Spaces:
Runtime error
Runtime error
File size: 9,681 Bytes
548d634 |
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 |
from pathlib import Path
from typing import Dict, List
import torchvision
import torch
import tops
import torchvision.transforms.functional as F
from .functional import hflip
class RandomHorizontalFlip(torch.nn.Module):
def __init__(self, p: float, flip_map=None,**kwargs):
super().__init__()
self.flip_ratio = p
self.flip_map = flip_map
if self.flip_ratio is None:
self.flip_ratio = 0.5
assert 0 <= self.flip_ratio <= 1
def forward(self, container: Dict[str, torch.Tensor]) -> Dict[str, torch.Tensor]:
if torch.rand(1) > self.flip_ratio:
return container
return hflip(container, self.flip_map)
class CenterCrop(torch.nn.Module):
"""
Performs the transform on the image.
NOTE: Does not transform the mask to improve runtime.
"""
def __init__(self, size: List[int]):
super().__init__()
self.size = tuple(size)
def forward(self, container: Dict[str, torch.Tensor]) -> Dict[str, torch.Tensor]:
min_size = min(container["img"].shape[1], container["img"].shape[2])
if min_size < self.size[0]:
container["img"] = F.center_crop(container["img"], min_size)
container["img"] = F.resize(container["img"], self.size)
return container
container["img"] = F.center_crop(container["img"], self.size)
return container
class Resize(torch.nn.Module):
"""
Performs the transform on the image.
NOTE: Does not transform the mask to improve runtime.
"""
def __init__(self, size, interpolation=F.InterpolationMode.BILINEAR):
super().__init__()
self.size = tuple(size)
self.interpolation = interpolation
def forward(self, container: Dict[str, torch.Tensor]) -> Dict[str, torch.Tensor]:
container["img"] = F.resize(container["img"], self.size, self.interpolation, antialias=True)
if "semantic_mask" in container:
container["semantic_mask"] = F.resize(
container["semantic_mask"], self.size, F.InterpolationMode.NEAREST)
if "embedding" in container:
container["embedding"] = F.resize(
container["embedding"], self.size, self.interpolation)
if "mask" in container:
container["mask"] = F.resize(
container["mask"], self.size, F.InterpolationMode.NEAREST)
if "E_mask" in container:
container["E_mask"] = F.resize(
container["E_mask"], self.size, F.InterpolationMode.NEAREST)
if "maskrcnn_mask" in container:
container["maskrcnn_mask"] = F.resize(
container["maskrcnn_mask"], self.size, F.InterpolationMode.NEAREST)
if "vertices" in container:
container["vertices"] = F.resize(
container["vertices"], self.size, F.InterpolationMode.NEAREST)
return container
def __repr__(self):
repr = super().__repr__()
vars_ = dict(size=self.size, interpolation=self.interpolation)
return repr + " " + " ".join([f"{k}: {v}" for k, v in vars_.items()])
class InsertHRImage(torch.nn.Module):
"""
Resizes mask by maxpool and assumes condition is already created
"""
def __init__(self, size, interpolation=F.InterpolationMode.BILINEAR):
super().__init__()
self.size = tuple(size)
self.interpolation = interpolation
def forward(self, container: Dict[str, torch.Tensor]) -> Dict[str, torch.Tensor]:
assert container["img"].dtype == torch.float32
container["img_hr"] = F.resize(container["img"], self.size, self.interpolation, antialias=True)
container["condition_hr"] = F.resize(container["condition"], self.size, self.interpolation, antialias=True)
mask = container["mask"] > 0
container["mask_hr"] = (torch.nn.functional.adaptive_max_pool2d(mask.logical_not().float(), output_size=self.size) > 0).logical_not().float()
container["condition_hr"] = container["condition_hr"] * (1 - container["mask_hr"]) + container["img_hr"] * container["mask_hr"]
return container
def __repr__(self):
repr = super().__repr__()
vars_ = dict(size=self.size, interpolation=self.interpolation)
return repr + " "
class CopyHRImage(torch.nn.Module):
def __init__(self) -> None:
super().__init__()
def forward(self, container: Dict[str, torch.Tensor]) -> Dict[str, torch.Tensor]:
container["img_hr"] = container["img"]
container["condition_hr"] = container["condition"]
container["mask_hr"] = container["mask"]
return container
class Resize2(torch.nn.Module):
"""
Resizes mask by maxpool and assumes condition is already created
"""
def __init__(self, size, interpolation=F.InterpolationMode.BILINEAR, downsample_condition: bool = True, mask_condition= True):
super().__init__()
self.size = tuple(size)
self.interpolation = interpolation
self.downsample_condition = downsample_condition
self.mask_condition = mask_condition
def forward(self, container: Dict[str, torch.Tensor]) -> Dict[str, torch.Tensor]:
# assert container["img"].dtype == torch.float32
container["img"] = F.resize(container["img"], self.size, self.interpolation, antialias=True)
mask = container["mask"] > 0
container["mask"] = (torch.nn.functional.adaptive_max_pool2d(mask.logical_not().float(), output_size=self.size) > 0).logical_not().float()
if self.downsample_condition:
container["condition"] = F.resize(container["condition"], self.size, self.interpolation, antialias=True)
if self.mask_condition:
container["condition"] = container["condition"] * (1 - container["mask"]) + container["img"] * container["mask"]
return container
def __repr__(self):
repr = super().__repr__()
vars_ = dict(size=self.size, interpolation=self.interpolation)
return repr + " " + " ".join([f"{k}: {v}" for k, v in vars_.items()])
class Normalize(torch.nn.Module):
"""
Performs the transform on the image.
