def __init__(self, classes):
super(SSD, self).__init__()
self.size = cfg.TRAIN.COMMON.INPUT_SIZE
self.classes = classes
self.num_classes = len(self.classes)
self.priors_cfg = self._init_prior_cfg()
self.priorbox = PriorBox(self.priors_cfg)
self.priors_xywh = Variable(self.priorbox.forward(), volatile=True)
self.priors = torch.cat([
self.priors_xywh[:, 0:1] - 0.5 * self.priors_xywh[:, 2:3],
self.priors_xywh[:, 1:2] - 0.5 * self.priors_xywh[:, 3:4],
self.priors_xywh[:, 0:1] + 0.5 * self.priors_xywh[:, 2:3],
self.priors_xywh[:, 1:2] + 0.5 * self.priors_xywh[:, 3:4]
], 1)
self.priors = self.priors * self.size
# Layer learns to scale the l2 normalized features from conv4_3
self.L2Norm = L2Norm(512, 20)
self.softmax = nn.Softmax(dim=-1)
self._isex = cfg.TRAIN.VMRN.ISEX
self.VMRN_rel_op2l = _OP2L(cfg.VMRN.OP2L_POOLING_SIZE, cfg.VMRN.OP2L_POOLING_SIZE, 1.0/8.0, True)
self._train_iter_conter = 0
self.criterion = MultiBoxLoss(self.num_classes)
def __init__(self, classes):
super(_SSD, self).__init__()
self.size = cfg.TRAIN.COMMON.INPUT_SIZE
self.classes = classes
self.num_classes = len(self.classes)
self.priors_cfg = self._init_prior_cfg()
self.priorbox = PriorBox(self.priors_cfg)
self.priors_xywh = Variable(self.priorbox.forward())
self.priors_xywh.detach()
self.priors = torch.cat([
self.priors_xywh[:, 0:1] - 0.5 * self.priors_xywh[:, 2:3],
self.priors_xywh[:, 1:2] - 0.5 * self.priors_xywh[:, 3:4],
self.priors_xywh[:, 0:1] + 0.5 * self.priors_xywh[:, 2:3],
self.priors_xywh[:, 1:2] + 0.5 * self.priors_xywh[:, 3:4]
], 1)
self.priors = self.priors * self.size
# Layer learns to scale the l2 normalized features from conv4_3
self.L2Norm = L2Norm(512, 20)
self.softmax = nn.Softmax(dim=-1)
self.criterion = MultiBoxLoss(self.num_classes)
class SSD(nn.Module):
"""Single Shot Multibox Architecture
The network is composed of a base VGG network followed by the
added multibox conv layers. Each multibox layer branches into
1) conv2d for class conf scores
2) conv2d for localization predictions
3) associated priorbox layer to produce default bounding
boxes specific to the layer's feature map size.
See: https://arxiv.org/pdf/1512.02325.pdf for more details.
Args:
phase: (string) Can be "test" or "train"
size: input image size
base: VGG16 layers for input, size of either 300 or 500
extras: extra layers that feed to multibox loc and conf layers
head: "multibox head" consists of loc and conf conv layers
"""
def __init__(self, classes):
super(SSD, self).__init__()
self.size = cfg.TRAIN.COMMON.INPUT_SIZE
self.classes = classes
self.num_classes = len(self.classes)
self.priors_cfg = self._init_prior_cfg()
self.priorbox = PriorBox(self.priors_cfg)
self.priors_xywh = Variable(self.priorbox.forward(), volatile=True)
self.priors = torch.cat([
self.priors_xywh[:, 0:1] - 0.5 * self.priors_xywh[:, 2:3],
self.priors_xywh[:, 1:2] - 0.5 * self.priors_xywh[:, 3:4],
self.priors_xywh[:, 0:1] + 0.5 * self.priors_xywh[:, 2:3],
self.priors_xywh[:, 1:2] + 0.5 * self.priors_xywh[:, 3:4]
], 1)
self.priors = self.priors * self.size
# Layer learns to scale the l2 normalized features from conv4_3
self.L2Norm = L2Norm(512, 20)
self.softmax = nn.Softmax(dim=-1)
self._isex = cfg.TRAIN.VMRN.ISEX
self.VMRN_rel_op2l = _OP2L(cfg.VMRN.OP2L_POOLING_SIZE, cfg.VMRN.OP2L_POOLING_SIZE, 1.0/8.0, True)
self._train_iter_conter = 0
self.criterion = MultiBoxLoss(self.num_classes)
def forward(self, x, im_info, gt_boxes, num_boxes, rel_mat):
"""Applies network layers and ops on input image(s) x.
