def AddExtraTopDownLayers(net, use_batchnorm=True, lr_mult=1):
# mbox_source_layers = ['conv4_3', 'fc7', 'conv6_2', 'conv7_2', 'conv8_2', 'conv9_2']
bbox = "cls_specific_bbox"
use_relu = True
# 1 x 1
from_layer = net.keys()[-1]
out_layer = "decls"
DeConvBNLayer(net, from_layer, out_layer, use_batchnorm, use_relu, 256, 1, 0, 1, lr_mult=lr_mult)
# crop feature form bottom-up net
from_layer = "conv9_2"
out_layer = "conv9_2_crop"
net[out_layer] = L.CropBBox(net[from_layer], net[bbox])
# concatenate the cropped feature and the class-specific top-down signals
out_layer = "deconv9_2_concat"
net[out_layer] = L.Concat(net["conv9_2_crop"], net["decls"])
from_layer = out_layer
out_layer = "deconv9_2"
DeConvBNLayer(net, from_layer, out_layer, use_batchnorm, use_relu, 128, 3, 0, 1,
lr_mult=lr_mult)
from_layer = out_layer
out_layer = "deconv9_1"
DeConvBNLayer(net, from_layer, out_layer, use_batchnorm, use_relu, 256, 1, 0, 1,
lr_mult=lr_mult)
# crop feature form bottom-up net
from_layer = "conv8_2"
out_layer = "conv8_2_crop"
net[out_layer] = L.CropBBox(net[from_layer], net[bbox])
# concatenate the cropped feature and the class-specific top-down signals
out_layer = "deconv8_2_concat"
net[out_layer] = L.Concat(net["conv8_2_crop"], net["deconv9_1"])
# 3 x 3
from_layer = out_layer
out_layer = "deconv8_2"
DeConvBNLayer(net, from_layer, out_layer, use_batchnorm, use_relu, 128, 3, 0, 1,
lr_mult=lr_mult)
from_layer = out_layer
out_layer = "deconv8_1"
DeConvBNLayer(net, from_layer, out_layer, use_batchnorm, use_relu, 256, 1, 0, 1,
lr_mult=lr_mult)
# crop feature form bottom-up net
from_layer = "conv7_2"
out_layer = "conv7_2_crop"
net[out_layer] = L.CropBBox(net[from_layer], net[bbox])
# concatenate the cropped feature and the class-specific top-down signals
out_layer = "deconv7_2_concat"
net[out_layer] = L.Concat(net["conv7_2_crop"], net["deconv8_1"])
# 5 x 5
from_layer = out_layer
out_layer = "deconv7_2"
DeConvBNLayer(net, from_layer, out_layer, use_batchnorm, use_relu, 128, 3, 1, 2,
lr_mult=lr_mult)
from_layer = out_layer
out_layer = "deconv7_1"
DeConvBNLayer(net, from_layer, out_layer, use_batchnorm, use_relu, 512, 1, 0, 1,
lr_mult=lr_mult)
# crop feature form bottom-up net
from_layer = "conv6_2"
out_layer = "conv6_2_crop"
net[out_layer] = L.CropBBox(net[from_layer], net[bbox])
# concatenate the cropped feature and the class-specific top-down signals
out_layer = "deconv6_2_concat"
net[out_layer] = L.Concat(net["conv6_2_crop"], net["deconv7_1"])
# 10 x 10
from_layer = out_layer
out_layer = "deconv6_2"
DeConvBNLayer(net, from_layer, out_layer, use_batchnorm, use_relu, 256, 3, 1, 2,
lr_mult=lr_mult)
from_layer = out_layer
out_layer = "deconv6_1"
DeConvBNLayer(net, from_layer, out_layer, use_batchnorm, use_relu, 1024, 1, 0, 1,
lr_mult=lr_mult)
return net
n = caffe.NetSpec()
n.data = L.DummyData(shape=dict(dim=[2, 2, 6, 6]),
data_filler=dict(type='constant', value=1))
n.data_conv = L.Convolution(n.data,
kernel_size=3,
pad=1,
stride=1,
num_output=2)
n.bbox = L.DummyData(shape=dict(dim=[1, 1, 2, 8]),
data_filler=dict(type='constant', value=0))
n.data_crop = L.CropBBox(n.data_conv, n.bbox)
n.dummy_loss = L.EuclideanLoss(n.data_conv, n.data_crop)
net_def = "test_crop_bbox_simple.prototxt"
