def postprocess_detections(
self, class_logits: Tensor, box_regression: Tensor,
proposals: List[Tensor], image_shapes: List[Tuple[int, int]]
) -> Tuple[List[Tensor], List[Tensor], List[Tensor]]:
device = class_logits.device
num_classes = class_logits.shape[-1]
boxes_per_image = [
boxes_in_image.shape[0] for boxes_in_image in proposals
]
pred_boxes = self.box_coder.decode(box_regression, proposals)
pred_scores = F.softmax(class_logits, -1)
pred_boxes_list = pred_boxes.split(boxes_per_image, 0)
pred_scores_list = pred_scores.split(boxes_per_image, 0)
all_boxes = []
all_scores = []
all_labels = []
for boxes, scores, image_shape in zip(pred_boxes_list,
pred_scores_list, image_shapes):
boxes = clip_boxes_to_image(boxes, image_shape)
# create labels for each prediction
labels = torch.arange(num_classes, device=device)
labels = labels.view(1, -1).expand_as(scores)
# remove predictions with the background label
boxes = boxes[:, 1:]
scores = scores[:, 1:]
labels = labels[:, 1:]
# batch everything, by making every class prediction be a separate instance
boxes = boxes.reshape(-1, 4)
scores = scores.reshape(-1)
labels = labels.reshape(-1)
# remove low scoring boxes
inds = torch.where(scores > self.score_thresh)[0]
boxes, scores, labels = boxes[inds], scores[inds], labels[inds]
# remove empty boxes
keep = remove_small_boxes(boxes, min_size=1e-2)
boxes, scores, labels = boxes[keep], scores[keep], labels[keep]
# non-maximum suppression, independently done per class
keep = batched_nms(boxes, scores, labels, self.nms_thresh)
# keep only topk scoring predictions
keep = keep[:self.detections_per_img]
boxes, scores, labels = boxes[keep], scores[keep], labels[keep]
all_boxes.append(boxes)
all_scores.append(scores)
all_labels.append(labels)
return all_boxes, all_scores, all_labels
def postprocess_detections(self, class_logits, box_regression, proposals, image_shapes):
device = class_logits.device
num_classes = class_logits.shape[-1]
boxes_per_image = [len(boxes_in_image) for boxes_in_image in proposals]
pred_boxes = self.box_coder.decode(box_regression, proposals)
pred_scores = F.softmax(class_logits, -1)
# split boxes and scores per image
if len(boxes_per_image) == 1:
# TODO : remove this when ONNX support dynamic split sizes
pred_boxes = (pred_boxes,)
pred_scores = (pred_scores,)
else:
pred_boxes = pred_boxes.split(boxes_per_image, 0)
pred_scores = pred_scores.split(boxes_per_image, 0)
all_boxes = []
all_scores = []
all_labels = []
for boxes, scores, image_shape in zip(pred_boxes, pred_scores, image_shapes):
boxes = box_ops.clip_boxes_to_image(boxes, image_shape)
# create labels for each prediction
labels = torch.arange(num_classes, device=device)
labels = labels.view(1, -1).expand_as(scores)
# remove predictions with the background label
boxes = boxes[:, 1:]
scores = scores[:, 1:]
labels = labels[:, 1:]
# batch everything, by making every class prediction be a separate instance
boxes = boxes.reshape(-1, 4)
scores = scores.reshape(-1)
labels = labels.reshape(-1)
# remove low scoring boxes
inds = torch.nonzero(scores > self.score_thresh).squeeze(1)
boxes, scores, labels = boxes[inds], scores[inds], labels[inds]
# remove empty boxes
keep = box_ops.remove_small_boxes(boxes, min_size=1e-2)
boxes, scores, labels = boxes[keep], scores[keep], labels[keep]
# non-maximum suppression, independently done per class
keep = box_ops.batched_nms(boxes, scores, labels, self.nms_thresh)
# keep only topk scoring predictions
keep = keep[:self.detections_per_img]
boxes, scores, labels = boxes[keep], scores[keep], labels[keep]
all_boxes.append(boxes)
all_scores.append(scores)
all_labels.append(labels)
return all_boxes, all_scores, all_labels
def filter_proposals(self, proposals, objectness, image_shapes,
num_anchors_per_level):
# type: (Tensor, Tensor, List[Tuple[int, int]], List[int]) -> Tuple[List[Tensor], List[Tensor]]
num_images = proposals.shape[0]
device = proposals.device
# do not backprop throught objectness
objectness = objectness.detach()
objectness = objectness.reshape(num_images, -1)
levels = [
torch.full((n, ), idx, dtype=torch.int64, device=device)
for idx, n in enumerate(num_anchors_per_level)
]
levels = torch.cat(levels, 0)
levels = levels.reshape(1, -1).expand_as(objectness)
# select top_n boxes independently per level before applying nms
top_n_idx = self._get_top_n_idx(objectness, num_anchors_per_level)
image_range = torch.arange(num_images, device=device)
batch_idx = image_range[:, None]
objectness = objectness[batch_idx, top_n_idx]
levels = levels[batch_idx, top_n_idx]
proposals = proposals[batch_idx, top_n_idx]
objectness_prob = torch.sigmoid(objectness)
final_boxes = []
final_scores = []
for boxes, scores, lvl, img_shape in zip(proposals, objectness_prob,
levels, image_shapes):
boxes = box_ops.clip_boxes_to_image(boxes, img_shape)
# remove small boxes
keep = box_ops.remove_small_boxes(boxes, self.min_size)
boxes, scores, lvl = boxes[keep], scores[keep], lvl[keep]
# remove low scoring boxes
# use >= for Backwards compatibility
keep = torch.where(scores >= self.score_thresh)[0]
boxes, scores, lvl = boxes[keep], scores[keep], lvl[keep]
# non-maximum suppression, independently done per level
keep = box_ops.batched_nms(boxes, scores, lvl, self.nms_thresh)
# keep only topk scoring predictions
keep = keep[:self.post_nms_top_n()]
boxes, scores = boxes[keep], scores[keep]
final_boxes.append(boxes)
final_scores.append(scores)
return final_boxes, final_scores
def postprocess_detections(
self,
pred_scores, # type: Tensor
pred_boxes, # type: Tensor
proposals, # type: List[Tensor]
image_shapes # type: List[Tuple[int, int]]
):
# type: (...) -> Tuple[List[Tensor], List[Tensor], List[Tensor]]
# device = class_logits.device
# num_classes = class_logits.shape[-1]
boxes_per_image = [
boxes_in_image.shape[0] for boxes_in_image in proposals
]
pred_boxes_list = pred_boxes.split(boxes_per_image, 0)
pred_scores_list = pred_scores.split(boxes_per_image, 0)
all_boxes = []
all_scores = []
all_labels = []
for boxes, scores, image_shape in zip(pred_boxes_list,
pred_scores_list, image_shapes):
boxes = box_ops.clip_boxes_to_image(boxes, image_shape)
#去掉背景类的得分
scores = scores[:, 1:]
# labels = labels[:, 1:]
scores, labels = scores.max(dim=1)
labels += 1 #目标的标签是从1开始的
# remove low scoring boxes
inds = torch.where(scores > self.score_thresh)[0]
boxes, scores, labels = boxes[inds], scores[inds], labels[inds]
# remove empty boxes
keep = box_ops.remove_small_boxes(boxes, min_size=1e-2)
boxes, scores, labels = boxes[keep], scores[keep], labels[keep]
# non-maximum suppression, independently done per class
keep = box_ops.batched_nms(boxes, scores, labels, self.nms_thresh)
# keep only topk scoring predictions
keep = keep[:self.detections_per_img]
boxes, scores, labels = boxes[keep], scores[keep], labels[keep]
all_boxes.append(boxes)
all_scores.append(scores)
all_labels.append(labels)
return all_boxes, all_scores, all_labels
def generate_anchors(self, x: Tensor) -> None:
