Spaces:
Runtime error
Runtime error
File size: 20,320 Bytes
583456e |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 |
# Copyright (c) Facebook, Inc. and its affiliates.
# Copyright (c) Meta Platforms, Inc. All Rights Reserved
# Modified by Feng Liang from
# https://github.com/MendelXu/zsseg.baseline/blob/master/mask_former/zero_shot_mask_former_model.py
import logging
from typing import Tuple
import numpy as np
import torch
from torch import nn
from torch.nn import functional as F
from detectron2.config import configurable
from detectron2.data import MetadataCatalog
from detectron2.modeling import META_ARCH_REGISTRY
from detectron2.modeling.backbone import Backbone
from detectron2.modeling.postprocessing import sem_seg_postprocess
from detectron2.structures import ImageList
from detectron2.utils.logger import log_first_n
from .modeling.clip_adapter import (
ClipAdapter,
MaskFormerClipAdapter,
build_text_prompt,
)
from .mask_former_model import MaskFormer
from .utils.misc import get_gt_binary_masks
@META_ARCH_REGISTRY.register()
class OVSeg(MaskFormer):
"""
Main class for zero shot mask classification semantic segmentation architectures.
"""
@configurable
def __init__(
self,
*,
backbone: Backbone,
sem_seg_head: nn.Module,
clip_adapter: nn.Module,
criterion: nn.Module,
num_queries: int,
panoptic_on: bool,
object_mask_threshold: float,
overlap_threshold: float,
metadata,
size_divisibility: int,
sem_seg_postprocess_before_inference: bool,
clip_ensemble: bool,
clip_ensemble_weight: float,
pixel_mean: Tuple[float],
pixel_std: Tuple[float],
):
"""
Args:
backbone: a backbone module, must follow detectron2's backbone interface
sem_seg_head: a module that predicts semantic segmentation from backbone features
criterion: a module that defines the loss
clip_adapter: adapter for clip-based mask classification
num_queries: int, number of queries
panoptic_on: bool, whether to output panoptic segmentation prediction
object_mask_threshold: float, threshold to filter query based on classification score
for panoptic segmentation inference
overlap_threshold: overlap threshold used in general inference for panoptic segmentation
metadata: dataset meta, get `thing` and `stuff` category names for panoptic
segmentation inference
size_divisibility: Some backbones require the input height and width to be divisible by a
specific integer. We can use this to override such requirement.
sem_seg_postprocess_before_inference: whether to resize the prediction back
to original input size before semantic segmentation inference or after.
For high-resolution dataset like Mapillary, resizing predictions before
inference will cause OOM error.
pixel_mean, pixel_std: list or tuple with #channels element, representing
the per-channel mean and std to be used to normalize the input image
"""
super().__init__(
backbone=backbone,
sem_seg_head=sem_seg_head,
criterion=criterion,
num_queries=num_queries,
panoptic_on=panoptic_on,
object_mask_threshold=object_mask_threshold,
overlap_threshold=overlap_threshold,
metadata=metadata,
size_divisibility=size_divisibility,
sem_seg_postprocess_before_inference=sem_seg_postprocess_before_inference,
pixel_mean=pixel_mean,
pixel_std=pixel_std,
)
self.clip_adapter: ClipAdapter = clip_adapter
self.clip_ensemble: bool = clip_ensemble
self.clip_ensemble_weight: float = clip_ensemble_weight
@classmethod
def from_config(cls, cfg):
init_kwargs = MaskFormer.from_config(cfg)
text_templates = build_text_prompt(cfg.MODEL.CLIP_ADAPTER)
clip_adapter = MaskFormerClipAdapter(
cfg.MODEL.CLIP_ADAPTER.CLIP_MODEL_NAME,
text_templates,
mask_fill=cfg.MODEL.CLIP_ADAPTER.MASK_FILL,
mask_expand_ratio=cfg.MODEL.CLIP_ADAPTER.MASK_EXPAND_RATIO,
mask_thr=cfg.MODEL.CLIP_ADAPTER.MASK_THR,
mask_matting=cfg.MODEL.CLIP_ADAPTER.MASK_MATTING,
region_resized=cfg.MODEL.CLIP_ADAPTER.REGION_RESIZED,
mask_prompt_depth=cfg.MODEL.CLIP_ADAPTER.MASK_PROMPT_DEPTH,
mask_prompt_fwd=cfg.MODEL.CLIP_ADAPTER.MASK_PROMPT_FWD,
)
init_kwargs["clip_adapter"] = clip_adapter
init_kwargs["clip_ensemble"] = cfg.MODEL.CLIP_ADAPTER.CLIP_ENSEMBLE
init_kwargs[
"clip_ensemble_weight"
] = cfg.MODEL.CLIP_ADAPTER.CLIP_ENSEMBLE_WEIGHT
return init_kwargs
def forward(self, batched_inputs):
"""
Args:
batched_inputs: a list, batched outputs of :class:`DatasetMapper`.
