def forward(self, images, targets=None):
"""
Arguments:
images (list[Tensor] or ImageList): images to be processed
targets (list[BoxList]): ground-truth boxes present in the image (optional)
Returns:
result (list[BoxList] or dict[Tensor]): the output from the model.
During training, it returns a dict[Tensor] which contains the losses.
During testing, it returns list[BoxList] contains additional fields
like `scores`, `labels` and `mask` (for Mask R-CNN models).
"""
if self.training and targets is None:
raise ValueError("In training mode, targets should be passed")
images = to_image_list(images)
features = self.backbone(images.tensors)
proposals, proposal_losses = self.rpn(images, features, targets)
if self.roi_heads:
x, result, detector_losses = self.roi_heads(features, proposals, targets)
else:
# RPN-only models don't have roi_heads
x = features
result = proposals
detector_losses = {}
if self.training:
losses = {}
losses.update(detector_losses)
losses.update(proposal_losses)
return losses
return result
def get_detectron_features(self, image_paths):
img_tensor, im_scales = [], []
for image_path in image_paths:
im, im_scale = self._image_transform(image_path)
img_tensor.append(im)
im_scales.append(im_scale)
# Image dimensions should be divisible by 32, to allow convolutions
# in detector to work
current_img_list = to_image_list(img_tensor, size_divisible=32)
current_img_list = current_img_list.to("cuda")
with torch.no_grad():
output = self.detection_model(current_img_list)
feat_list = self._process_feature_extraction(
output, im_scales, self.args.feature_name,
self.args.confidence_threshold)
return feat_list
def compute_prediction(self, img, img_h, img_w):
"""
Arguments:
imgs, imgs_h, imgs_w (np.ndarray): imgs as returned by data_loader
Returns:
top_predictions ([BoxList]): the detected objects. Additional information
of the detection properties can be found in the fields of
the BoxList via `prediction.fields()`
"""
img_list = to_image_list(img, self.cfg.DATALOADER.SIZE_DIVISIBILITY)
img_list = img_list.to(self.device)
# compute predictions
with torch.no_grad():
predictions = self.model(img_list)
predictions = [o.to(self.cpu_device) for o in predictions]
prediction = predictions[0].resize((img_w, img_h))
top_prediction = self.select_top_predictions(prediction, img_w, img_h)
return top_prediction
def forward(self, images, targets=None, force_boxes=False):
"""
Arguments:
images (list[Tensor] or ImageList): images to be processed
targets (list[BoxList]): ground-truth boxes present in the image (optional)
Returns:
result (list[BoxList] or dict[Tensor]): the output from the model.
During training, it returns a dict[Tensor] which contains the losses.
During testing, it returns list[BoxList] contains additional fields
like `scores`, `labels` and `mask` (for Mask R-CNN models).
"""
if self.training and targets is None:
raise ValueError("In training mode, targets should be passed")
images = to_image_list(images)
features = self.backbone(images.tensors)
if (not self.training) and targets and force_boxes:
# if in the reference model, we want to force boxes given in the target
proposals = [
BoxList(target.bbox, target.size, target.mode)
for target in targets
]
proposal_losses = {}
else:
proposals, proposal_losses = self.rpn(images, features, targets)
if self.roi_heads:
x, result, detector_losses = self.roi_heads(
features, proposals, targets, force_boxes)
else:
# RPN-only models don't have roi_heads
x = features
result = proposals
detector_losses = {}
if self.training:
losses = {}
losses.update(detector_losses)
losses.update(proposal_losses)
return losses
return result
def __call__(self, batch):
transposed_batch = list(zip(*batch))
img_lists = list(zip(*(transposed_batch[0])))
# img_lists = transposed_batch[0]
target_lists = list(zip(*(transposed_batch[1])))
# target_lists = transposed_batch[1]
imgs = []
tgts = []
for img_l in img_lists:
imgs += img_l
# imgs.append(img_l[0])
for tgt in target_lists:
tgts += tgt
# tgts.append(tgt[0])
images = to_image_list(imgs, self.size_divisible)
targets = tgts
img_ids = transposed_batch[2]
return images, targets, img_ids
def forward(self, images, targets=None):
"""
Arguments:
images (list[Tensor] or ImageList): images to be processed
targets (list[BoxList]): ground-truth boxes present in the image (optional)
Returns:
result (list[BoxList] or dict[Tensor]): the output from the model.
During training, it returns a dict[Tensor] which contains the losses.