anchors = torch.cat([
_generate_anchors(self.input_size, x.size(-1),
listify(anchor_sizes), self.aspect_ratios,
stride)
for anchor_sizes, stride in zip(self.anchor_sizes, self.strides)
],
dim=0)
# Filter anchors
anchors = box_ops.clip_boxes_to_image(
anchors, (self.input_size, self.input_size))
keep = box_ops.remove_small_boxes(anchors, 1e-3)
self.anchors = anchors[keep]
def filter_predictions(self, boxes, labels, scores):
inds = torch.nonzero(scores > self.score_thresh).squeeze(1)
boxes, scores, labels = boxes[inds], scores[inds], labels[inds]
# remove empty boxes
keep = box_ops.remove_small_boxes(boxes, min_size=1e-2)
boxes, scores, labels = boxes[keep], scores[keep], labels[keep]
# non-maximum suppression, independently done per class
keep = box_ops.batched_nms(boxes, scores, labels, self.nms_thresh)
# keep only topk scoring predictions
keep = keep[:self.num_boxes_per_img]
boxes, scores, labels = boxes[keep], scores[keep], labels[keep]
return boxes, scores, labels
def filter_proposals_patch(self, proposals, objectness, image_shapes,
num_anchors_per_level):
# type: (Tensor, Tensor, List[Tuple[int, int]], List[int])
print("patch:filter_proposals")
num_images = proposals.shape[0]
device = proposals.device
objectness = objectness.detach()
objectness = objectness.reshape(num_images, -1)
levels = [
torch.full((n, ), idx, dtype=torch.int64, device=device)
for idx, n in enumerate(num_anchors_per_level)
]
levels = torch.cat(levels, 0)
levels = levels.reshape(1, -1).expand_as(objectness)
top_n_idx = self._get_top_n_idx(objectness,
num_anchors_per_level) #All call the patch
image_range = torch.arange(num_images, device=device)
batch_idx = image_range[:, None]
objectness = objectness[batch_idx, top_n_idx]
levels = levels[batch_idx, top_n_idx]
proposals = proposals[batch_idx, top_n_idx]
final_boxes = []
final_scores = []
for boxes, scores, lvl, img_shape in zip(proposals, objectness, levels,
image_shapes):
# boxes = box_ops.clip_boxes_to_image(boxes, img_shape)
boxes = clip_boxes_to_image_patch(boxes, img_shape)
keep = box_ops.remove_small_boxes(boxes, self.min_size)
boxes, scores, lvl = boxes[keep], scores[keep], lvl[keep]
# non-maximum suppression, independently done per level
keep = box_ops.batched_nms(boxes, scores, lvl, self.nms_thresh)
# keep only topk scoring predictions
keep = keep[:self.post_nms_top_n()]
boxes, scores = boxes[keep], scores[keep]
final_boxes.append(boxes)
final_scores.append(scores)
return final_boxes, final_scores
def crop(img: Image, target: Dict[str, Any],
region: Tuple[int]) -> Tuple[Image, Dict[str, Any]]:
"""
Args:
region: [Top, Left, H, W]
"""
# crop image
src_w, src_h = img.size
img = TF.crop(img, *region)
target = deepcopy(target)
top, left, h, w = region
# set new image size
if "size" in target.keys():
target["size"] = (h, w)
fields: List[str] = list()
for k, v in target.items():
if isinstance(v, Tensor):
fields.append(k)
# crop bounding boxes
if "boxes" in target:
boxes = target["boxes"]
boxes[:, [0, 2]] *= src_w
boxes[:, [1, 3]] *= src_h
boxes = box_op.box_convert(boxes, "cxcywh", "xyxy")
boxes -= torch.tensor([left, top, left, top])
boxes = box_op.clip_boxes_to_image(boxes, (h, w))
keep = box_op.remove_small_boxes(boxes, 1)
boxes[:, [0, 2]] /= w
boxes[:, [1, 3]] /= h
boxes = box_op.box_convert(boxes, "xyxy", "cxcywh")
target["boxes"] = boxes
for field in fields:
target[field] = target[field][keep]
if "masks" in target:
target['masks'] = target['masks'][:, top:top + h, left:left + w]
keep = target['masks'].flatten(1).any(1)
for field in fields:
target[field] = target[field][keep]
return img, target
def filter_proposals(self, proposals, objectness, image_shapes,
num_anchors_per_level):
# type: (Tensor, Tensor, List[Tuple[int, int]], List[int])
num_images = proposals.shape[0]
device = proposals.device
# do not backprop throught objectness
objectness = objectness.detach()
objectness = objectness.reshape(num_images, -1)
levels = [
torch.full((n, ), idx, dtype=torch.int64, device=device)
for idx, n in enumerate(num_anchors_per_level)
]
levels = torch.cat(levels, 0)
levels = levels.reshape(1, -1).expand_as(objectness)
# select top_n boxes independently per level before applying nms
top_n_idx = self._get_top_n_idx(objectness, num_anchors_per_level)
image_range = torch.arange(num_images, device=device)
batch_idx = image_range[:, None]
objectness = objectness[batch_idx, top_n_idx]
levels = levels[batch_idx, top_n_idx]
proposals = proposals[batch_idx, top_n_idx]
final_boxes = []
final_scores = []
for boxes, scores, lvl, img_shape in zip(proposals, objectness, levels,
image_shapes):
# For onnx export, Clip's min max can not be traced as tensor.
if torchvision._is_tracing():
boxes = _onnx_clip_boxes_to_image(boxes, img_shape)
else:
boxes = box_ops.clip_boxes_to_image(boxes, img_shape)
keep = box_ops.remove_small_boxes(boxes, self.min_size)
boxes, scores, lvl = boxes[keep], scores[keep], lvl[keep]
# non-maximum suppression, independently done per level
keep = box_ops.batched_nms(boxes, scores, lvl, self.nms_thresh)
# keep only topk scoring predictions
keep = keep[:self.post_nms_top_n()]
boxes, scores = boxes[keep], scores[keep]
final_boxes.append(boxes)
final_scores.append(scores)
return final_boxes, final_scores
def filter_proposals(self, proposals, objectness, image_shapes,
num_anchors_per_level):
num_images = proposals.shape[0]
device = proposals.device
# do not backprop throught objectness
objectness = objectness.detach()
objectness = objectness.reshape(num_images, -1)
levels = [
torch.full((n, ), idx, dtype=torch.int64, device=device)
for idx, n in enumerate(num_anchors_per_level)
]
levels = torch.cat(levels, 0)
levels = levels.reshape(1, -1).expand_as(objectness)
# select top_n boxes independently per level before applying nms
top_n_idx = self._get_top_n_idx(objectness, num_anchors_per_level)
batch_idx = torch.arange(num_images, device=device)[:, None]
objectness = objectness[batch_idx, top_n_idx]
levels = levels[batch_idx, top_n_idx]
proposals = proposals[batch_idx, top_n_idx]
final_boxes = []
final_scores = []
for boxes, scores, lvl, img_shape in zip(proposals, objectness, levels,
image_shapes):
boxes = box_ops.clip_boxes_to_image(boxes, img_shape)
keep = box_ops.remove_small_boxes(boxes, self.min_size)
boxes, scores, lvl = boxes[keep], scores[keep], lvl[keep]
# non-maximum suppression, independently done per level
#lvl=torch.tensor(np.arange(len(lvl))).to(device)
keep = box_ops.batched_nms(boxes, scores, lvl, self.nms_thresh)
# keep only topk scoring predictions
keep = keep[:self.post_nms_top_n]
boxes, scores = boxes[keep], scores[keep]
final_boxes.append(boxes)
final_scores.append(scores)
return final_boxes, final_scores
def postprocess_boxes(
self,
class_logits,
box_regression,
embeddings,
proposals,
image_shapes,
fcs=None,
gt_det=None,
cws=True,
):
"""
Similar to RoIHeads.postprocess_detections, but can handle embeddings and implement
First Classification Score (FCS).