Args:
x: input image or batch of images. Shape: [batch,3,300,300].
Return:
Depending on phase:
test:
Variable(tensor) of output class label predictions,
confidence score, and corresponding location predictions for
each object detected. Shape: [batch,topk,7]
train:
list of concat outputs from:
1: confidence layers, Shape: [batch*num_priors,num_classes]
2: localization layers, Shape: [batch,num_priors*4]
3: priorbox layers, Shape: [2,num_priors*4]
"""
self._train_iter_conter += 1
sources = list()
loc = list()
conf = list()
self.batch_size = x.size(0)
# apply vgg up to conv4_3 relu
if isinstance(self.base, nn.ModuleList):
for k,v in enumerate(self.base):
x = v(x)
else:
x = self.base(x)
s = self.L2Norm(x)
sources.append(s)
base_feat = s
# apply vgg up to fc7
if isinstance(self.conv5, nn.ModuleList):
for k,v in enumerate(self.conv5):
x = v(x)
else:
x = self.conv5(x)
sources.append(x)
# apply extra layers and cache source layer outputs
for k, v in enumerate(self.extras):
x = F.relu(v(x), inplace=True)
if k % 2 == 1:
sources.append(x)
# apply multibox head to source layers
for (x, l, c) in zip(sources, self.loc, self.conf):
loc.append(l(x).permute(0, 2, 3, 1).contiguous())
conf.append(c(x).permute(0, 2, 3, 1).contiguous())
loc = torch.cat([o.view(o.size(0), -1) for o in loc], 1)
conf = torch.cat([o.view(o.size(0), -1) for o in conf], 1)
loc = loc.view(loc.size(0), -1, 4)
conf = conf.view(conf.size(0), -1, self.num_classes)
SSD_loss_cls = 0
SSD_loss_bbox = 0
if self.training:
predictions = (
loc,
conf,
self.priors.type_as(loc)
)
# targets = torch.cat([gt_boxes[:,:,:4] / self.size, gt_boxes[:,:,4:5]],dim=2)
targets = gt_boxes
SSD_loss_bbox, SSD_loss_cls = self.criterion(predictions, targets, num_boxes)
conf = self.softmax(conf)
# online data
if self.training:
if self._train_iter_conter > cfg.TRAIN.VMRN.ONLINEDATA_BEGIN_ITER:
obj_rois, obj_num = self._obj_det(conf, loc, self.batch_size, im_info)
obj_rois = obj_rois.type_as(gt_boxes)
obj_num = obj_num.type_as(num_boxes)
else:
obj_rois = torch.FloatTensor([]).type_as(gt_boxes)
obj_num = torch.LongTensor([]).type_as(num_boxes)
obj_labels = None
else:
# when testing, this is object detection results
# TODO: SUPPORT MULTI-IMAGE BATCH
obj_rois, obj_num = self._obj_det(conf, loc, self.batch_size, im_info)
if obj_rois.numel() > 0:
obj_labels = obj_rois[:, 5]
obj_rois = obj_rois[:, :5]
obj_rois = obj_rois.type_as(gt_boxes)
obj_num = obj_num.type_as(num_boxes)
else:
# there is no object detected
obj_labels = torch.Tensor([]).type_as(gt_boxes).long()
obj_rois = obj_rois.type_as(gt_boxes)
obj_num = obj_num.type_as(num_boxes)
if self.training:
# offline data
for i in range(self.batch_size):
obj_rois = torch.cat([obj_rois,
torch.cat([(i * torch.ones(num_boxes[i].item(), 1)).type_as(gt_boxes),
(gt_boxes[i][:num_boxes[i]][:, 0:4])], 1)
])
obj_num = torch.cat([obj_num, torch.Tensor([num_boxes[i]]).type_as(obj_num)])
obj_rois = Variable(obj_rois)
VMRN_rel_loss_cls = 0
rel_cls_prob = torch.Tensor([]).type_as(obj_rois)