During testing, it returns list[BoxList] contains additional fields
like `scores`, `labels` and `mask` (for Mask R-CNN models).
"""
if self.training and targets is None:
raise ValueError("In training mode, targets should be passed")
images = to_image_list(images)
features = self.backbone(images.tensors)
proposals, proposal_losses = self.rpn(images, features, targets)
if self.roi_heads:
x, result, detector_losses = self.roi_heads(
features, proposals, targets)
else:
# RPN-only models don't have roi_heads
result = proposals
detector_losses = {}
if self.training:
losses = {}
# Prepend the key name so as to allow better visualisation in tensorboard
detector_losses_tb = {
'detector_losses/' + k: v
for k, v in detector_losses.items()
}
proposal_losses_tb = {
'proposal_losses/' + k: v
for k, v in proposal_losses.items()
}
losses.update(detector_losses_tb)
losses.update(proposal_losses_tb)
return losses
return result
def forward(self, images, targets=None):
"""
Arguments:
images (list[Tensor] or ImageList): images to be processed
targets (list[BoxList]): ground-truth boxes present in the image (optional)
Returns:
result (list[BoxList] or dict[Tensor]): the output from the model.
During training, it returns a dict[Tensor] which contains the losses.
During testing, it returns list[BoxList] contains additional fields
like `scores`, `labels` and `mask` (for Mask R-CNN models).
"""
if self.training and targets is None:
raise ValueError("In training mode, targets should be passed")
images = to_image_list(images)
features = self.backbone(images.tensors)
# conduct multi-label classification on the top of features
multilable_loss = {}
if self.training:
labels_list = [target.get_field("labels") for target in targets]
multilabel_loss = self.multilabel_cls(features[-1], labels_list)
proposals, proposal_losses = self.rpn(images, features, targets)
if self.roi_heads:
x, result, detector_losses = self.roi_heads(
features, proposals, targets)
else:
# RPN-only models don't have roi_heads
x = features
result = proposals
detector_losses = {}
if self.training:
losses = {}
losses.update(detector_losses)
losses.update(proposal_losses)
losses.update(multilabel_loss)
return losses
return result
def forward(self, images, targets=None, centerness_pack=None, iter=None):
"""
Arguments:
images (list[Tensor] or ImageList): images to be processed
targets (list[BoxList]): ground-truth boxes present in the image (optional)
Returns:
result (list[BoxList] or dict[Tensor]): the output from the model.
During training, it returns a dict[Tensor] which contains the losses.
During testing, it returns list[BoxList] contains additional fields
like `scores`, `labels` and `mask` (for Mask R-CNN models).
"""
if self.training and targets is None:
raise ValueError("In training mode, targets should be passed")
images = to_image_list(images)
features = self.backbone(images.tensors)
# print([image.shape for image in images.tensors])
# print([feature.shape for feature in features])
# print(targets)
proposals, proposal_losses, pred_targets = self.rpn(
images, features[:-1], targets, features[-1], centerness_pack,
iter)
if self.roi_heads:
x, result, detector_losses = self.roi_heads(
features[:-1], proposals, targets, pred_targets)
else:
# RPN-only models don't have roi_heads
x = features
result = proposals
detector_losses = {}
if self.training:
losses = {}
losses.update(detector_losses)
losses.update(proposal_losses)
if iter <= self.warmup:
for key, loss in losses.items():
if key != 'loss_centerness' and key != 'loss_rpn_center_box_reg':
loss *= 0.0
return losses
return result
def forward(self, images, targets=None):
"""
Arguments:
images (list[Tensor] or ImageList): images to be processed
targets (list[BoxList]): ground-truth boxes present in the image (optional)
Returns:
result (list[BoxList] or dict[Tensor]): the output from the model.
During training, it returns a dict[Tensor] which contains the losses.
During testing, it returns list[BoxList] contains additional fields
like `scores`, `labels` and `mask` (for Mask R-CNN models).