"""
device = class_logits.device
boxes_per_image = [len(boxes_in_image) for boxes_in_image in proposals]
pred_boxes = self.box_coder.decode(box_regression, proposals)
if fcs is not None:
# Fist Classification Score (FCS)
pred_scores = fcs[0]
else:
pred_scores = torch.sigmoid(class_logits)
if cws:
# Confidence Weighted Similarity (CWS)
embeddings = embeddings * pred_scores.view(-1, 1)
# split boxes and scores per image
pred_boxes = pred_boxes.split(boxes_per_image, 0)
pred_scores = pred_scores.split(boxes_per_image, 0)
pred_embeddings = embeddings.split(boxes_per_image, 0)
all_boxes = []
all_scores = []
all_labels = []
all_embeddings = []
for boxes, scores, embeddings, image_shape in zip(
pred_boxes, pred_scores, pred_embeddings, image_shapes):
boxes = box_ops.clip_boxes_to_image(boxes, image_shape)
# create labels for each prediction
labels = torch.ones(scores.size(0), device=device)
# remove predictions with the background label
boxes = boxes[:, 1:]
scores = scores.unsqueeze(1)
labels = labels.unsqueeze(1)
# batch everything, by making every class prediction be a separate instance
boxes = boxes.reshape(-1, 4)
scores = scores.flatten()
labels = labels.flatten()
embeddings = embeddings.reshape(-1, self.embedding_head.dim)
# remove low scoring boxes
inds = torch.nonzero(scores > self.score_thresh).squeeze(1)
boxes, scores, labels, embeddings = (
boxes[inds],
scores[inds],
labels[inds],
embeddings[inds],
)
# remove empty boxes
keep = box_ops.remove_small_boxes(boxes, min_size=1e-2)
boxes, scores, labels, embeddings = (
boxes[keep],
scores[keep],
labels[keep],
embeddings[keep],
)
if gt_det is not None:
# include GT into the detection results
boxes = torch.cat((boxes, gt_det["boxes"]), dim=0)
labels = torch.cat((labels, torch.tensor([1.0]).to(device)),
dim=0)
scores = torch.cat((scores, torch.tensor([1.0]).to(device)),
dim=0)
embeddings = torch.cat((embeddings, gt_det["embeddings"]),
dim=0)
# non-maximum suppression, independently done per class
keep = box_ops.batched_nms(boxes, scores, labels, self.nms_thresh)
# keep only topk scoring predictions
keep = keep[:self.detections_per_img]
boxes, scores, labels, embeddings = (
boxes[keep],
scores[keep],
labels[keep],
embeddings[keep],
)
all_boxes.append(boxes)
all_scores.append(scores)
all_labels.append(labels)
all_embeddings.append(embeddings)
return all_boxes, all_scores, all_embeddings, all_labels
def roi_postprocess_detections(self, class_logits, box_regression, proposals,
image_shapes, *extra_tensors):
"""Hack into the torchvision model to obtain features for
training caption model; training is assumed to be false
https://github.com/pytorch/vision/blob/master/
torchvision/models/detection/roi_heads.py"""
device = class_logits.device
num_classes = class_logits.shape[-1]
boxes_per_image = [len(boxes_in_image) for boxes_in_image in proposals]
pred_boxes = self.box_coder.decode(box_regression, proposals)
pred_scores = F.softmax(class_logits, -1)
# split boxes and scores per image
if len(boxes_per_image) == 1:
# TODO : remove this when ONNX support dynamic split sizes
# and just assign to pred_boxes instead of pred_boxes_list
pred_boxes_list = [pred_boxes]
pred_scores_list = [pred_scores]
extra_tensors_list = [[x] for x in extra_tensors]
else:
pred_boxes_list = pred_boxes.split(boxes_per_image, 0)
pred_scores_list = pred_scores.split(boxes_per_image, 0)
extra_tensors_list = [
x.split(boxes_per_image, 0) for x in extra_tensors
]
all_boxes = []
all_scores = []
all_labels = []
all_extras = [[] for _ in extra_tensors]
for boxes, scores, image_shape, *extras in zip(pred_boxes_list,
pred_scores_list,
image_shapes,
*extra_tensors_list):
boxes = box_ops.clip_boxes_to_image(boxes, image_shape)
# create labels for each prediction
labels = torch.arange(num_classes, device=device)
labels = labels.view(1, -1).expand_as(scores)
# remove predictions with the background label
boxes = boxes[:, 1:]
scores = scores[:, 1:]
labels = labels[:, 1:]
# batch everything, by making every class prediction be a separate instance
boxes = boxes.reshape(-1, 4)
scores = scores.reshape(-1)
labels = labels.reshape(-1)
# remove low scoring boxes
inds = torch.nonzero(scores > self.score_thresh).squeeze(1)
boxes, scores, labels = boxes[inds], scores[inds], labels[inds]
# each feature vector is used for all 91 class predictions
# there are 90 classes (minus the background)
# take the feature vector corresponding to each class
extras = [x[inds // (num_classes - 1)] for x in extras]
# remove empty boxes
keep = box_ops.remove_small_boxes(boxes, min_size=1e-2)
boxes, scores, labels = boxes[keep], scores[keep], labels[keep]
# each feature vector is used for all 91 class predictions
# there are 90 classes (minus the background)
# take the feature vector corresponding to each class
extras = [x[keep] for x in extras]
# non-maximum suppression, independently done per class
keep = box_ops.batched_nms(boxes, scores, labels, self.nms_thresh)
# keep only topk scoring predictions
keep = keep[:self.detections_per_img]
boxes, scores, labels = boxes[keep], scores[keep], labels[keep]
# each feature vector is used for all 91 class predictions
# there are 90 classes (minus the background)
# take the feature vector corresponding to each class
extras = [x[keep] for x in extras]
all_boxes.append(boxes)
all_scores.append(scores)
all_labels.append(labels)
for x, y in zip(all_extras, extras):
x.append(y)
return [all_boxes, all_scores, all_labels, *all_extras]
def box_features_hook(self, module, input, output):
'''
hook for extracting features from MaskRCNN
'''
features, proposals, image_shapes, targets = input
box_features = module.box_roi_pool(features, proposals, image_shapes)
box_features = module.box_head(box_features)
class_logits, box_regression = module.box_predictor(box_features)
device = class_logits.device
num_classes = class_logits.shape[-1]
boxes_per_image = [len(boxes_in_image) for boxes_in_image in proposals]
pred_boxes = module.box_coder.decode(box_regression, proposals)
pred_scores = F.softmax(class_logits, -1)
# split boxes and scores per image
pred_boxes = pred_boxes.split(boxes_per_image, 0)
pred_scores = pred_scores.split(boxes_per_image, 0)
all_boxes = []
all_scores = []
all_labels = []
all_keeps = []
for boxes, scores, image_shape in zip(pred_boxes, pred_scores,
image_shapes):
boxes = box_ops.clip_boxes_to_image(boxes, image_shape)
# create labels for each prediction
labels = torch.arange(num_classes, device=device)
labels = labels.view(1, -1).expand_as(scores)
# remove predictions with the background label
boxes = boxes[:, 1:]
scores = scores[:, 1:]
labels = labels[:, 1:]
# batch everything, by making every class prediction be a separate instance
boxes = boxes.reshape(-1, 4)
scores = scores.flatten()
labels = labels.flatten()
# remove low scoring boxes
inds = torch.nonzero(scores > module.score_thresh).squeeze(1)
boxes, scores, labels = boxes[inds], scores[inds], labels[inds]
# remove empty boxes
keep = box_ops.remove_small_boxes(boxes, min_size=1e-2)
boxes, scores, labels = boxes[keep], scores[keep], labels[keep]
# non-maximum suppression, independently done per class
keep = box_ops.batched_nms(boxes, scores, labels,
module.nms_thresh)
# keep only topk scoring predictions
keep = keep[:self.mask_rcnn_top_k_boxes]
boxes, scores, labels = boxes[keep], scores[keep], labels[keep]
all_boxes.append(boxes)
all_scores.append(scores)
all_labels.append(labels)
all_keeps.append(keep)
box_features_per_image = []
for keep in all_keeps:
box_features_per_image.append(box_features[keep])
self.detection_box_features = box_features_per_image
self.fpn_pooled_features = self.avg2dpool(
features['pool']).squeeze(-1).squeeze(-1)
def postprocess_detections(self, pred_scores, box_regression, embeddings_,
proposals, image_shapes):
device = pred_scores.device
num_classes = pred_scores.shape[-1]
boxes_per_image = [len(boxes_in_image) for boxes_in_image in proposals]
pred_boxes = self.box_coder.decode(box_regression, proposals)
# split boxes and scores per image
pred_boxes = pred_boxes.split(boxes_per_image, 0)
pred_scores = pred_scores.split(boxes_per_image, 0)
pred_embeddings = embeddings_.split(boxes_per_image, 0)
all_boxes = []
all_scores = []
all_labels = []
all_embeddings = []
for boxes, scores, embeddings, image_shape in zip(
pred_boxes, pred_scores, pred_embeddings, image_shapes):
boxes = box_ops.clip_boxes_to_image(boxes, image_shape)
# create labels for each prediction
labels = torch.arange(num_classes, device=device)
labels = labels.view(1, -1).expand_as(scores)
# remove predictions with the background label
boxes = boxes[:, 1:]
scores = scores[:, 1:]
labels = labels[:, 1:]