Each item in the list contains the inputs for one image.
For now, each item in the list is a dict that contains:
* "image": Tensor, image in (C, H, W) format.
* "instances": per-region ground truth
* Other information that's included in the original dicts, such as:
"height", "width" (int): the output resolution of the model (may be different
from input resolution), used in inference.
Returns:
list[dict]:
each dict has the results for one image. The dict contains the following keys:
* "sem_seg":
A Tensor that represents the
per-pixel segmentation prediced by the head.
The prediction has shape KxHxW that represents the logits of
each class for each pixel.
* "panoptic_seg":
A tuple that represent panoptic output
panoptic_seg (Tensor): of shape (height, width) where the values are ids for each segment.
segments_info (list[dict]): Describe each segment in `panoptic_seg`.
Each dict contains keys "id", "category_id", "isthing".
"""
dataset_name = [x["meta"]["dataset_name"] for x in batched_inputs]
assert len(set(dataset_name)) == 1
dataset_name = dataset_name[0]
images = [x["image"].to(self.device) for x in batched_inputs]
images = [(x - self.pixel_mean) / self.pixel_std for x in images]
images = ImageList.from_tensors(images, self.size_divisibility)
features = self.backbone(images.tensor)
outputs = self.sem_seg_head(features)
class_names = self.get_class_name_list(dataset_name)
text_features = self.clip_adapter.get_text_features(class_names)
outputs["pred_logits"] = self.clip_adapter.get_sim_logits(
text_features, self.clip_adapter.normalize_feature(outputs["pred_logits"])
)
if self.training:
if "aux_outputs" in outputs.keys():
for i in range(len(outputs["aux_outputs"])):
outputs["aux_outputs"][i][
"pred_logits"
] = self.clip_adapter.get_sim_logits(
text_features,
self.clip_adapter.normalize_feature(
outputs["aux_outputs"][i]["pred_logits"]
),
)
# mask classification target
if "instances" in batched_inputs[0]:
gt_instances = [x["instances"].to(self.device) for x in batched_inputs]
targets = self.prepare_targets(gt_instances, images)
else:
targets = None
# bipartite matching-based loss
losses = self.criterion(outputs, targets)
for k in list(losses.keys()):
if k in self.criterion.weight_dict:
losses[k] *= self.criterion.weight_dict[k]
else:
# remove this loss if not specified in `weight_dict`
losses.pop(k)
return losses
else:
mask_cls_results = outputs["pred_logits"]
mask_pred_results = outputs["pred_masks"]
# upsample masks
mask_pred_results = F.interpolate(
mask_pred_results,
size=(images.tensor.shape[-2], images.tensor.shape[-1]),
mode="bilinear",
align_corners=False,
)
processed_results = []
for mask_cls_result, mask_pred_result, input_per_image, image_size in zip(
mask_cls_results, mask_pred_results, batched_inputs, images.image_sizes
):
height = image_size[0]
width = image_size[1]
mask_pred_result = sem_seg_postprocess(
mask_pred_result, image_size, height, width
)
image = input_per_image["image"].to(self.device)
r, regions = self.semantic_inference(
mask_cls_result, mask_pred_result, image, class_names
)
height = input_per_image.get("height", image_size[0])
width = input_per_image.get("width", image_size[1])
r = sem_seg_postprocess(r, image_size, height, width)
processed_results.append({"sem_seg": r})
# panoptic segmentation inference
if self.panoptic_on:
panoptic_r = self.panoptic_inference(
mask_cls_result, mask_pred_result
)
processed_results[-1]["panoptic_seg"] = panoptic_r
return processed_results
def semantic_inference(self, mask_cls, mask_pred, image, class_names):
mask_cls = F.softmax(mask_cls, dim=-1)[..., :-1]
mask_pred = mask_pred.sigmoid()
regions = None
if self.clip_ensemble:
clip_cls, regions, valid_flag = self.clip_adapter(
image, class_names, mask_pred, normalize=True
)
if clip_cls is None:
clip_cls = torch.empty(0, mask_cls.shape[-1] + 1, device=self.device)