NOTE: Does not transform the mask to improve runtime.
"""
def __init__(self, mean, std, inplace, keys=["img"]):
super().__init__()
self.mean = mean
self.std = std
self.inplace = inplace
self.keys = keys
def forward(self, container: Dict[str, torch.Tensor]) -> Dict[str, torch.Tensor]:
for key in self.keys:
container[key] = F.normalize(container[key], self.mean, self.std, self.inplace)
return container
def __repr__(self):
repr = super().__repr__()
vars_ = dict(mean=self.mean, std=self.std, inplace=self.inplace)
return repr + " " + " ".join([f"{k}: {v}" for k, v in vars_.items()])
class ToFloat(torch.nn.Module):
def __init__(self, keys=["img"], norm=True) -> None:
super().__init__()
self.keys = keys
self.gain = 255 if norm else 1
def forward(self, container: Dict[str, torch.Tensor]) -> Dict[str, torch.Tensor]:
for key in self.keys:
container[key] = container[key].float() / self.gain
return container
class RandomCrop(torchvision.transforms.RandomCrop):
"""
Performs the transform on the image.
NOTE: Does not transform the mask to improve runtime.
"""
def forward(self, container: Dict[str, torch.Tensor]) -> Dict[str, torch.Tensor]:
container["img"] = super().forward(container["img"])
return container
class CreateCondition(torch.nn.Module):
def forward(self, container: Dict[str, torch.Tensor]) -> Dict[str, torch.Tensor]:
if container["img"].dtype == torch.uint8:
container["condition"] = container["img"] * container["mask"].byte() + (1-container["mask"].byte()) * 127
return container
container["condition"] = container["img"] * container["mask"]
return container
class CreateEmbedding(torch.nn.Module):
def __init__(self, embed_path: Path, cuda=True) -> None:
super().__init__()
self.embed_map = torch.load(embed_path, map_location=torch.device("cpu"))
if cuda:
self.embed_map = tops.to_cuda(self.embed_map)
def forward(self, container: Dict[str, torch.Tensor]) -> Dict[str, torch.Tensor]:
vertices = container["vertices"]
if vertices.ndim == 3:
embedding = self.embed_map[vertices.long()].squeeze(dim=0)
embedding = embedding.permute(2, 0, 1) * container["E_mask"]
pass
else:
assert vertices.ndim == 4
embedding = self.embed_map[vertices.long()].squeeze(dim=1)
embedding = embedding.permute(0, 3, 1, 2) * container["E_mask"]
container["embedding"] = embedding
container["embed_map"] = self.embed_map.clone()
return container
class UpdateMask(torch.nn.Module):
def forward(self, container: Dict[str, torch.Tensor]) -> Dict[str, torch.Tensor]:
container["mask"] = (container["img"] == container["condition"]).any(dim=1, keepdims=True).float()
return container
class LoadClassEmbedding(torch.nn.Module):
def __init__(self, embedding_path: Path) -> None:
super().__init__()
self.embedding = torch.load(embedding_path, map_location="cpu")
def forward(self, container: Dict[str, torch.Tensor]) -> Dict[str, torch.Tensor]:
key = "_".join(container["__key__"].split("train/")[-1].split("/")[:-1])
container["class_embedding"] = self.embedding[key].view(-1)
return container
|