# embeddings are already personized.
# batch everything, by making every class prediction be a separate
# instance
boxes = boxes.reshape(-1, 4)
scores = scores.flatten()
labels = labels.flatten()
embeddings = embeddings.reshape(-1, self.embedding_head.dim)
# remove low scoring boxes
inds = torch.nonzero(scores > self.score_thresh).squeeze(1)
boxes, scores, labels, embeddings = boxes[inds], scores[
inds], labels[inds], embeddings[inds]
# remove empty boxes
keep = box_ops.remove_small_boxes(boxes, min_size=1e-2)
boxes, scores, labels, embeddings = boxes[keep], scores[keep], \
labels[keep], embeddings[keep]
# non-maximum suppression, independently done per class
keep = box_ops.batched_nms(boxes, scores, labels, self.nms_thresh)
# keep only topk scoring predictions
keep = keep[:self.detections_per_img]
boxes, scores, labels, embeddings = boxes[keep], scores[keep], \
labels[keep], embeddings[keep]
all_boxes.append(boxes)
all_scores.append(scores)
all_labels.append(labels)
all_embeddings.append(embeddings)
return all_boxes, all_scores, all_embeddings, all_labels
def post_processor(self, class_logits, box_regression, proposals,
img_metas):
num_classes = class_logits.shape[1]
device = class_logits.device
boxes_per_image = [box.shape[0] for box in proposals]
proposals = cat([box for box in proposals])
pred_boxes = self.box_coder.decode(
box_regression.view(sum(boxes_per_image), -1), proposals)
pred_boxes = pred_boxes.reshape(sum(boxes_per_image), -1, 4)
pred_scores = F.softmax(class_logits, -1)
# split boxes and scores per image
if len(boxes_per_image) == 1:
pred_boxes = (pred_boxes, )
pred_scores = (pred_scores, )
else:
pred_boxes = pred_boxes.split(boxes_per_image,
dim=0) # (N, #CLS, 4)
pred_scores = pred_scores.split(boxes_per_image,
dim=0) # (N, #CLS)
results = []
for scores, boxes, img_meta in zip(pred_scores, pred_boxes, img_metas):
width, height = img_meta['img_shape']
boxes = box_ops.clip_boxes_to_image(boxes, (height, width))
# create labels for each prediction
labels = torch.arange(num_classes, device=device)
labels = labels.view(1, -1).expand_as(scores)
# remove predictions with the background label
boxes = boxes[:, 1:]
scores = scores[:, 1:]
labels = labels[:, 1:]
# batch everything, by making every class prediction be a separate instance
boxes = boxes.reshape(-1, 4)
scores = scores.reshape(-1)
labels = labels.reshape(-1)
# remove low scoring boxes
inds = torch.nonzero(scores > self.score_thresh).squeeze(1)
boxes, scores, labels = boxes[inds], scores[inds], labels[inds]
# remove empty boxes
keep = box_ops.remove_small_boxes(boxes, min_size=1)
boxes, scores, labels = boxes[keep], scores[keep], labels[keep]
# non-maximum suppression, independently done per class
keep = box_ops.batched_nms(boxes, scores, labels, self.nms_thresh)
# keep only topk scoring predictions
keep = keep[:self.detections_per_img]
boxes, scores, labels = boxes[keep], scores[keep], labels[keep]
result = {
'boxes': boxes,
'scores': scores,
'labels': labels,
}
results.append(result)
return results
def postprocess_detections(self, class_logits, attr_logits, box_regression,
proposals, image_shapes):
# type: (Tensor, Tensor, List[Tensor], List[Tuple[int, int]])
# class_logits: (1000,47), attr-logits: (1000,295)
device = class_logits.device
num_classes = class_logits.shape[-1]
num_attrs = attr_logits.shape[-1]
boxes_per_image = [len(boxes_in_image) for boxes_in_image in proposals]
pred_boxes = self.box_coder.decode(box_regression, proposals)
pred_scores = F.softmax(class_logits, -1)
pred_ascores = F.sigmoid(attr_logits)
# print(pred_ascores)
# set_trace()
# split boxes and scores per image
if len(boxes_per_image) == 1:
# TODO : remove this when ONNX support dynamic split sizes
# and just assign to pred_boxes instead of pred_boxes_list
pred_boxes_list = [pred_boxes]
pred_scores_list = [pred_scores]
pred_ascores_list = [pred_ascores]
else:
pred_boxes_list = pred_boxes.split(boxes_per_image,
0) # (bs,1000,47,4)
pred_scores_list = pred_scores.split(boxes_per_image,
0) # (bs, 1000, 47)
pred_ascores_list = pred_ascores.split(boxes_per_image, 0)
all_boxes = []
all_scores = []
all_labels = []
all_attrs = []
for boxes, scores, ascores, image_shape in zip(
pred_boxes_list, pred_scores_list, pred_ascores_list,
image_shapes): # for each image
boxes = box_ops.clip_boxes_to_image(boxes, image_shape)
# create labels for each prediction
labels = torch.arange(num_classes, device=device)
labels = labels.view(1,
-1).expand_as(scores) # (1000,47) 都是从0,46结束
# remove predictions with the background label
boxes = boxes[:, 1:]
scores = scores[:, 1:]
labels = labels[:, 1:]
ascores = ascores[:, 1:] # (1000,294)
# batch everything, by making every class prediction be a separate instance
boxes = boxes.reshape(-1, 4) # (46000,4)
scores = scores.reshape(-1) # (46000,)
labels = labels.reshape(-1) # (46000,)
ascores = ascores.unsqueeze(0).repeat(46, 1, 1).reshape(
-1, num_attrs - 1) # (46,1000,294)->(46000,294)
# remove low scoring boxes
inds = torch.nonzero(scores > self.score_thresh).squeeze(
1) # 压缩成1d的，用thresh去卡
boxes, scores, labels, ascores = boxes[inds], scores[inds], labels[
inds], ascores[inds]
# remove empty boxes
# attention!!! 压缩成1d的，用thresh去卡（每个box会预测46个类别的box，所以一个box对一个46个box，因此卡thresh是对类别单位的box进行的）
keep = box_ops.remove_small_boxes(boxes, min_size=1e-2)
boxes, scores, labels, ascores = boxes[keep], scores[keep], labels[
keep], ascores[keep]
# non-maximum suppression, independently done per class
keep = box_ops.batched_nms(boxes, scores, labels, self.nms_thresh)
# keep only topk scoring predictions
keep = keep[:self.detections_per_img]
boxes, scores, labels, ascores = boxes[keep], scores[keep], labels[
keep], ascores[keep]
all_boxes.append(boxes)
all_scores.append(scores)
all_labels.append(labels)
# TODO: add attributes-wise thresh
# ascores: (post_rois, 294)
# brute force
ascore_list = []
for ascore in ascores:
ascore_list.append(
torch.where(ascore > self.attr_score_thresh)
[0]) # list(Tensor()) # 长度不均匀的一维tensor
all_attrs.append(ascore_list) # list[(post_rois, 294)]
# all_scores: list[(post_rois,)], all_labels: list[(post_rois,)], all_attrs: list[list[attr1,attr2], (...)] 每个list里有每个roi的attr tensor
return all_boxes, all_scores, all_labels, all_attrs
def postprocess_detections(
self,
class_logits, # type: Tensor
box_regression, # type: Tensor
proposals, # type: List[Tensor]
attr_logits,
image_shapes # type: List[Tuple[int, int]]
):