# softmax before index or after?
clip_cls = F.softmax(clip_cls[:, :-1], dim=-1)
if self.clip_ensemble_weight > 0:
map_back_clip_cls = mask_cls.new_ones(mask_cls.shape)
map_back_clip_cls[valid_flag] = clip_cls
mask_cls = torch.pow(mask_cls, 1 - self.clip_ensemble_weight) * \
torch.pow(map_back_clip_cls, self.clip_ensemble_weight)
else:
# only clip model predictions are used
mask_cls = clip_cls
mask_pred = mask_pred[valid_flag]
semseg = torch.einsum("qc,qhw->chw", mask_cls, mask_pred)
return semseg, regions
def get_class_name_list(self, dataset_name):
class_names = [
c.strip() for c in MetadataCatalog.get(dataset_name).stuff_classes
]
return class_names
@META_ARCH_REGISTRY.register()
class OVSegDEMO(MaskFormer):
"""
Main class for zero shot mask classification semantic segmentation architectures.
"""
@configurable
def __init__(
self,
*,
backbone: Backbone,
sem_seg_head: nn.Module,
clip_adapter: nn.Module,
criterion: nn.Module,
num_queries: int,
panoptic_on: bool,
object_mask_threshold: float,
overlap_threshold: float,
metadata,
size_divisibility: int,
sem_seg_postprocess_before_inference: bool,
clip_ensemble: bool,
clip_ensemble_weight: float,
pixel_mean: Tuple[float],
pixel_std: Tuple[float],
):
"""
Args:
backbone: a backbone module, must follow detectron2's backbone interface
sem_seg_head: a module that predicts semantic segmentation from backbone features
criterion: a module that defines the loss
clip_adapter: adapter for clip-based mask classification
num_queries: int, number of queries
panoptic_on: bool, whether to output panoptic segmentation prediction
object_mask_threshold: float, threshold to filter query based on classification score
for panoptic segmentation inference
overlap_threshold: overlap threshold used in general inference for panoptic segmentation
metadata: dataset meta, get `thing` and `stuff` category names for panoptic
segmentation inference
size_divisibility: Some backbones require the input height and width to be divisible by a
specific integer. We can use this to override such requirement.
sem_seg_postprocess_before_inference: whether to resize the prediction back
to original input size before semantic segmentation inference or after.
For high-resolution dataset like Mapillary, resizing predictions before
inference will cause OOM error.
pixel_mean, pixel_std: list or tuple with #channels element, representing
the per-channel mean and std to be used to normalize the input image
"""
super().__init__(
backbone=backbone,
sem_seg_head=sem_seg_head,
criterion=criterion,
num_queries=num_queries,
panoptic_on=panoptic_on,
object_mask_threshold=object_mask_threshold,
overlap_threshold=overlap_threshold,
metadata=metadata,
size_divisibility=size_divisibility,
sem_seg_postprocess_before_inference=sem_seg_postprocess_before_inference,
pixel_mean=pixel_mean,
pixel_std=pixel_std,
)
self.clip_adapter: ClipAdapter = clip_adapter
self.clip_ensemble: bool = clip_ensemble
self.clip_ensemble_weight: float = clip_ensemble_weight
@classmethod
def from_config(cls, cfg):
init_kwargs = MaskFormer.from_config(cfg)
text_templates = build_text_prompt(cfg.MODEL.CLIP_ADAPTER)
clip_adapter = MaskFormerClipAdapter(
cfg.MODEL.CLIP_ADAPTER.CLIP_MODEL_NAME,
text_templates,
mask_fill=cfg.MODEL.CLIP_ADAPTER.MASK_FILL,
mask_expand_ratio=cfg.MODEL.CLIP_ADAPTER.MASK_EXPAND_RATIO,
mask_thr=cfg.MODEL.CLIP_ADAPTER.MASK_THR,
mask_matting=cfg.MODEL.CLIP_ADAPTER.MASK_MATTING,
region_resized=cfg.MODEL.CLIP_ADAPTER.REGION_RESIZED,
mask_prompt_depth=cfg.MODEL.CLIP_ADAPTER.MASK_PROMPT_DEPTH,
mask_prompt_fwd=cfg.MODEL.CLIP_ADAPTER.MASK_PROMPT_FWD,
)
init_kwargs["clip_adapter"] = clip_adapter
init_kwargs["clip_ensemble"] = cfg.MODEL.CLIP_ADAPTER.CLIP_ENSEMBLE
init_kwargs[
"clip_ensemble_weight"
] = cfg.MODEL.CLIP_ADAPTER.CLIP_ENSEMBLE_WEIGHT
return init_kwargs
def forward(self, batched_inputs):
"""
Args:
batched_inputs: a list, batched outputs of :class:`DatasetMapper`.
