Spaces:
Sleeping
Sleeping
File size: 8,542 Bytes
c608946 c74a070 c608946 c74a070 c608946 c74a070 c608946 c74a070 c608946 c74a070 c608946 c74a070 c608946 c74a070 c608946 c74a070 c608946 c74a070 c608946 c74a070 c608946 c74a070 c608946 c74a070 c608946 c74a070 c608946 c74a070 c608946 c74a070 c608946 c74a070 c608946 c74a070 c608946 c74a070 c608946 c74a070 c608946 c74a070 c608946 c74a070 c608946 c74a070 c608946 c74a070 c608946 c74a070 c608946 c74a070 c608946 c74a070 c608946 |
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 |
from einops.einops import rearrange
import torch
import torch.nn as nn
import torch.nn.functional as F
from roma.utils.utils import get_gt_warp
import wandb
import roma
import math
class RobustLosses(nn.Module):
def __init__(
self,
robust=False,
center_coords=False,
scale_normalize=False,
ce_weight=0.01,
local_loss=True,
local_dist=4.0,
local_largest_scale=8,
smooth_mask=False,
depth_interpolation_mode="bilinear",
mask_depth_loss=False,
relative_depth_error_threshold=0.05,
alpha=1.0,
c=1e-3,
):
super().__init__()
self.robust = robust # measured in pixels
self.center_coords = center_coords
self.scale_normalize = scale_normalize
self.ce_weight = ce_weight
self.local_loss = local_loss
self.local_dist = local_dist
self.local_largest_scale = local_largest_scale
self.smooth_mask = smooth_mask
self.depth_interpolation_mode = depth_interpolation_mode
self.mask_depth_loss = mask_depth_loss
self.relative_depth_error_threshold = relative_depth_error_threshold
self.avg_overlap = dict()
self.alpha = alpha
self.c = c
def gm_cls_loss(self, x2, prob, scale_gm_cls, gm_certainty, scale):
with torch.no_grad():
B, C, H, W = scale_gm_cls.shape
device = x2.device
cls_res = round(math.sqrt(C))
G = torch.meshgrid(
*[
torch.linspace(
-1 + 1 / cls_res, 1 - 1 / cls_res, steps=cls_res, device=device
)
for _ in range(2)
]
)
G = torch.stack((G[1], G[0]), dim=-1).reshape(C, 2)
GT = (
(G[None, :, None, None, :] - x2[:, None])
.norm(dim=-1)
.min(dim=1)
.indices
)
cls_loss = F.cross_entropy(scale_gm_cls, GT, reduction="none")[prob > 0.99]
if not torch.any(cls_loss):
cls_loss = certainty_loss * 0.0 # Prevent issues where prob is 0 everywhere
certainty_loss = F.binary_cross_entropy_with_logits(gm_certainty[:, 0], prob)
losses = {
f"gm_certainty_loss_{scale}": certainty_loss.mean(),
f"gm_cls_loss_{scale}": cls_loss.mean(),
}
wandb.log(losses, step=roma.GLOBAL_STEP)
return losses
def delta_cls_loss(
self, x2, prob, flow_pre_delta, delta_cls, certainty, scale, offset_scale
):
with torch.no_grad():
B, C, H, W = delta_cls.shape
device = x2.device
cls_res = round(math.sqrt(C))
G = torch.meshgrid(
*[
torch.linspace(
-1 + 1 / cls_res, 1 - 1 / cls_res, steps=cls_res, device=device
)
for _ in range(2)
]
)
G = torch.stack((G[1], G[0]), dim=-1).reshape(C, 2) * offset_scale
GT = (
(G[None, :, None, None, :] + flow_pre_delta[:, None] - x2[:, None])
.norm(dim=-1)
.min(dim=1)
.indices
)
cls_loss = F.cross_entropy(delta_cls, GT, reduction="none")[prob > 0.99]
if not torch.any(cls_loss):
cls_loss = certainty_loss * 0.0 # Prevent issues where prob is 0 everywhere
certainty_loss = F.binary_cross_entropy_with_logits(certainty[:, 0], prob)
losses = {
f"delta_certainty_loss_{scale}": certainty_loss.mean(),