# type: (...) -> Tuple[List[Tensor], List[Tensor], List[Tensor]]
device = class_logits.device
num_classes = class_logits.shape[-1]
#print(class_logits.shape)
boxes_per_image = [
boxes_in_image.shape[0] for boxes_in_image in proposals
]
#print(boxes_per_image)
pred_boxes = self.box_coder.decode(box_regression, proposals)
pred_scores = F.softmax(class_logits, -1)
#print(pred_scores.shape)
pred_attr_scores = F.sigmoid(attr_logits)
#print(pred_attr_scores.shape)
pred_boxes_list = pred_boxes.split(boxes_per_image, 0)
pred_scores_list = pred_scores.split(boxes_per_image, 0)
pred_attr_scores_list = pred_attr_scores.split(boxes_per_image, 0)
all_boxes = []
all_scores = []
all_labels = []
all_attr_scores = []
all_attr_labels = []
for boxes, scores, attr_scores, image_shape in zip(
pred_boxes_list, pred_scores_list, pred_attr_scores_list,
image_shapes):
boxes = box_ops.clip_boxes_to_image(boxes, image_shape)
# create labels for each prediction
labels = torch.arange(num_classes, device=device)
#print(labels)
#print(scores.shape)
labels = labels.view(1, -1).expand_as(scores)
#print(labels.shape)
#print(attr_labels.shape)
detection_per_image = attr_scores.shape[0]
attr_classes = 341
curr_attr_shape = (detection_per_image, attr_classes)
new_attr_shape = (detection_per_image, num_classes - 1,
attr_classes)
#print(attr_scores)
attr_scores = attr_scores.unsqueeze(1)
#print("Unsqueeze: ", attr_scores.shape)
attr_scores = attr_scores.expand(*new_attr_shape)
#print("Expand: ", attr_scores.shape)
#print("New Val: ", attr_scores)
#print("Attr Shape: ", attr_labels.shape)
#print(attr_labels)
attr_labels = torch.arange(341, device=device)
attr_labels = attr_labels.view(1, 1, -1).expand_as(attr_scores)
#print(attr_labels.shape)
#print(attr_labels)
# remove predictions with the background label
boxes = boxes[:, 1:]
scores = scores[:, 1:]
labels = labels[:, 1:]
# batch everything, by making every class prediction be a separate instance
boxes = boxes.reshape(-1, 4)
scores = scores.reshape(-1)
labels = labels.reshape(-1)
attr_labels = attr_labels.reshape(-1, 341)
attr_scores = attr_scores.reshape(-1, 341)
#print(scores.shape)
#print(labels.shape)
# remove low scoring boxes
inds = torch.nonzero(scores > self.score_thresh).squeeze(1)
#print(inds)
boxes, scores, labels, attr_scores, attr_labels = boxes[
inds], scores[inds], labels[inds], attr_scores[
inds], attr_labels[inds]
# remove empty boxes
keep = box_ops.remove_small_boxes(boxes, min_size=1e-2)
boxes, scores, labels, attr_scores, attr_labels = boxes[
keep], scores[keep], labels[keep], attr_scores[
keep], attr_labels[keep]
# non-maximum suppression, independently done per class
keep = box_ops.batched_nms(boxes, scores, labels, self.nms_thresh)
# keep only topk scoring predictions
keep = keep[:self.detections_per_img]
boxes, scores, labels, attr_scores, attr_labels = boxes[
keep], scores[keep], labels[keep], attr_scores[
keep], attr_labels[keep]
all_boxes.append(boxes)
all_scores.append(scores)
all_labels.append(labels)
all_attr_scores.append(attr_scores)
all_attr_labels.append(attr_labels)
return all_boxes, all_scores, all_labels, all_attr_scores, all_attr_labels
def ssm_postprocess_detections(self, head_outputs, anchors, image_shapes):
# type: (Dict[str, Tensor], List[Tensor], List[Tuple[int, int]]) -> List[Dict[str, Tensor]]
# TODO: Merge this with roi_heads.RoIHeads.postprocess_detections ?
class_logits = head_outputs.pop('cls_logits')
box_regression = head_outputs.pop('bbox_regression')
other_outputs = head_outputs
device = class_logits.device
num_classes = class_logits.shape[-1]
scores = torch.sigmoid(class_logits)
# create labels for each score
labels = torch.arange(num_classes, device=device)
labels = labels.view(1, -1).expand_as(scores)
detections = torch.jit.annotate(List[Dict[str, Tensor]], [])
al_idx = 0
all_boxes = torch.empty([0, 4]).cuda()
all_scores = torch.tensor([]).cuda()
all_labels = []
CONF_THRESH = 0.5 # bigger leads more active learning samples
for index, (box_regression_per_image, scores_per_image, labels_per_image, anchors_per_image, image_shape) in \
enumerate(zip(box_regression, scores, labels, anchors, image_shapes)):
if torch.max(scores_per_image) < CONF_THRESH:
# print(scores)
al_idx = 1
detections.append({
"boxes": all_boxes,
"labels": all_labels,
"scores": all_scores,
'al': al_idx,
})
continue
boxes_per_image = self.box_coder.decode_single(
box_regression_per_image, anchors_per_image)
boxes_per_image = box_ops.clip_boxes_to_image(
boxes_per_image, image_shape)
other_outputs_per_image = [(k, v[index])
for k, v in other_outputs.items()]
image_boxes = []
image_scores = []
image_labels = []
image_other_outputs = torch.jit.annotate(Dict[str, List[Tensor]],
{})
for class_index in range(num_classes):
# remove low scoring boxes
inds = torch.gt(scores_per_image[:, class_index],
self.score_thresh)
boxes_per_class, scores_per_class, scores_all_class, labels_per_class = \
boxes_per_image[inds], scores_per_image[inds, class_index], scores_per_image[inds], \
labels_per_image[inds, class_index]
other_outputs_per_class = [(k, v[inds])
for k, v in other_outputs_per_image]
keep = [i for i in range(len(boxes_per_class))]
random.shuffle(keep)
keep = keep[:500]
boxes_per_class, scores_per_class, scores_all_class, labels_per_class = \
boxes_per_class[keep], scores_per_class[keep], scores_all_class[keep], labels_per_class[keep]
other_outputs_per_class = [(k, v[keep])
for k, v in other_outputs_per_class]
# remove empty boxes
keep = box_ops.remove_small_boxes(boxes_per_class,
min_size=1e-2)
boxes_per_class, scores_per_class, scores_all_class, labels_per_class = \
boxes_per_class[keep], scores_per_class[keep], scores_all_class[keep], labels_per_class[keep]
other_outputs_per_class = [(k, v[keep])
for k, v in other_outputs_per_class]
# non-maximum suppression, independently done per class
keep = box_ops.nms(boxes_per_class, scores_per_class,
self.nms_thresh)
# keep only topk scoring predictions
keep = keep[:self.detections_per_img]
boxes_per_class, scores_per_class, scores_all_class, labels_per_class = \
boxes_per_class[keep], scores_per_class[keep], scores_all_class[keep], labels_per_class[keep]
other_outputs_per_class = [(k, v[keep])
for k, v in other_outputs_per_class]
image_boxes.append(boxes_per_class)
image_scores.append(scores_per_class)
image_labels.append(labels_per_class)
for k, v in other_outputs_per_class:
if k not in image_other_outputs:
image_other_outputs[k] = []
image_other_outputs[k].append(v)
for i in range(len(boxes_per_class)):