if (obj_num > 1).sum().item() > 0:
obj_pair_feat = self.VMRN_rel_op2l(base_feat, obj_rois, self.batch_size, obj_num)
# obj_pair_feat = obj_pair_feat.detach()
obj_pair_feat = self._rel_head_to_tail(obj_pair_feat)
rel_cls_score = self.VMRN_rel_cls_score(obj_pair_feat)
rel_cls_prob = F.softmax(rel_cls_score)
self.rel_batch_size = obj_pair_feat.size(0)
if self.training:
obj_pair_rel_label = self._generate_rel_labels(obj_rois, gt_boxes, obj_num, rel_mat)
obj_pair_rel_label = obj_pair_rel_label.type_as(gt_boxes).long()
rel_not_keep = (obj_pair_rel_label == 0)
# no relationship is kept
if (rel_not_keep == 0).sum().item() > 0:
rel_keep = torch.nonzero(rel_not_keep == 0).view(-1)
rel_cls_score = rel_cls_score[rel_keep]
obj_pair_rel_label = obj_pair_rel_label[rel_keep]
obj_pair_rel_label -= 1
VMRN_rel_loss_cls = F.cross_entropy(rel_cls_score, obj_pair_rel_label)
else:
if (not cfg.TEST.VMRN.ISEX) and cfg.TRAIN.VMRN.ISEX:
rel_cls_prob = rel_cls_prob[::2, :]
rel_result = None
if not self.training:
if obj_rois.numel() > 0:
pred_boxes = obj_rois.data[:,1:5]
pred_boxes[:, 0::2] /= im_info[0][3].item()
pred_boxes[:, 1::2] /= im_info[0][2].item()
rel_result = (pred_boxes, obj_labels, rel_cls_prob.data)
else:
rel_result = (obj_rois.data, obj_labels, rel_cls_prob.data)
return loc, conf, rel_result, SSD_loss_bbox, SSD_loss_cls, VMRN_rel_loss_cls
def _generate_rel_labels(self, obj_rois, gt_boxes, obj_num, rel_mat):
obj_pair_rel_label = torch.Tensor(self.rel_batch_size).type_as(gt_boxes).zero_().long()
# generate online data labels
cur_pair = 0
for i in range(obj_num.size(0)):
img_index = i % self.batch_size
if obj_num[i] <=1 :
continue
begin_ind = torch.sum(obj_num[:i])
overlaps = bbox_overlaps(obj_rois[begin_ind:begin_ind + obj_num[i]][:, 1:5],
gt_boxes[img_index][:, 0:4])
max_overlaps, max_inds = torch.max(overlaps, 1)
for o1ind in range(obj_num[i]):
for o2ind in range(o1ind + 1, obj_num[i]):
o1_gt = int(max_inds[o1ind].item())
o2_gt = int(max_inds[o2ind].item())
if o1_gt == o2_gt:
# skip invalid pairs
if self._isex:
cur_pair += 2
else:
cur_pair += 1
continue
# some labels are leaved out when labeling
if rel_mat[img_index][o1_gt, o2_gt].item() == 0:
if rel_mat[img_index][o2_gt, o1_gt].item() == 3:
rel_mat[img_index][o1_gt, o2_gt] = rel_mat[img_index][o2_gt, o1_gt]
else:
rel_mat[img_index][o1_gt, o2_gt] = 3 - rel_mat[img_index][o2_gt, o1_gt]
obj_pair_rel_label[cur_pair] = rel_mat[img_index][o1_gt, o2_gt]
cur_pair += 1
if self._isex:
# some labels are leaved out when labeling
if rel_mat[img_index][o2_gt, o1_gt].item() == 0:
if rel_mat[img_index][o1_gt, o2_gt].item() == 3:
rel_mat[img_index][o2_gt, o1_gt] = rel_mat[img_index][o1_gt, o2_gt]
else:
rel_mat[img_index][o2_gt, o1_gt] = 3 - rel_mat[img_index][o1_gt, o2_gt]
obj_pair_rel_label[cur_pair] = rel_mat[img_index][o2_gt, o1_gt]
cur_pair += 1
return obj_pair_rel_label
def load_weights(self, base_file):
other, ext = os.path.splitext(base_file)
if ext == '.pkl' or '.pth':
print('Loading weights into state dict...')
