loss = p_loss + (1 - seg_mask) # b, l
p_loss = p_loss.min(dim = -1)[0] # b,
loss += -p_loss
count += 1
loss = loss / count
return loss
def get_local_loss(self, attn_map_cache, seg, seg_mask):
loss = 0
count = 0
for item in attn_map_cache:
heads = item["heads"]
size = item["size"]
attn_map = item["attn_map"]
if size < self.min_attn_size: continue
seg_l = seg_mask.shape[1]
bh, n, l = attn_map.shape # bh: batch size * heads / n : pixel length(h*w) / l: token length
attn_map = attn_map.reshape((-1, heads, n, l)) # b, h, n, l
assert seg_l <= l
attn_map = attn_map[..., :seg_l]
attn_map = attn_map.permute(0, 1, 3, 2) # b, h, l, n
attn_map = attn_map.mean(dim = 1) # b, l, n
attn_map = attn_map.reshape((-1, seg_l, size, size)) # b, l, s, s
attn_map = F.conv2d(attn_map, self.g_kernel, padding = self.gaussian_kernel_size//2, groups=seg_l) # gaussian blur on each channel
attn_map = attn_map.reshape((-1, seg_l, n)) # b, l, n
seg_map = F.interpolate(seg, (size, size))
seg_map = seg_map.reshape((-1, seg_l, n)) # b, l, n
n_seg_map = 1 - seg_map
p_loss = (seg_map * attn_map).max(dim = -1)[0] # b, l
n_loss = (n_seg_map * attn_map).max(dim = -1)[0] # b, l
p_loss = p_loss * seg_mask # b, l
n_loss = n_loss * seg_mask # b, l
p_loss = p_loss.sum(dim = -1) / seg_mask.sum(dim = -1) # b,
n_loss = n_loss.sum(dim = -1) / seg_mask.sum(dim = -1) # b,
f_loss = n_loss - p_loss # b,
loss += f_loss
count += 1
loss = loss / count
return loss |