f"delta_cls_loss_{scale}": cls_loss.mean(),
}
wandb.log(losses, step=roma.GLOBAL_STEP)
return losses
def regression_loss(self, x2, prob, flow, certainty, scale, eps=1e-8, mode="delta"):
epe = (flow.permute(0, 2, 3, 1) - x2).norm(dim=-1)
if scale == 1:
pck_05 = (epe[prob > 0.99] < 0.5 * (2 / 512)).float().mean()
wandb.log({"train_pck_05": pck_05}, step=roma.GLOBAL_STEP)
ce_loss = F.binary_cross_entropy_with_logits(certainty[:, 0], prob)
a = self.alpha
cs = self.c * scale
x = epe[prob > 0.99]
reg_loss = cs**a * ((x / (cs)) ** 2 + 1**2) ** (a / 2)
if not torch.any(reg_loss):
reg_loss = ce_loss * 0.0 # Prevent issues where prob is 0 everywhere
losses = {
f"{mode}_certainty_loss_{scale}": ce_loss.mean(),
f"{mode}_regression_loss_{scale}": reg_loss.mean(),
}
wandb.log(losses, step=roma.GLOBAL_STEP)
return losses
def forward(self, corresps, batch):
scales = list(corresps.keys())
tot_loss = 0.0
# scale_weights due to differences in scale for regression gradients and classification gradients
scale_weights = {1: 1, 2: 1, 4: 1, 8: 1, 16: 1}
for scale in scales:
scale_corresps = corresps[scale]
(
scale_certainty,
flow_pre_delta,
delta_cls,
offset_scale,
scale_gm_cls,
scale_gm_certainty,
flow,
scale_gm_flow,
) = (
scale_corresps["certainty"],
scale_corresps["flow_pre_delta"],
scale_corresps.get("delta_cls"),
scale_corresps.get("offset_scale"),
scale_corresps.get("gm_cls"),
scale_corresps.get("gm_certainty"),
scale_corresps["flow"],
scale_corresps.get("gm_flow"),
)
flow_pre_delta = rearrange(flow_pre_delta, "b d h w -> b h w d")
b, h, w, d = flow_pre_delta.shape
gt_warp, gt_prob = get_gt_warp(
batch["im_A_depth"],
batch["im_B_depth"],
batch["T_1to2"],
batch["K1"],
batch["K2"],
H=h,
W=w,
)
x2 = gt_warp.float()
prob = gt_prob
if self.local_largest_scale >= scale:
prob = prob * (
F.interpolate(prev_epe[:, None], size=(h, w), mode="nearest-exact")[
:, 0
]
< (2 / 512) * (self.local_dist[scale] * scale)
)
if scale_gm_cls is not None:
gm_cls_losses = self.gm_cls_loss(
x2, prob, scale_gm_cls, scale_gm_certainty, scale
)
gm_loss = (
self.ce_weight * gm_cls_losses[f"gm_certainty_loss_{scale}"]
+ gm_cls_losses[f"gm_cls_loss_{scale}"]
)
tot_loss = tot_loss + scale_weights[scale] * gm_loss
elif scale_gm_flow is not None:
gm_flow_losses = self.regression_loss(
x2, prob, scale_gm_flow, scale_gm_certainty, scale, mode="gm"
)
gm_loss = (
self.ce_weight * gm_flow_losses[f"gm_certainty_loss_{scale}"]
+ gm_flow_losses[f"gm_regression_loss_{scale}"]
)
tot_loss = tot_loss + scale_weights[scale] * gm_loss
if delta_cls is not None:
delta_cls_losses = self.delta_cls_loss(
x2,
prob,
flow_pre_delta,
delta_cls,
scale_certainty,
scale,
offset_scale,
)
delta_cls_loss = (
self.ce_weight * delta_cls_losses[f"delta_certainty_loss_{scale}"]
+ delta_cls_losses[f"delta_cls_loss_{scale}"]
)
tot_loss = tot_loss + scale_weights[scale] * delta_cls_loss
else:
delta_regression_losses = self.regression_loss(
x2, prob, flow, scale_certainty, scale
)
reg_loss = (
self.ce_weight
* delta_regression_losses[f"delta_certainty_loss_{scale}"]
+ delta_regression_losses[f"delta_regression_loss_{scale}"]
)
tot_loss = tot_loss + scale_weights[scale] * reg_loss
prev_epe = (flow.permute(0, 2, 3, 1) - x2).norm(dim=-1).detach()
return tot_loss
|