"""
if self.training and targets is None:
raise ValueError("In training mode, targets should be passed")
images = to_image_list(images)
features = self.backbone(images.tensors)
if self.fp4p_on:
# get you C4
proposals, proposal_losses = self.rpn(images, (features[-1], ),
targets)
else:
proposals, proposal_losses = self.rpn(images, features, targets)
# features = [feature.detach() for feature in features]
if self.roi_heads:
x, result, detector_losses = self.roi_heads(
features, proposals, targets)
else:
# RPN-only models don't have roi_heads
x = features
result = proposals
detector_losses = {}
if self.training:
losses = {}
losses.update(detector_losses)
losses.update(proposal_losses)
return losses
return result
def compute_prediction(self, original_image):
"""
Arguments:
original_image (np.ndarray): an image as returned by OpenCV
Returns:
prediction (BoxList): the detected objects. Additional information
of the detection properties can be found in the fields of
the BoxList via `prediction.fields()`
"""
# apply pre-processing to image
#print(len(original_image))
image = [self.transforms(original_image[i]) for i in range(len(original_image))]
'''
with open('test.txt','a') as f:
np.set_printoptions(threshold='nan')
i = torch.LongTensor([[1,1],[1,2],[2,1],[2,2]])
print(image[0].shape)
a = image[0][0]
b = image[0][1]
c = image[0][2]
f.write(str(a))
f.write(str(b))
f.write(str(c))
sys.exit(0)
'''
# convert to an ImageList, padded so that it is divisible by
# cfg.DATALOADER.SIZE_DIVISIBILITY
image_list = to_image_list(image, self.cfg.DATALOADER.SIZE_DIVISIBILITY)
#print(image_list)
image_list = image_list.to(self.device)
# compute predictions
predictions = self.model(image_list)
predictions = [o.to(self.cpu_device) for o in predictions]
#print("predictions: {}".format(predictions))
# always single image is passed at a time
#prediction = predictions[0]
# reshape prediction (a BoxList) into the original image size
height = width = original_image[0].shape[-2]
prediction = [predictions[i].resize((width, height)) for i in range(len(predictions))]
return prediction
def compute_prediction(self, original_image):
"""
Arguments:
original_image (np.ndarray): an image as returned by OpenCV
Returns:
prediction (BoxList): the detected objects. Additional information
of the detection properties can be found in the fields of
the BoxList via `prediction.fields()`
"""
self.pool_feature = []
self.global_feature = []
def hook(module, input, output):
self.global_feature = module.backbone.fpn.global_feature
# apply pre-processing to image
image = self.transforms(original_image)
# convert to an ImageList, padded so that it is divisible by
# cfg.DATALOADER.SIZE_DIVISIBILITY
image_list = to_image_list(image, self.cfg.DATALOADER.SIZE_DIVISIBILITY)
image_list = image_list.to(self.device)
# compute predictions
with torch.no_grad():
handle = self.model.register_forward_hook(hook)
predictions = self.model(image_list)
print(self.global_feature.shape)
predictions = [o.to(self.cpu_device) for o in predictions]
# always single image is passed at a time
prediction = predictions[0]
# reshape prediction (a BoxList) into the original image size
height, width = original_image.shape[:-1]
prediction = prediction.resize((width, height))
if prediction.has_field("mask"):
# if we have masks, paste the masks in the right position
# in the image, as defined by the bounding boxes
masks = prediction.get_field("mask")
# always single image is passed at a time
masks = self.masker([masks], [prediction])[0]
prediction.add_field("mask", masks)
return prediction
def compute_prediction(self, original_image):
"""
Arguments:
original_image (np.ndarray): an image as returned by OpenCV
Returns:
prediction (BoxList): the detected objects. Additional information
of the detection properties can be found in the fields of
the BoxList via `prediction.fields()`
"""
# apply pre-processing to image
image = self.transforms(original_image)
# convert to an ImageList, padded so that it is divisible by
# cfg.DATALOADER.SIZE_DIVISIBILITY
image_list = to_image_list(image, self.cfg.DATALOADER.SIZE_DIVISIBILITY)
image_list = image_list.to(self.device)
# compute predictions
height, width = original_image.shape[:-1]
with torch.no_grad():
predictions = self.model(image_list)
predictions = [o.to(self.cpu_device) for o in predictions]
# always single image is passed at a time
prediction = predictions[0]
# reshape prediction (a BoxList) into the original image size
prediction = prediction.resize((width, height))
if prediction.has_field("mask"):
# if we have masks, paste the masks in the right position
# in the image, as defined by the bounding boxes
masks = prediction.get_field("mask")
# always single image is passed at a time
masks = self.masker([masks], [prediction])[0]
prediction.add_field("mask", masks)
if prediction.has_field("ke") and self.cfg.MODEL.KE_ON:
kes = prediction.get_field("ke")
mty = prediction.get_field("mty")
prediction.add_field("ke", kes.kes)
prediction.add_field("mty", mty)
return prediction
def forward(self, images, targets=None, logits_only=False, adv_patch=None):
"""
Arguments:
images (list[Tensor] or ImageList): images to be processed
targets (list[BoxList]): ground-truth boxes present in the image (optional)
Returns:
result (list[BoxList] or dict[Tensor]): the output from the model.