self.load_state_dict(torch.load(base_file,
map_location=lambda storage, loc: storage))
print('Finished!')
else:
print('Sorry only .pth and .pkl files supported.')
def _obj_det(self, conf, loc, batch_size, im_info):
det_results = torch.Tensor([]).type_as(loc)
obj_num = []
if not self.training:
det_labels = torch.Tensor([]).type_as(loc).long()
for i in range(batch_size):
cur_cls_prob = conf[i:i + 1]
cur_bbox_pred = loc[i:i + 1]
cur_im_info = im_info[i:i + 1]
obj_boxes = self._get_single_obj_det_results(cur_cls_prob, cur_bbox_pred, cur_im_info)
obj_num.append(obj_boxes.size(0))
if obj_num[-1] > 0:
det_results = torch.cat([det_results,
torch.cat([i * torch.ones(obj_boxes.size(0), 1).type_as(det_results),
obj_boxes], 1)
], 0)
return det_results, torch.LongTensor(obj_num)
def _get_single_obj_det_results(self, cls_prob, bbox_pred, im_info):
scores = cls_prob.data
thresh = 0.05 # filter out low confidence boxes for acceleration
results = []
if cfg.TEST.COMMON.BBOX_REG:
# Apply bounding-box regression deltas
box_deltas = bbox_pred.data
if cfg.TRAIN.COMMON.BBOX_NORMALIZE_TARGETS_PRECOMPUTED:
# Optionally normalize targets by a precomputed mean and stdev
box_deltas = box_deltas.view(-1, 4) * torch.FloatTensor(cfg.TRAIN.COMMON.BBOX_NORMALIZE_STDS).type_as(box_deltas) \
+ torch.FloatTensor(cfg.TRAIN.COMMON.BBOX_NORMALIZE_MEANS).type_as(box_deltas)
box_deltas = box_deltas.view(1, -1, 4)
pred_boxes = bbox_transform_inv(self.priors.type_as(bbox_pred).data, box_deltas, 1)
pred_boxes = clip_boxes(pred_boxes, im_info.data, 1)
else:
# Simply repeat the boxes, once for each class
pred_boxes = np.tile(self.priors.data, (1, scores.shape[1]))
scores = scores.squeeze()
pred_boxes = pred_boxes.squeeze()
for j in xrange(1, self.num_classes):
inds = torch.nonzero(scores[:, j] > thresh).view(-1)
# if there is det
if inds.numel() > 0:
cls_scores = scores[:, j][inds]
_, order = torch.sort(cls_scores, 0, True)
cls_boxes = pred_boxes[inds, :]
cls_dets = torch.cat((cls_boxes, cls_scores.unsqueeze(1)), 1)
# cls_dets = torch.cat((cls_boxes, cls_scores), 1)
cls_dets = cls_dets[order]
keep = nms(cls_dets, cfg.TEST.COMMON.NMS)
cls_dets = cls_dets[keep.view(-1).long()]
final_keep = torch.nonzero(cls_dets[:, -1] > cfg.TEST.COMMON.OBJ_DET_THRESHOLD).squeeze()
result = cls_dets[final_keep]
if result.numel()>0 and result.dim() == 1:
result = result.unsqueeze(0)
# in testing, concat object labels
if final_keep.numel() > 0:
if self.training:
result = result[:,:4]
else:
result = torch.cat([result[:,:4],
j * torch.ones(result.size(0),1).type_as(result)],1)
if result.numel() > 0:
results.append(result)
if len(results):
final = torch.cat(results, 0)
else:
final = torch.Tensor([]).type_as(bbox_pred)
return final
def create_architecture(self):
self._init_modules()
def weights_init(m):
def xavier(param):
init.xavier_uniform(param)
if isinstance(m, nn.Conv2d):
xavier(m.weight.data)
m.bias.data.zero_()
# initialize newly added layers' weights with xavier method
self.extras.apply(weights_init)
self.loc.apply(weights_init)
self.conf.apply(weights_init)
def _init_prior_cfg(self):