all_boxes = torch.cat(
(all_boxes, boxes_per_class[i].unsqueeze(0)), 0)
all_scores = torch.cat(
(all_scores, scores_per_class[i].unsqueeze(0)), 0)
all_labels.append(judge_y(scores_all_class[i][1:]))
detections.append({
"boxes": all_boxes,
"labels": all_labels,
"scores": all_scores,
'al': al_idx,
})
for k, v in image_other_outputs.items():
detections[-1].update({k: torch.cat(v, dim=0)})
return detections
def postprocess_detections(
self,
class_logits, # type: Tensor
box_regression, # type: Tensor
proposals, # type: List[Tensor]
image_shapes # type: List[Tuple[int, int]]
):
# type: (...) -> Tuple[List[Tensor], List[Tensor], List[Tensor]]
device = class_logits.device
num_classes = class_logits.shape[-1]
boxes_per_image = [
boxes_in_image.shape[0] for boxes_in_image in proposals
]
pred_boxes = self.box_coder.decode(box_regression, proposals)
pred_scores = F.softmax(class_logits, -1)
pred_boxes_list = pred_boxes.split(boxes_per_image, 0)
pred_scores_list = pred_scores.split(boxes_per_image, 0)
all_boxes = []
all_scores = []
all_labels = []
for boxes, scores, image_shape in zip(pred_boxes_list,
pred_scores_list, image_shapes):
boxes = box_ops.clip_boxes_to_image(boxes, image_shape)
# create labels for each prediction
labels = torch.arange(num_classes, device=device)
labels = labels.view(1, -1).expand_as(scores)
# remove predictions with the background label
boxes = boxes[:, 1:]
scores = scores[:, 1:]
labels = labels[:, 1:]
# batch everything, by making every class prediction be a separate instance
boxes = boxes.reshape(-1, 4)
scores = scores.reshape(-1)
labels = labels.reshape(-1)
# remove low scoring boxes
inds = torch.nonzero(scores > cfg.BOX.SCORE_THRESH).squeeze(1)
boxes, scores, labels = boxes[inds], scores[inds], labels[inds]
# remove empty boxes
keep = box_ops.remove_small_boxes(boxes, min_size=1e-2)
boxes, scores, labels = boxes[keep], scores[keep], labels[keep]
# non-maximum suppression, independently done per class
keep = box_ops.batched_nms(boxes, scores, labels,
cfg.BOX.NMS_THRESH)
# keep only topk scoring predictions
keep = keep[:cfg.BOX.DETECTIONS_PER_IMG]
boxes, scores, labels = boxes[keep], scores[keep], labels[keep]
all_boxes.append(boxes)
all_scores.append(scores)
all_labels.append(labels)
return all_boxes, all_scores, all_labels
def generate_proposals(self,
anchors,
objectness,
box_regression,
img_metas,
is_target_domain=False):
"""
Args:
anchors:
objectness: (N, A, H, W)
box_regression: (N, A * 4, H, W)
img_metas:
is_target_domain:
Returns:
"""
pre_nms_top_n = self.pre_nms_top_n[self.training]
post_nms_top_n = self.post_nms_top_n[self.training]
if is_target_domain:
post_nms_top_n = self.cfg.MODEL.ROI_HEADS.BATCH_SIZE_PER_IMAGE
nms_thresh = self.nms_thresh
device = objectness.device
N, A, H, W = objectness.shape
objectness = objectness.permute(0, 2, 3, 1).reshape(N, H * W * A)
objectness = objectness.sigmoid()
box_regression = box_regression.permute(0, 2, 3,
1).reshape(N, H * W * A, 4)
concat_anchors = cat(anchors, dim=0)
concat_anchors = concat_anchors.reshape(N, A * H * W, 4)
num_anchors = A * H * W
pre_nms_top_n = min(pre_nms_top_n, num_anchors)
objectness, topk_idx = objectness.topk(pre_nms_top_n,
dim=1,
sorted=True)
batch_idx = torch.arange(N, device=device)[:, None]
box_regression = box_regression[batch_idx, topk_idx]
concat_anchors = concat_anchors[batch_idx, topk_idx]
proposals = self.box_coder.decode(box_regression.view(-1, 4),
concat_anchors.view(-1, 4))
proposals = proposals.view(N, -1, 4)
results = []
for proposal, score, img_meta in zip(proposals, objectness, img_metas):
img_width, img_height = img_meta['img_shape']
proposal = box_ops.clip_boxes_to_image(proposal,
(img_height, img_width))
keep = box_ops.remove_small_boxes(proposal, self.min_size)
proposal = proposal[keep]
score = score[keep]
keep = ops.nms(proposal, score, nms_thresh)
keep = keep[:post_nms_top_n]
proposal = proposal[keep]
score = score[keep]
results.append(proposal) # (N, 4)
return results
def postprocess_detections(self, class_logits, box_regression, anchors,
image_shapes):
# type: (Tensor, Tensor, List[Tensor], List[Tuple[int, int]])
device = class_logits.device
num_classes = class_logits.shape[-1]
#$
boxes_per_image = [len(boxes_in_image) for boxes_in_image in anchors]
pred_boxes = self.box_coder.decode(box_regression, anchors)
pred_scores = F.softmax(class_logits, -1)
# split boxes and scores per image
if len(boxes_per_image) == 1:
# TODO : remove this when ONNX support dynamic split sizes
# and just assign to pred_boxes instead of pred_boxes_list
pred_boxes_list = [pred_boxes]
pred_scores_list = [pred_scores]
else:
pred_boxes_list = pred_boxes.split(boxes_per_image, 0)
pred_scores_list = pred_scores.split(boxes_per_image, 0)
all_boxes = []
all_scores = []
all_labels = []
for boxes, scores, image_shape in zip(pred_boxes_list,
pred_scores_list, image_shapes):
boxes = box_ops.clip_boxes_to_image(boxes, image_shape)
# create labels for each prediction
labels = torch.arange(num_classes, device=device)
row_to, collumn = 1, -1
labels = labels.view(row_to, collumn).expand_as(scores)
along_class_prediction = 1
pred_class_labels = torch.argmax(scores,
dim=along_class_prediction)
prediction_num = torch.arange(boxes.shape[0], device=device)
flat_pred_class_prob_idxs = prediction_num * num_classes + pred_class_labels
top1_scores = torch.take(scores, flat_pred_class_prob_idxs)
labels = pred_class_labels
# remove predictions with the background label
#boxes = boxes[:, 1:]
#scores = scores[:, 1:]
#labels = labels[:, 1:]
# batch everything, by making every class prediction be a separate instance
#boxes = boxes.reshape(-1, 4)
#scores = scores.reshape(-1)
#labels = labels.reshape(-1)
# remove low scoring boxes
inds = torch.nonzero(top1_scores > self.score_thresh).squeeze(1)
boxes, scores, labels = boxes[inds].squeeze(
1), top1_scores[inds], labels[inds]
# remove empty boxes
keep = box_ops.remove_small_boxes(boxes, min_size=1e-2)
boxes, scores, labels = boxes[keep], scores[keep], labels[keep]
# non-maximum suppression, independently done per class
keep = box_ops.batched_nms(boxes, scores, labels, self.nms_thresh)
# keep only topk scoring predictions
keep = keep[:self.detections_per_img]
boxes, scores, labels = boxes[keep], scores[keep], labels[keep]
all_boxes.append(boxes)
all_scores.append(scores)
all_labels.append(labels)
return all_boxes, all_scores, all_labels
def postprocess_detections(self, head_outputs, anchors, image_shapes):