During training, it returns a dict[Tensor] which contains the losses.
During testing, it returns list[BoxList] contains additional fields
like `scores`, `labels` and `mask` (for Mask R-CNN models).
"""
if self.training and targets is None:
raise ValueError("In training mode, targets should be passed")
images = to_image_list(images)
features = self.backbone(images.tensors)
proposals, proposal_losses = self.rpn(images, features, targets,
logits_only)
if logits_only:
if self.training:
tv_losses = {"loss_tv": 2.5 * SmoothTV(adv_patch)}
losses = {}
losses.update(proposal_losses)
losses.update(tv_losses)
return losses
if self.roi_heads:
x, result, detector_losses = self.roi_heads(
features, proposals, targets)
else:
# RPN-only models don't have roi_heads
x = features
result = proposals
detector_losses = {}
if self.training:
losses = {}
losses.update(detector_losses)
losses.update(proposal_losses)
return losses
return result
def compute_prediction(self, original_image):
"""
Arguments:
original_image (np.ndarray): an image as returned by OpenCV
Returns:
prediction (BoxList): the detected objects. Additional information
of the detection properties can be found in the fields of
the BoxList via `prediction.fields()`
"""
# apply pre-processing to image
image = self.transforms(original_image)
# convert to an ImageList, padded so that it is divisible by
# cfg.DATALOADER.SIZE_DIVISIBILITY
image_list = to_image_list(image, self.cfg.DATALOADER.SIZE_DIVISIBILITY)
image_list = image_list.to(self.device)
# compute predictions
with torch.no_grad():
predictions = self.model(image_list)
predictions = [o.to(self.cpu_device) for o in predictions]
# always single image is passed at a time
prediction = predictions[0]
# reshape prediction (a BoxList) into the original image size
height, width = original_image.shape[:-1]
prediction = prediction.resize((width, height))
# print('from compute_prediction:', prediction.bbox,prediction.get_field("scores"))
if prediction.has_field("mask"):
# if we have masks, paste the masks in the right position
# in the image, as defined by the bounding boxes
masks = prediction.get_field("mask")
# always single image is passed at a time
masks, res_polys, res_maskscore = self.masker([masks], [prediction])
# print("masks:", len(masks), masks[0][0][0].shape, res_polys[0][0], prediction.bbox[0].data.cpu().numpy())
# cv2.imshow("win", masks[0][0][0].data.cpu().numpy().astype(np.uint8))
# cv2.waitKey(0)
prediction.add_field("mask", masks[0])
prediction.add_field("mask_score", res_maskscore[0])
return prediction
def forward(self, images, targets=None, features=None):
"""
Arguments:
images (list[Tensor] or ImageList): images to be processed
targets (list[BoxList]): ground-truth boxes present in the image (optional)
features (list[Tensor]): encoder output features (optional)
Returns:
result (list[BoxList] or dict[Tensor]): the output from the model.
During training, it returns a dict[Tensor] which contains the losses.
During testing, it returns list[BoxList] contains additional fields
like `scores`, `labels` and `mask` (for Mask R-CNN models).
"""
if self.training and targets is None:
raise ValueError("In training mode, targets should be passed")
if features is None:
images = to_image_list(images)
features = self.encoder(images.tensors)
return self.decoder(images, features, targets)
def get_vinvl_features(self, image_paths):
img_tensor, im_infos = [], []
for image_path in image_paths:
im, im_info = self._image_transform(image_path)
img_tensor.append(im)
im_infos.append(im_info)
current_img_list = to_image_list(img_tensor, size_divisible=32)
current_img_list = current_img_list.to(DEVICE)
torch.manual_seed(0)
with torch.no_grad():
output = self.detection_model(current_img_list)
feat_list = self._process_feature_extraction(
output,
im_infos,
)
return feat_list
def get_detectron_features(self, image):
start = time.time()
image, scale = self.image_transform(image)
images = to_image_list([image], size_divisible=32)
images = images.to(self.device.type)
with torch.no_grad():
output = self.detectron_model(images)
features = self.feature_extract(output, [scale], 'fc6', 0.2)
end = time.time()
print(
f'Tiki : Getting Features : Detectron - Finished in {end-start:7.3f} Seconds'
)
processing['Detectron'] = end - start
return features[0]
def im_detect_bbox_hflip(model, images, target_scale, target_max_size, device):
"""
Performs bbox detection on the horizontally flipped image.