prior_cfg = {
'min_dim': self.size,
'feature_maps': cfg.SSD.FEATURE_MAPS,
'min_sizes': cfg.SSD.PRIOR_MIN_SIZE,
'max_sizes': cfg.SSD.PRIOR_MAX_SIZE,
'steps': cfg.SSD.PRIOR_STEP,
'aspect_ratios':cfg.SSD.PRIOR_ASPECT_RATIO,
'clip':cfg.SSD.PRIOR_CLIP
}
return prior_cfg
def resume_iter(self, epoch, iter_per_epoch):
self._train_iter_conter = epoch * iter_per_epoch
def __init__(self, classes, class_agnostic, feat_name, feat_list=('conv3', 'conv4'), pretrained = True):
super(SSD, self).__init__(classes, class_agnostic, feat_name, feat_list, pretrained)
self.FeatExt.feat_layer["conv3"][0].ceil_mode = True
##### Important to set model to eval mode before evaluation ####
self.FeatExt.eval()
rand_img = torch.Tensor(1, 3, 300, 300)
rand_feat = self.FeatExt(rand_img)
self.FeatExt.train()
n_channels = [f.size(1) for f in rand_feat]
self.size = cfg.SCALES[0]
self.priors_cfg = self._init_prior_cfg()
self.priorbox = PriorBox(self.priors_cfg)
self.priors_xywh = Variable(self.priorbox.forward())
self.priors_xywh.detach()
self.priors = torch.cat([
self.priors_xywh[:, 0:1] - 0.5 * self.priors_xywh[:, 2:3],
self.priors_xywh[:, 1:2] - 0.5 * self.priors_xywh[:, 3:4],
self.priors_xywh[:, 0:1] + 0.5 * self.priors_xywh[:, 2:3],
self.priors_xywh[:, 1:2] + 0.5 * self.priors_xywh[:, 3:4]
], 1)
self.priors = self.priors * self.size
# Layer learns to scale the l2 normalized features from conv4_3
self.L2Norm = L2Norm(512, 20)
self.softmax = nn.Softmax(dim=-1)
self.criterion = MultiBoxLoss(self.n_classes)
mbox_cfg = []
for i in cfg.SSD.PRIOR_ASPECT_RATIO:
mbox_cfg.append(2 * len(i) + 2)
self.extra_conv = nn.ModuleList()
self.loc = nn.ModuleList()
self.conf = nn.ModuleList()
# conv 4_3 detector
self.loc.append(
nn.Conv2d(n_channels[0], mbox_cfg[0] * 4 if self.class_agnostic else mbox_cfg[0] * 4 * self.n_classes
, kernel_size=3, padding=1))
self.conf.append(nn.Conv2d(n_channels[0], mbox_cfg[0] * self.n_classes, kernel_size=3, padding=1))
# conv 7 detector
self.extra_conv.append(nn.Sequential(
nn.MaxPool2d(kernel_size=3, stride=1, padding=1),
nn.Conv2d(512, 1024, kernel_size=3, padding=6, dilation=6),
nn.ReLU(inplace=True),
nn.Conv2d(1024, 1024, kernel_size=1),
nn.ReLU(inplace=True)))
self.loc.append(nn.Conv2d(1024, mbox_cfg[1] * 4 if self.class_agnostic else mbox_cfg[1] * 4 * self.n_classes,
kernel_size=3, padding=1))
self.conf.append(nn.Conv2d(1024, mbox_cfg[1] * self.n_classes, kernel_size=3, padding=1))
def add_extra_conv(extra_conv, loc, conf, in_c, mid_c, out_c, downsamp, mbox, n_cls, cag):
extra_conv.append(nn.Sequential(
nn.Conv2d(in_c, mid_c, kernel_size=1),
nn.ReLU(inplace=True),
nn.Conv2d(mid_c, out_c, kernel_size=3, stride=2 if downsamp else 1, padding=1 if downsamp else 0),
nn.ReLU(inplace=True),
))
loc.append(nn.Conv2d(out_c, mbox * 4 if cag else mbox * 4 * n_cls, kernel_size=3, padding=1))
conf.append(nn.Conv2d(out_c, mbox * n_cls, kernel_size=3, padding=1))
add_extra_conv(self.extra_conv, self.loc, self.conf, 1024, 256, 512, True, mbox_cfg[2], self.n_classes, self.class_agnostic)