# type: (Dict[str, Tensor], List[Tensor], List[Tuple[int, int]]) -> List[Dict[str, Tensor]]
# TODO: Merge this with roi_heads.RoIHeads.postprocess_detections ?
class_logits = head_outputs.pop('cls_logits')
box_regression = head_outputs.pop('bbox_regression')
other_outputs = head_outputs
device = class_logits.device
num_classes = class_logits.shape[-1]
scores = torch.sigmoid(class_logits)
# create labels for each score
labels = torch.arange(num_classes, device=device)
labels = labels.view(1, -1).expand_as(scores)
detections = torch.jit.annotate(List[Dict[str, Tensor]], [])
for index, (box_regression_per_image, scores_per_image, labels_per_image, anchors_per_image, image_shape) in \
enumerate(zip(box_regression, scores, labels, anchors, image_shapes)):
boxes_per_image = self.box_coder.decode_single(
box_regression_per_image, anchors_per_image)
boxes_per_image = box_ops.clip_boxes_to_image(
boxes_per_image, image_shape)
other_outputs_per_image = [(k, v[index])
for k, v in other_outputs.items()]
image_boxes = []
image_scores = []
image_labels = []
image_other_outputs = torch.jit.annotate(Dict[str, List[Tensor]],
{})
for class_index in range(num_classes):
# remove low scoring boxes
inds = torch.gt(scores_per_image[:, class_index],
self.score_thresh)
boxes_per_class, scores_per_class, labels_per_class = \
boxes_per_image[inds], scores_per_image[inds, class_index], labels_per_image[inds, class_index]
other_outputs_per_class = [(k, v[inds])
for k, v in other_outputs_per_image]
# remove empty boxes
keep = box_ops.remove_small_boxes(boxes_per_class,
min_size=1e-2)
boxes_per_class, scores_per_class, labels_per_class = \
boxes_per_class[keep], scores_per_class[keep], labels_per_class[keep]
other_outputs_per_class = [(k, v[keep])
for k, v in other_outputs_per_class]
# non-maximum suppression, independently done per class
keep = box_ops.nms(boxes_per_class, scores_per_class,
self.nms_thresh)
# keep only topk scoring predictions
keep = keep[:self.detections_per_img]
boxes_per_class, scores_per_class, labels_per_class = \
boxes_per_class[keep], scores_per_class[keep], labels_per_class[keep]
other_outputs_per_class = [(k, v[keep])
for k, v in other_outputs_per_class]
image_boxes.append(boxes_per_class)
image_scores.append(scores_per_class)
image_labels.append(labels_per_class)
for k, v in other_outputs_per_class:
if k not in image_other_outputs:
image_other_outputs[k] = []
image_other_outputs[k].append(v)
detections.append({
'boxes': torch.cat(image_boxes, dim=0),
'scores': torch.cat(image_scores, dim=0),
'labels': torch.cat(image_labels, dim=0),
})
for k, v in image_other_outputs.items():
detections[-1].update({k: torch.cat(v, dim=0)})
return detections
def postprocess_detections(
self,
class_logits, # type: Tensor
dof_regression, # type: Tensor
proposals, # type: List[Tensor]
image_shapes, # type: List[Tuple[int, int]]
):
# type: (...) -> Tuple[List[Tensor], List[Tensor], List[Tensor]]
device = class_logits.device
num_classes = class_logits.shape[-1]
boxes_per_image = [
boxes_in_image.shape[0] for boxes_in_image in proposals
]
pred_boxes = torch.cat(proposals, dim=0)
N = dof_regression.shape[0]
pred_boxes = pred_boxes.reshape(N, -1, 4)
pred_dofs = dof_regression.reshape(N, -1, 6)
pred_scores = F.softmax(class_logits, -1)
pred_boxes_list = pred_boxes.split(boxes_per_image, 0)
pred_scores_list = pred_scores.split(boxes_per_image, 0)
pred_dofs_list = pred_dofs.split(boxes_per_image, 0)
all_boxes = []
all_scores = []
all_labels = []
all_dofs = []
for boxes, dofs, scores, image_shape in zip(pred_boxes_list,
pred_dofs_list,
pred_scores_list,
image_shapes):
boxes = box_ops.clip_boxes_to_image(boxes, image_shape)
# create labels for each prediction
labels = torch.arange(num_classes, device=device)
labels = labels.view(1, -1).expand_as(scores)
# remove predictions with the background label
dofs = dofs[:, 1:]
scores = scores[:, 1:]
labels = labels[:, 1:]
# batch everything, by making every class prediction be a separate instance
boxes = boxes.reshape(-1, 4)
dofs = dofs.reshape(-1, 6)
scores = scores.reshape(-1)
labels = labels.reshape(-1)
# remove low scoring boxes
inds = torch.nonzero(scores > self.score_thresh).squeeze(1)
boxes, dofs, scores, labels = (
boxes[inds],
dofs[inds],
scores[inds],
labels[inds],
)
# remove empty boxes
keep = box_ops.remove_small_boxes(boxes, min_size=1e-2)
boxes, dofs, scores, labels = (
boxes[keep],
dofs[keep],
scores[keep],
labels[keep],
)
# create boxes from the predicted poses
boxes, dofs = transform_pose_global_project_bbox(
boxes,
dofs,
self.pose_mean,
self.pose_stddev,
image_shape,
self.threed_68_points,
bbox_x_factor=self.bbox_x_factor,
bbox_y_factor=self.bbox_y_factor,
expand_forehead=self.expand_forehead,
)
# non-maximum suppression, independently done per class
keep = box_ops.batched_nms(boxes, scores, labels, self.nms_thresh)
boxes, dofs, scores, labels = (
boxes[keep],
dofs[keep],
scores[keep],
labels[keep],
)
# keep only topk scoring predictions
keep = keep[:self.detections_per_img]
all_boxes.append(boxes)
all_scores.append(scores)
all_labels.append(labels)
all_dofs.append(dofs)
return all_boxes, all_dofs, all_scores, all_labels
def process_detections(
self,
outputs: Dict[str, Tensor],
anchors: List[Tensor],
im_szs: List[Tuple[int, int]],
) -> Tuple[List[Tensor], List[Tensor], List[Tensor]]:
" Process `outputs` and return the predicted bboxes, score, clas_labels above `detect_thres` "
class_logits = outputs.pop("cls_preds")
bboxes = outputs.pop("bbox_preds")
scores = torch.sigmoid(class_logits)
device = class_logits.device
num_classes = class_logits.shape[-1]
# create labels for each score
labels = torch.arange(num_classes, device=device)
labels = labels.view(1, -1).expand_as(scores)
detections = torch.jit.annotate(List[Dict[str, Tensor]], [])
for bb_per_im, sc_per_im, ancs_per_im, im_sz, lbl_per_im in zip(bboxes, scores, anchors, im_szs, labels):
all_boxes = []
all_scores = []
all_labels = []
# convert the activation i.e, outputs of the model to bounding boxes
bb_per_im = activ_2_bbox(bb_per_im, ancs_per_im)
# clip the bounding boxes to the image size
bb_per_im = ops.clip_boxes_to_image(bb_per_im, im_sz)
# Iterate over each `cls_idx` in `num_classes` and do nms
# to each class individually
for cls_idx in range(num_classes):
# remove low predicitons with scores < score_thres
# and grab the predictions corresponding to the cls_idx
inds = torch.gt(sc_per_im[:, cls_idx], self.score_thres)