Function signature is the same as for im_detect_bbox.
"""
transform = TT.Compose([
T.Resize(target_scale, target_max_size),
TT.RandomHorizontalFlip(1.0),
TT.ToTensor(),
T.Normalize(mean=cfg.INPUT.PIXEL_MEAN,
std=cfg.INPUT.PIXEL_STD,
to_bgr255=cfg.INPUT.TO_BGR255)
])
images = [transform(image) for image in images]
images = to_image_list(images, cfg.DATALOADER.SIZE_DIVISIBILITY)
boxlists = model(images.to(device))
# Invert the detections computed on the flipped image
boxlists_inv = [boxlist.transpose(0) for boxlist in boxlists]
return boxlists_inv
def forward(self, images, targets=None, vis=False):
"""
Arguments:
images (list[Tensor] or ImageList): images to be processed
targets (list[BoxList]): ground-truth boxes present in the image (optional)
vis (bool): not used
Returns:
result (list[BoxList] or dict[Tensor]): the output from the model.
During training, it returns a dict[Tensor] which contains the losses.
During testing, it returns list[BoxList] contains additional fields
like `scores`, `labels` and `mask` (for Mask R-CNN models).
"""
if self.training and targets is None:
raise ValueError("In training mode, targets should be passed")
if self.training and self.has_aux_heads:
targets, targets_aux = targets
images = to_image_list(images)
features = self.neck(self.backbone(images.tensors))
proposals, proposal_losses = self.rpn(images,
features,
targets,
vis=vis)
if self.roi_heads:
x, result, detector_losses = self.roi_heads(
features, proposals, targets)
else:
# self.warm_start -= 1
# RPN-only models don't have roi_heads
result = proposals
detector_losses = {}
if self.training:
losses = {}
losses.update(detector_losses)
losses.update(proposal_losses)
return losses
return result
def forward(self, images, targets=None):
"""
Arguments:
images (list[Tensor] or ImageList): images to be processed
targets (list[BoxList]): ground-truth boxes present in the image (optional)
Returns:
result (list[BoxList] or dict[Tensor]): the output from the model.
During training, it returns a dict[Tensor] which contains the losses.
During testing, it returns list[BoxList] contains additional fields
like `scores`, `labels` and `mask` (for Mask R-CNN models).
"""
assert self.training is False
all_anchors, all_box_cls, all_box_regression = [], [], []
for resizer in self.resizers:
image_size = images.image_sizes[0]
size = resizer.get_size(image_size[::-1])
aug_images = interpolate(
images.tensors[:, :, :image_size[0], :image_size[1]],
size,
mode='bilinear',
align_corners=True)[0]
aug_images = to_image_list(
aug_images,
size_divisible=self.retinanet.cfg.DATALOADER.SIZE_DIVISIBILITY)
features = self.retinanet.backbone(aug_images.tensors)
if self.retinanet.cfg.RETINANET.BACKBONE == "p2p7":
features = features[1:]
box_cls, box_regression = self.retinanet.rpn.head(features)
anchors = self.retinanet.rpn.anchor_generator(
aug_images, features)[0]
all_anchors.extend(anchors), all_box_cls.extend(
box_cls), all_box_regression.extend(box_regression)
detections = self.retinanet.rpn.box_selector_test([all_anchors],
all_box_cls,
all_box_regression)
return detections
def forward(self, images, targets=None):
"""
Arguments:
images (list[Tensor] or ImageList): images to be processed
targets (list[BoxList]): ground-truth boxes present in the image (optional)
Returns:
result (list[BoxList] or dict[Tensor]): the output from the model.
During training, it returns a dict[Tensor] which contains the losses.
During testing, it returns list[BoxList] contains additional fields
like `scores`, `labels` and `mask` (for Mask R-CNN models).
"""
if self.training and targets is None:
raise ValueError("In training mode, targets should be passed")
images = to_image_list(images)
# R-50-FPN-RETINANET FPN
backbone_features = self.backbone(images.tensors)
# for f in backbone_features:
# print(f.shape)
# exit(0)
fpn_features = self.fpn(backbone_features)
proposals, proposal_losses = self.rpn(images, backbone_features[1:],
fpn_features, targets)
if self.roi_heads:
x, result, detector_losses = self.roi_heads(
features, proposals, targets)
else:
# RPN-only models don't have roi_heads
#FOCS 代替rpn
x = fpn_features
result = proposals
detector_losses = {}
if self.training:
losses = {}
losses.update(detector_losses)
losses.update(proposal_losses)
return losses
return result
def forward(self, images, targets=None, query=False):
"""
Arguments:
images (list[Tensor] or ImageList): images to be processed
targets (list[BoxList]): ground-truth boxes present in the image (optional)
Returns:
result (list[BoxList] or dict[Tensor]): the output from the model.