add_extra_conv(self.extra_conv, self.loc, self.conf, 512, 128, 256, True, mbox_cfg[3], self.n_classes, self.class_agnostic)
add_extra_conv(self.extra_conv, self.loc, self.conf, 256, 128, 256, False, mbox_cfg[4], self.n_classes, self.class_agnostic)
add_extra_conv(self.extra_conv, self.loc, self.conf, 256, 128, 256, False, mbox_cfg[5], self.n_classes, self.class_agnostic)
self.iter_counter = 0
class SSD(objectDetector):
def __init__(self, classes, class_agnostic, feat_name, feat_list=('conv3', 'conv4'), pretrained = True):
super(SSD, self).__init__(classes, class_agnostic, feat_name, feat_list, pretrained)
self.FeatExt.feat_layer["conv3"][0].ceil_mode = True
##### Important to set model to eval mode before evaluation ####
self.FeatExt.eval()
rand_img = torch.Tensor(1, 3, 300, 300)
rand_feat = self.FeatExt(rand_img)
self.FeatExt.train()
n_channels = [f.size(1) for f in rand_feat]
self.size = cfg.SCALES[0]
self.priors_cfg = self._init_prior_cfg()
self.priorbox = PriorBox(self.priors_cfg)
self.priors_xywh = Variable(self.priorbox.forward())
self.priors_xywh.detach()
self.priors = torch.cat([
self.priors_xywh[:, 0:1] - 0.5 * self.priors_xywh[:, 2:3],
self.priors_xywh[:, 1:2] - 0.5 * self.priors_xywh[:, 3:4],
self.priors_xywh[:, 0:1] + 0.5 * self.priors_xywh[:, 2:3],
self.priors_xywh[:, 1:2] + 0.5 * self.priors_xywh[:, 3:4]
], 1)
self.priors = self.priors * self.size
# Layer learns to scale the l2 normalized features from conv4_3
self.L2Norm = L2Norm(512, 20)
self.softmax = nn.Softmax(dim=-1)
self.criterion = MultiBoxLoss(self.n_classes)
mbox_cfg = []
for i in cfg.SSD.PRIOR_ASPECT_RATIO:
mbox_cfg.append(2 * len(i) + 2)
self.extra_conv = nn.ModuleList()
self.loc = nn.ModuleList()
self.conf = nn.ModuleList()
# conv 4_3 detector
self.loc.append(
nn.Conv2d(n_channels[0], mbox_cfg[0] * 4 if self.class_agnostic else mbox_cfg[0] * 4 * self.n_classes
, kernel_size=3, padding=1))
self.conf.append(nn.Conv2d(n_channels[0], mbox_cfg[0] * self.n_classes, kernel_size=3, padding=1))
# conv 7 detector
self.extra_conv.append(nn.Sequential(
nn.MaxPool2d(kernel_size=3, stride=1, padding=1),
nn.Conv2d(512, 1024, kernel_size=3, padding=6, dilation=6),
nn.ReLU(inplace=True),
nn.Conv2d(1024, 1024, kernel_size=1),
nn.ReLU(inplace=True)))
self.loc.append(nn.Conv2d(1024, mbox_cfg[1] * 4 if self.class_agnostic else mbox_cfg[1] * 4 * self.n_classes,
kernel_size=3, padding=1))
self.conf.append(nn.Conv2d(1024, mbox_cfg[1] * self.n_classes, kernel_size=3, padding=1))
def add_extra_conv(extra_conv, loc, conf, in_c, mid_c, out_c, downsamp, mbox, n_cls, cag):
extra_conv.append(nn.Sequential(
nn.Conv2d(in_c, mid_c, kernel_size=1),
nn.ReLU(inplace=True),
nn.Conv2d(mid_c, out_c, kernel_size=3, stride=2 if downsamp else 1, padding=1 if downsamp else 0),
nn.ReLU(inplace=True),
))
loc.append(nn.Conv2d(out_c, mbox * 4 if cag else mbox * 4 * n_cls, kernel_size=3, padding=1))
conf.append(nn.Conv2d(out_c, mbox * n_cls, kernel_size=3, padding=1))
add_extra_conv(self.extra_conv, self.loc, self.conf, 1024, 256, 512, True, mbox_cfg[2], self.n_classes, self.class_agnostic)