Each item in the list contains the inputs for one image.
For now, each item in the list is a dict that contains:
* "image": Tensor, image in (C, H, W) format.
* "instances": per-region ground truth
* Other information that's included in the original dicts, such as:
"height", "width" (int): the output resolution of the model (may be different
from input resolution), used in inference.
Returns:
list[dict]:
each dict has the results for one image. The dict contains the following keys:
* "sem_seg":
A Tensor that represents the
per-pixel segmentation prediced by the head.
The prediction has shape KxHxW that represents the logits of
each class for each pixel.
* "panoptic_seg":
A tuple that represent panoptic output
panoptic_seg (Tensor): of shape (height, width) where the values are ids for each segment.
segments_info (list[dict]): Describe each segment in `panoptic_seg`.
Each dict contains keys "id", "category_id", "isthing".
"""
images = [x["image"].to(self.device) for x in batched_inputs]
images = [(x - self.pixel_mean) / self.pixel_std for x in images]
images = ImageList.from_tensors(images, self.size_divisibility)
features = self.backbone(images.tensor)
outputs = self.sem_seg_head(features)
class_names = batched_inputs[0]["class_names"]
if len(class_names) == 1:
# Because classification is performed in a 'contrastive' manner, adding others to represent other concepts
class_names.append('others')
text_features = self.clip_adapter.get_text_features(class_names)
outputs["pred_logits"] = self.clip_adapter.get_sim_logits(
text_features, self.clip_adapter.normalize_feature(outputs["pred_logits"])
)
mask_cls_results = outputs["pred_logits"]
mask_pred_results = outputs["pred_masks"]
# upsample masks
mask_pred_results = F.interpolate(
mask_pred_results,
size=(images.tensor.shape[-2], images.tensor.shape[-1]),
mode="bilinear",
align_corners=False,
)
processed_results = []
for mask_cls_result, mask_pred_result, input_per_image, image_size in zip(
mask_cls_results, mask_pred_results, batched_inputs, images.image_sizes
):
height = image_size[0]
width = image_size[1]
mask_pred_result = sem_seg_postprocess(
mask_pred_result, image_size, height, width
)
image = input_per_image["image"].to(self.device)
r, regions = self.demo_inference(mask_cls_result, mask_pred_result, image, class_names)
height = input_per_image.get("height", image_size[0])
width = input_per_image.get("width", image_size[1])
r = sem_seg_postprocess(r, image_size, height, width)
processed_results.append({"sem_seg": r})
return processed_results
def demo_inference(self, mask_cls, mask_pred, image, class_names):
mask_cls = F.softmax(mask_cls, dim=-1)[..., :-1]
mask_pred = mask_pred.sigmoid()
regions = None
if self.clip_ensemble:
clip_cls, regions, valid_flag = self.clip_adapter(
image, class_names, mask_pred, normalize=True
)
if clip_cls is None:
clip_cls = torch.empty(0, mask_cls.shape[-1] + 1, device=self.device)
# softmax before index or after?
clip_cls = F.softmax(clip_cls[:, :-1], dim=-1)
if self.clip_ensemble_weight > 0:
map_back_clip_cls = mask_cls.new_ones(mask_cls.shape)
map_back_clip_cls[valid_flag] = clip_cls
mask_cls = torch.pow(mask_cls, 1 - self.clip_ensemble_weight) * \
torch.pow(map_back_clip_cls, self.clip_ensemble_weight)
else:
# only clip model predictions are used
mask_cls = clip_cls
mask_pred = mask_pred[valid_flag]
bin_mask = mask_pred > self.clip_adapter.mask_thr
select_cls = torch.zeros(sum(valid_flag), mask_cls.shape[-1], device=self.device)
select_mask = torch.argmax(mask_cls, dim=0)
if len(class_names) == 2 and class_names[-1] == 'others':
select_mask = select_mask[:-1]
for idx in select_mask:
select_cls[idx] = mask_cls[idx]
semseg = torch.einsum("qc,qhw->chw", select_cls, bin_mask.float())
return semseg, regions
|