During training, it returns a dict[Tensor] which contains the losses.
During testing, it returns list[BoxList] contains additional fields
like `scores`, `labels` and `mask` (for Mask R-CNN models).
"""
if self.training and targets is None:
raise ValueError("In training mode, targets should be passed")
images = to_image_list(images)
features = self.backbone(images.tensors)
proposals, proposal_losses = self.rpn(images, features, targets)
if self.roi_heads:
if not self.training and query:
x, result, detector_losses = self.roi_heads(
features, targets, targets, query)
else:
x, result, detector_losses = self.roi_heads(
features, proposals, targets)
else:
# RPN-only models don't have roi_heads
x = features
result = proposals
detector_losses = {}
if self.training:
losses = {}
losses.update(detector_losses)
if not (self.cfg.MODEL.REID.TRAIN_PART
or self.cfg.MODEL.REID.TRAIN_PADREG):
losses.update(proposal_losses)
return losses
return result
def forward(self, images, targets=None):
# for target in targets:
# print(target.get_field('rotations'), '==========')
"""
Arguments:
images (list[Tensor] or ImageList): images to be processed
targets (list[BoxList]): ground-truth boxes present in the image (optional)
Returns:
result (list[BoxList] or dict[Tensor]): the output from the model.
During training, it returns a dict[Tensor] which contains the losses.
During testing, it returns list[BoxList] contains additional fields
like `scores`, `labels` and `mask` (for Mask R-CNN models).
"""
if self.training and targets is None:
raise ValueError("In training mode, targets should be passed")
imagen_tensor = images.tensors
images = to_image_list(images)
features = self.backbone(images.tensors)
# if not self.flag:
# self.save_feature_map(features,imagen_tensor) #=============================================== SAVING FEATURE MAPS
# print(images.tensors.size(),'=========================================')
proposals, proposal_losses = self.rpn(images, features, targets)
if self.roi_heads:
x, result, detector_losses = self.roi_heads(
features, proposals, targets)
else:
# RPN-only models don't have roi_heads
x = features
result = proposals
detector_losses = {}
if self.training:
losses = {}
losses.update(detector_losses)
losses.update(proposal_losses)
return losses
return result
def forward(self, images, targets=None):
# first forward pass of the Gnet
outputs = self.Gnet(images, targets)
if self.training:
losses = outputs
images = to_image_list(images)
# entering the discrimantor's loss
g_losses, d_losses = self.gan_loss_evaluator((torch.sigmoid(self.Gnet.roi_heads.mask.loss_evaluator.mask_dict['masks'].unsqueeze(1)), \
self.Gnet.roi_heads.mask.loss_evaluator.mask_dict['targets'].unsqueeze(1), \
self.Gnet.roi_heads.mask.loss_evaluator.positive_proposals, self.Gnet.features), \
(self.Gnet.roi_heads.box.bbox_dict['bbox'], self.Gnet.roi_heads.box.bbox_dict['target'], \
images.tensors))
losses.update(g_losses)
return losses, d_losses
return outputs
def get_detectron_features(image_paths, detection_model, get_boxes, feat_name,
device):
img_tensor, im_scales = [], []
for img_path in image_paths:
im = read_image(img_path)
im, im_scale = image_transform(im)
img_tensor.append(im)
im_scales.append(im_scale)
current_img_list = to_image_list(img_tensor, size_divisible=32)
current_img_list = current_img_list.to(device)
with torch.no_grad():
output = detection_model(current_img_list)
feat_list = process_feature_extraction(output,
im_scales,
get_boxes=get_boxes,
feat_name=feat_name,
conf_thresh=0.2)
return feat_list
def forward(self, images, targets=None, logger=None, ret_sg=None):
"""
Arguments:
images (list[Tensor] or ImageList): images to be processed
targets (list[BoxList]): ground-truth boxes present in the image (optional)
Returns:
result (list[BoxList] or dict[Tensor]): the output from the model.
During training, it returns a dict[Tensor] which contains the losses.
During testing, it returns list[BoxList] contains additional fields
like `scores`, `labels` and `mask` (for Mask R-CNN models).