add_extra_conv(self.extra_conv, self.loc, self.conf, 512, 128, 256, True, mbox_cfg[3], self.n_classes, self.class_agnostic)
add_extra_conv(self.extra_conv, self.loc, self.conf, 256, 128, 256, False, mbox_cfg[4], self.n_classes, self.class_agnostic)
add_extra_conv(self.extra_conv, self.loc, self.conf, 256, 128, 256, False, mbox_cfg[5], self.n_classes, self.class_agnostic)
self.iter_counter = 0
def _get_obj_det_result(self, sources):
loc = []
conf = []
# apply multibox head to source layers
for (x, l, c) in zip(sources, self.loc, self.conf):
loc.append(l(x).permute(0, 2, 3, 1).contiguous())
conf.append(c(x).permute(0, 2, 3, 1).contiguous())
loc = torch.cat([o.view(o.size(0), -1) for o in loc], 1)
conf = torch.cat([o.view(o.size(0), -1) for o in conf], 1)
loc = loc.view(loc.size(0), -1, 4)
conf = conf.view(conf.size(0), -1, self.n_classes)
return loc, conf
def forward(self, data_batch):
"""Applies network layers and ops on input image(s) x.
Args:
x: input image or batch of images. Shape: [batch,3,300,300].
Return:
Depending on phase:
test:
Variable(tensor) of output class label predictions,
confidence score, and corresponding location predictions for
each object detected. Shape: [batch,topk,7]
train:
list of concat outputs from:
1: confidence layers, Shape: [batch*num_priors,num_classes]
2: localization layers, Shape: [batch,num_priors*4]
3: priorbox layers, Shape: [2,num_priors*4]
"""
x = data_batch[0]
im_info = data_batch[1]
gt_boxes = data_batch[2]
num_boxes = data_batch[3]
if self.training:
self.iter_counter += 1
sources = []
s0, x = self.FeatExt(x)
s0 = self.L2Norm(s0)
sources.append(s0)
for m in self.extra_conv:
x = m(x)
sources.append(x)
loc, conf = self._get_obj_det_result(sources)
SSD_loss_cls, SSD_loss_bbox = 0, 0
if self.training:
predictions = (
loc,
conf,
self.priors.type_as(loc)
)
SSD_loss_bbox, SSD_loss_cls = self.criterion(predictions, gt_boxes, num_boxes)
conf = self.softmax(conf)
return loc, conf, SSD_loss_bbox, SSD_loss_cls
def create_architecture(self):
self._init_modules()
self._init_weights()
def _init_modules(self):
pass
def _init_weights(self):
def weights_init(m):
def xavier(param):
init.xavier_uniform(param)
if isinstance(m, nn.Conv2d):
xavier(m.weight.data)
m.bias.data.zero_()
# initialize newly added layers' weights with xavier method
self.extra_conv.apply(weights_init)
self.loc.apply(weights_init)
self.conf.apply(weights_init)
def _init_prior_cfg(self):
prior_cfg = {
'min_dim': self.size,
'feature_maps': cfg.SSD.FEATURE_MAPS,
'min_sizes': cfg.SSD.PRIOR_MIN_SIZE,
'max_sizes': cfg.SSD.PRIOR_MAX_SIZE,
'steps': cfg.SSD.PRIOR_STEP,
'aspect_ratios':cfg.SSD.PRIOR_ASPECT_RATIO,
'clip':cfg.SSD.PRIOR_CLIP
}
return prior_cfg
class _SSD(nn.Module):
"""Single Shot Multibox Architecture
The network is composed of a base VGG network followed by the
added multibox conv layers. Each multibox layer branches into
1) conv2d for class conf scores
2) conv2d for localization predictions
3) associated priorbox layer to produce default bounding
boxes specific to the layer's feature map size.
See: https://arxiv.org/pdf/1512.02325.pdf for more details.