bb_per_cls, sc_per_cls, lbl_per_cls = (
bb_per_im[inds],
sc_per_im[inds, cls_idx],
lbl_per_im[inds, cls_idx],
)
# remove boxes that are too small ~(1-02)
keep = ops.remove_small_boxes(bb_per_cls, min_size=1e-2)
bb_per_cls, sc_per_cls, lbl_per_cls = (
bb_per_cls[keep],
sc_per_cls[keep],
lbl_per_cls[keep],
)
# compute non max supression to supress overlapping boxes
keep = ops.nms(bb_per_cls, sc_per_cls, self.nms_thres)
bb_per_cls, sc_per_cls, lbl_per_cls = (
bb_per_cls[keep],
sc_per_cls[keep],
lbl_per_cls[keep],
)
all_boxes.append(bb_per_cls)
all_scores.append(sc_per_cls)
all_labels.append(lbl_per_cls)
# Convert to tensors
all_boxes = torch.cat(all_boxes, dim=0)
all_scores = torch.cat(all_scores, dim=0)
all_labels = torch.cat(all_labels, dim=0)
# model is going to predict classes which are going to be in the range of [0, num_classes]
# 0 is reserved for the background class for which no loss is calculate , so
# we will add 1 to all the class_predictions to shift the predicitons range from
# [0, num_classes) -> [1, num_classes]
all_labels = all_labels + 1
# Sort by scores and
# Grab the idxs from the corresponding to the topk predictions
_, topk_idxs = all_scores.sort(descending=True)
topk_idxs = topk_idxs[: self.detections_per_img]
all_boxes, all_scores, all_labels = (
all_boxes[topk_idxs],
all_scores[topk_idxs],
all_labels[topk_idxs],
)
detections.append({"boxes": all_boxes, "scores": all_scores, "labels": all_labels,})
return detections
def post_process(self, cls_logits: torch.Tensor, reg_deltas: torch.Tensor,
batched_rois: List[torch.Tensor]):
nms_threshold = self._params['nms_threshold']
conf_threshold = self._params['conf_threshold']
keep_top_n = self._params['keep_top_n']
batched_dets: List[torch.Tensor] = []
current = 0
for rois in batched_rois:
N = rois.size(0)
if N == 0:
print("warning! found empty rois")
batched_dets.append(
torch.empty(0,
6,
dtype=reg_deltas.dtype,
device=reg_deltas.device))
continue
logits = cls_logits[current:current + N]
offsets = reg_deltas[current:current + N]
current += N
# logits: torch.Tensor(N,)
# deltas: torch.Tensor(N,4)
# rois: torch.Tensor(N,4)
scores = torch.sigmoid(logits)
preds = torch.zeros(scores.shape,
dtype=torch.int64,
device=scores.device)
preds[scores >= 0.5] = 1
fg_preds_mask = preds != 0
# convert offsets to boxes
# N,4 | N,4 => N,4 as xmin,ymin,xmax,ymax
boxes = offsets2boxes(offsets.unsqueeze(0), rois).squeeze(0)
# extract bg predictions
boxes = boxes[fg_preds_mask]
preds = preds[fg_preds_mask]
scores = scores[fg_preds_mask]
# apply conf threshold
keep = scores >= conf_threshold
scores, preds, boxes = scores[keep], preds[keep], boxes[keep]
# remove small
keep = box_ops.remove_small_boxes(boxes, 1e-3) # TODO try 1
scores, preds, boxes = scores[keep], preds[keep], boxes[keep]
# batched nms
keep = box_ops.batched_nms(boxes, scores, preds, nms_threshold)
scores, preds, boxes = scores[keep], preds[keep], boxes[keep]
# select top n
keep_n = min(keep_top_n, scores.size(0))
_, selected_ids = scores.topk(keep_n)
scores, preds, boxes = scores[selected_ids], preds[
selected_ids], boxes[selected_ids]
scores.unsqueeze_(1)
preds = preds.unsqueeze(1).to(boxes.dtype)
dets = torch.cat([boxes, scores, preds], dim=-1)
batched_dets.append(dets)
return batched_dets
def _postprocess_detections(
self,
class_logits: Tensor,
box_features: Tensor,
box_regression: Tensor,
proposals: List[Tensor],
image_shapes: List[Tuple[int, int]],
) -> Tuple[List[Tensor], List[Tensor], List[Tensor], List[Tensor]]:
"""
Adapted from https://github.com/pytorch/vision/blob/
4521f6d152875974e317fa247a633e9ad1ea05c8/torchvision/models/detection/roi_heads.py#L664.
The only reason we have to re-implement this method is so we can pull out the box
features that we want.
"""
device = class_logits.device
num_classes = class_logits.shape[-1]
boxes_per_image = [boxes_in_image.shape[0] for boxes_in_image in proposals]
# shape: (batch_size * boxes_per_image, num_classes, 4)
pred_boxes = self.detector.roi_heads.box_coder.decode(box_regression, proposals)
pred_scores = F.softmax(class_logits, -1)
pred_boxes_list = pred_boxes.split(boxes_per_image, 0)
features_list = box_features.split(boxes_per_image, dim=0)
pred_scores_list = pred_scores.split(boxes_per_image, 0)
all_boxes = []
all_features = []
all_scores = []
all_labels = []
for boxes, features, scores, image_shape in zip(
pred_boxes_list, features_list, pred_scores_list, image_shapes
):
# shape: (boxes_per_image, num_classes, 4)
boxes = box_ops.clip_boxes_to_image(boxes, image_shape)
# shape: (boxes_per_image, num_classes, feature_size)
features = features.unsqueeze(1).expand(boxes.shape[0], boxes.shape[1], -1)
# create labels for each prediction
# shape: (num_classes,)
labels = torch.arange(num_classes, device=device)
# shape: (boxes_per_image, num_classes,)
labels = labels.view(1, -1).expand_as(scores)
# remove predictions with the background label
# shape: (boxes_per_image, num_classes - 1, 4)
boxes = boxes[:, 1:]
# shape: (boxes_per_image, num_classes, feature_size)
features = features[:, 1:]
# shape: (boxes_per_image, num_classes - 1,)
scores = scores[:, 1:]
# shape: (boxes_per_image, num_classes - 1,)
labels = labels[:, 1:]
# batch everything, by making every class prediction be a separate instance
# shape: (boxes_per_image * (num_classes - 1), 4)
boxes = boxes.reshape(-1, 4)
# shape: (boxes_per_image * (num_classes - 1), feature_size)
features = features.reshape(boxes.shape[0], -1)
# shape: (boxes_per_image * (num_classes - 1),)
scores = scores.reshape(-1)
# shape: (boxes_per_image * (num_classes - 1),)
labels = labels.reshape(-1)
# remove low scoring boxes
inds = torch.where(scores > self.detector.roi_heads.score_thresh)[0]
boxes, features, scores, labels = (
boxes[inds],
features[inds],
scores[inds],
labels[inds],
)
# remove empty boxes
keep = box_ops.remove_small_boxes(boxes, min_size=1e-2)
boxes, features, scores, labels = (
boxes[keep],
features[keep],
scores[keep],
labels[keep],
)
# non-maximum suppression, independently done per class
keep = box_ops.batched_nms(boxes, scores, labels, self.detector.roi_heads.nms_thresh)
# keep only topk scoring predictions
keep = keep[: self.detector.roi_heads.detections_per_img]
boxes, features, scores, labels = (
boxes[keep],
features[keep],
scores[keep],
labels[keep],
)
all_boxes.append(boxes)
all_features.append(features)
all_scores.append(scores)
all_labels.append(labels)
return all_boxes, all_features, all_scores, all_labels