"""
if self.training and targets is None:
raise ValueError("In training mode, targets should be passed")
images = to_image_list(images)
features = self.backbone(images.tensors)
proposals, proposal_losses = self.rpn(images, features, targets)
if self.roi_heads:
x, result, detector_losses, sg = self.roi_heads(
features, proposals, targets, logger)
else:
# RPN-only models don't have roi_heads
x = features
result = proposals
detector_losses = {}
if self.training:
losses = {}
losses.update(detector_losses)
if not self.cfg.MODEL.RELATION_ON:
# During the relationship training stage, the rpn_head should be fixed, and no loss.
losses.update(proposal_losses)
return losses, sg
if ret_sg:
return sg
return result
def compute_prediction(self, original_image, id):
"""
Arguments:
original_image (np.ndarray): an image as returned by OpenCV
Returns:
prediction (BoxList): the detected objects. Additional information
of the detection properties can be found in the fields of
the BoxList via `prediction.fields()`
"""
# apply pre-processing to image
image = self.transforms(original_image)
# depth1 = np.load("/home/zoey/ssds/data/Kitchen/simulator/rgbd/ADE_val_depth{0}.npy".format(id))
# imgWidth, imgHeight = 640, 480
# fov, aspect, nearplane, farplane = 45, imgWidth / imgHeight, 0.01, 100
# depth = farplane * nearplane / (farplane - (farplane - nearplane) * depth1)
# image = torch.cat((F.to_tensor(depth), image * 5))
# convert to an ImageList, padded so that it is divisible by
# cfg.DATALOADER.SIZE_DIVISIBILITY
image_list = to_image_list(image,
self.cfg.DATALOADER.SIZE_DIVISIBILITY)
image_list = image_list.to(self.device)
# compute predictions
with torch.no_grad():
predictions = self.model(image_list)
predictions = [o.to(self.cpu_device) for o in predictions]
# always single image is passed at a time
prediction = predictions[0]
# reshape prediction (a BoxList) into the original image size
height, width = original_image.shape[:-1]
prediction = prediction.resize((width, height))
if prediction.has_field("mask"):
# if we have masks, paste the masks in the right position
# in the image, as defined by the bounding boxes
masks = prediction.get_field("mask")
# always single image is passed at a time
masks = self.masker([masks], [prediction])[0]
prediction.add_field("mask", masks)
return prediction
def forward(self, images, targets=None, visualizer=None):
"""
Arguments:
images (list[Tensor] or ImageList): images to be processed
targets (list[BoxList]): ground-truth boxes present in the image (optional)
visualizer： for visualization
Returns:
result (list[BoxList] or dict[Tensor]): the output from the model.
During training, it returns a dict[Tensor] which contains the losses.
During testing, it returns list[BoxList] contains additional fields
like `scores`, `labels` and `mask` (for Mask R-CNN models).
"""
if self.training and targets is None:
raise ValueError("In training mode, targets should be passed")
images = to_image_list(images)
features = self.backbone(images.tensors)
proposals, proposal_losses = self.rpn(images, features, targets)
if self.roi_heads:
x, result, detector_losses = self.roi_heads(features, proposals, targets)
else:
# RPN-only models don't have roi_heads
x = features
result = proposals
detector_losses = {}
if not self.training:
# convert image format to opencv format
image = visualizer.transformBackForShowing(images.tensors[0].to("cpu"), None)[0].numpy().astype(np.uint8)
image = np.transpose(image, (1, 2, 0))
visualizer.run_on_opencv_image_prediction(image, result)
if self.training:
losses = {}
losses.update(detector_losses)
losses.update(proposal_losses)
return losses
return result
def prepare_images(self, inputCOCO_Image_maskrcnnTransform_list):
# Transform so that the min size is no smaller than cfg.INPUT.MIN_SIZE_TRAIN, and the max size is no larger than cfg.INPUT.MIN_SIZE_TRAIN
# image_batch = [self.transforms(original_image) for original_image in original_image_batch_list]
image_batch = inputCOCO_Image_maskrcnnTransform_list
image_sizes_after_transform = [(image_after.shape[2],
image_after.shape[1])
for image_after in image_batch]
# if self.training:
# for original_image, image_after, image_after_size in zip(inputCOCO_Image_maskrcnnTransform, image_batch, image_sizes_after_transform):
# self.printer.print('[generalized_rcnn_rui-prepare_images] Image sizes:', original_image.shape, '-->', image_after.shape, image_after_size)
# [Rui] PADDING
# convert to an ImageList, ``padded`` so that it is divisible by cfg.DATALOADER.SIZE_DIVISIBILITY
image_list = to_image_list(image_batch,
self.cfg.DATALOADER.SIZE_DIVISIBILITY)
# print(self.cfg.INPUT.MIN_SIZE_TRAIN, self.cfg.INPUT.MAX_SIZE_TRAIN, self.cfg.INPUT.MIN_SIZE_TEST, self.cfg.INPUT.MAX_SIZE_TEST)
if self.training:
self.printer.print(
'PADDED: image_list.tensors, image_list.image_sizes (before pad):',
image_list.tensors.shape, image_list.image_sizes)
image_list = image_list.to(self.device)
return image_list, image_sizes_after_transform
def forward(self, images, targets=None):
"""
Arguments:
images (list[Tensor] or ImageList): images to be processed
targets (list[BoxList]): ground-truth boxes present in the image (optional)
Returns:
result (list[BoxList] or dict[Tensor]): the output from the model.