Args:
phase: (string) Can be "test" or "train"
size: input image size
base: VGG16 layers for input, size of either 300 or 500
extras: extra layers that feed to multibox loc and conf layers
head: "multibox head" consists of loc and conf conv layers
"""
def __init__(self, classes):
super(_SSD, self).__init__()
self.size = cfg.TRAIN.COMMON.INPUT_SIZE
self.classes = classes
self.num_classes = len(self.classes)
self.priors_cfg = self._init_prior_cfg()
self.priorbox = PriorBox(self.priors_cfg)
self.priors_xywh = Variable(self.priorbox.forward())
self.priors_xywh.detach()
self.priors = torch.cat([
self.priors_xywh[:, 0:1] - 0.5 * self.priors_xywh[:, 2:3],
self.priors_xywh[:, 1:2] - 0.5 * self.priors_xywh[:, 3:4],
self.priors_xywh[:, 0:1] + 0.5 * self.priors_xywh[:, 2:3],
self.priors_xywh[:, 1:2] + 0.5 * self.priors_xywh[:, 3:4]
], 1)
self.priors = self.priors * self.size
# Layer learns to scale the l2 normalized features from conv4_3
self.L2Norm = L2Norm(512, 20)
self.softmax = nn.Softmax(dim=-1)
self.criterion = MultiBoxLoss(self.num_classes)
def forward(self, x, im_info, gt_boxes, num_boxes):
"""Applies network layers and ops on input image(s) x.
Args:
x: input image or batch of images. Shape: [batch,3,300,300].
Return:
Depending on phase:
test:
Variable(tensor) of output class label predictions,
confidence score, and corresponding location predictions for
each object detected. Shape: [batch,topk,7]
train:
list of concat outputs from:
1: confidence layers, Shape: [batch*num_priors,num_classes]
2: localization layers, Shape: [batch,num_priors*4]
3: priorbox layers, Shape: [2,num_priors*4]
"""
sources = list()
loc = list()
conf = list()
# apply vgg up to conv4_3 relu
if isinstance(self.base, nn.ModuleList):
for layer in self.base:
x = layer(x)
else:
x = self.base(x)
s = self.L2Norm(x)
sources.append(s)
# apply vgg up to fc7
for conv in self.SSD_feat_layers:
x = conv(x)
sources.append(x)
# apply extra layers and cache source layer outputs
for k, v in enumerate(self.extras):
x = F.relu(v(x), inplace=True)
if k % 2 == 1:
sources.append(x)
# apply multibox head to source layers
for (x, l, c) in zip(sources, self.loc, self.conf):
loc.append(l(x).permute(0, 2, 3, 1).contiguous())
conf.append(c(x).permute(0, 2, 3, 1).contiguous())
loc = torch.cat([o.view(o.size(0), -1) for o in loc], 1)
conf = torch.cat([o.view(o.size(0), -1) for o in conf], 1)
loc = loc.view(loc.size(0), -1, 4)
conf = conf.view(conf.size(0), -1, self.num_classes)
if self.training:
predictions = (
loc,
conf,
self.priors.type_as(loc)
)
# targets = torch.cat([gt_boxes[:,:,:4] / self.size, gt_boxes[:,:,4:5]],dim=2)
targets = gt_boxes
SSD_loss_bbox, SSD_loss_cls = self.criterion(predictions, targets, num_boxes)
else:
SSD_loss_cls = 0
SSD_loss_bbox = 0
conf = self.softmax(conf)
return loc, conf, SSD_loss_bbox, SSD_loss_cls
def create_architecture(self):
self._init_modules()
def weights_init(m):
def xavier(param):
init.xavier_uniform(param)
if isinstance(m, nn.Conv2d):
xavier(m.weight.data)
m.bias.data.zero_()
# initialize newly added layers' weights with xavier method
self.extras.apply(weights_init)
self.loc.apply(weights_init)
self.conf.apply(weights_init)
def _init_prior_cfg(self):
prior_cfg = {
'min_dim': self.size,
'feature_maps': cfg.SSD.FEATURE_MAPS,
'min_sizes': cfg.SSD.PRIOR_MIN_SIZE,
'max_sizes': cfg.SSD.PRIOR_MAX_SIZE,
'steps': cfg.SSD.PRIOR_STEP,
'aspect_ratios':cfg.SSD.PRIOR_ASPECT_RATIO,
'clip':cfg.SSD.PRIOR_CLIP
}
return prior_cfg