During training, it returns a dict[Tensor] which contains the losses.
During testing, it returns list[BoxList] contains additional fields
like `scores`, `labels` and `mask` (for Mask R-CNN models).
"""
if self.training and targets is None:
raise ValueError("In training mode, targets should be passed")
# usually, it seems this is already an ImageList.
images = to_image_list(images)
features = self.backbone(images.tensors)
# for panoptic FPN, it seems like we might need to disentangle this so we can
# feed the features to the semantic head.
proposals, proposal_losses = self.rpn(images, features, targets)
if self.roi_heads:
x, result, detector_losses = self.roi_heads(
features, proposals, targets)
else:
# RPN-only models don't have roi_heads
x = features
result = proposals
detector_losses = {}
if self.training:
losses = {}
losses.update(detector_losses)
losses.update(proposal_losses)
return losses
return result
def forward(self, images, targets=None):
"""
Arguments:
images (list[Tensor] or ImageList): images to be processed
targets (list[BoxList]): ground-truth boxes present in the image (optional)
Returns:
result (list[BoxList] or dict[Tensor]): the output from the model.
During training, it returns a dict[Tensor] which contains the losses.
During testing, it returns list[BoxList] contains additional fields
like `scores`, `labels` and `mask` (for Mask R-CNN models).
"""
if self.training and targets is None:
raise ValueError("In training mode, targets should be passed")
images = to_image_list(images)
features = self.backbone(images.tensors)
# import ipdb;ipdb.set_trace()
##self.rpn为build_rpn中的RPNModule(cfg, in_channels)
##RPNModule返回boxes, losses
proposals, proposal_losses = self.rpn(images, features, targets)
if self.roi_heads:
##self.roi_heads为build_roi_heads中的CombinedROIHeads(cfg, roi_heads)
##CombinedROIHeads返回x, detections, losses
x, result, detector_losses = self.roi_heads(
features, proposals, targets)
else:
# RPN-only models don't have roi_heads
x = features
result = proposals
detector_losses = {}
if self.training:
losses = {}
losses.update(detector_losses)
losses.update(proposal_losses)
return losses
return result
def compute_prediction(self, original_image):
"""
Arguments:
original_image (np.ndarray): an image as returned by OpenCV
Returns:
prediction (BoxList): the detected objects. Additional information
of the detection properties can be found in the fields of
the BoxList via `prediction.fields()`
"""
# apply pre-processing to image
image = self.transforms(original_image)
# convert to an ImageList, padded so that it is divisible by
# cfg.DATALOADER.SIZE_DIVISIBILITY
image_list = to_image_list(image, self.cfg.DATALOADER.SIZE_DIVISIBILITY)
image_list = image_list.to(self.device)
# compute predictions
with torch.no_grad():
predictions = self.model(image_list)
predictions = [o.to(self.cpu_device) for o in predictions]
# always single image is passed at a time
prediction = predictions[0]
# reshape prediction (a BoxList) into the original image size
height, width = original_image.shape[:-1]
prediction = prediction.resize((width, height))
if prediction.has_field("mask"):
# if we have masks, paste the masks in the right position
# in the image, as defined by the bounding boxes
masks = prediction.get_field("mask")
# always single image is passed at a time
masks = self.masker([masks], [prediction])[0]
prediction.add_field("mask", masks)
return prediction
def __call__(self, batch):
transposed_batch = list(zip(*batch))
images = to_image_list(transposed_batch[0], self.size_divisible)
targets = transposed_batch[1]
img_ids = transposed_batch[2]
return images, targets, img_ids