with open(net_def, 'w') as f:
f.write(str(n.to_proto()))
caffe.set_device(0)
caffe.set_mode_gpu()
net = caffe.Net(net_def, caffe.TRAIN)
img_labels = [[1], [2]]
bbox = np.zeros(net.blobs['bbox'].data.shape).astype(np.float)
bbox[0, 0, 0, 0] = 0
kwargs = {
'include': dict(phase=caffe_pb2.Phase.Value('TRAIN')),
'transform_param': transform_param
}
label_map_file = "data/VOC0712/labelmap_voc.prototxt"
bbox_seg_data_param = {
'label_map_file': label_map_file,
'batch_sampler': batch_sampler,
}
source = "/home/amax/NiuChuang/data/VOCdevkit/VOC0712/lmdb/VOC0712_trainval_lmdb_test2"
net.data, net.bbox, net.seg = L.BBoxSegData(
name="data",
annotated_data_param=bbox_seg_data_param,
data_param=dict(batch_size=8, backend=P.Data.LMDB, source=source),
ntop=3,
**kwargs)
net.cls_specific_bbox, net.binary_mask, net.cls = L.SelectBinary(
net.bbox, net.seg, random_select=True, num_class=20, ntop=3)
net = VGGNetBody(net, 'data', need_fc=False)
layer_name = net.keys()[-1]
# layer_crop_name = layer_name + '_crop'
net.conv_crop = L.CropBBox(net[layer_name], net.cls_specific_bbox)
with open("test_crop_bbox.prototxt", 'w') as f:
f.write(str(net.to_proto()))
def vgg16_ssd_seg(source,
bbox_seg_data_param,
kwargs,
use_batchnorm=False,
lr_mult=1,
crop_layers=[],
is_crop_last=False,
is_cls=False,
is_deploy=False,
is_crop_all=False,
is_crop_cls=False,
is_crop_merge_feature=False):
if crop_layers is None:
crop_layers = [
'conv4_3', 'fc7', 'conv6_2', 'conv7_2', 'conv8_2', 'conv9_2'
]
net = caffe.NetSpec()
if is_deploy:
net.data = L.Input(input_param=dict(shape=dict(dim=[1, 3, 320, 320])))
net.cls_specific_bbox = L.Input(input_param=dict(shape=dict(
dim=[1, 1, 1, 8])))
if is_cls:
net.cls = L.Input(input_param=dict(shape=dict(dim=[1, 20])))
else:
net.data, net.bbox, net.seg = L.BBoxSegData(
name="data",
annotated_data_param=bbox_seg_data_param,
data_param=dict(batch_size=8, backend=P.Data.LMDB, source=source),
ntop=3,
**kwargs)
net.cls_specific_bbox, net.binary_mask, net.cls = L.SelectBinary(
net.bbox, net.seg, random_select=True, num_class=20, ntop=3)
VGGNetBody(net,
from_layer='data',
fully_conv=True,
reduced=True,
dilated=True,
dropout=False,
pool_mask=True,
freeze_all=True)
AddExtraLayers(net, use_batchnorm, lr_mult=0)
if is_deploy:
# MultiBoxLoss parameters.
num_classes = 21
share_location = True
background_label_id = 0
train_on_diff_gt = True
normalization_mode = P.Loss.VALID
code_type = P.PriorBox.CENTER_SIZE
ignore_cross_boundary_bbox = False
mining_type = P.MultiBoxLoss.MAX_NEGATIVE
neg_pos_ratio = 3.
loc_weight = (neg_pos_ratio + 1.) / 4.
multibox_loss_param = {
'loc_loss_type': P.MultiBoxLoss.SMOOTH_L1,
'conf_loss_type': P.MultiBoxLoss.SOFTMAX,
'loc_weight': loc_weight,
'num_classes': num_classes,
'share_location': share_location,
'match_type': P.MultiBoxLoss.PER_PREDICTION,
'overlap_threshold': 0.5,
'use_prior_for_matching': True,
'background_label_id': background_label_id,
'use_difficult_gt': train_on_diff_gt,
'mining_type': mining_type,
'neg_pos_ratio': neg_pos_ratio,
'neg_overlap': 0.5,
'code_type': code_type,
'ignore_cross_boundary_bbox': ignore_cross_boundary_bbox,
}
# parameters for generating priors.
# minimum dimension of input image
min_dim = 320
# conv4_3 ==> 38 x 38
# fc7 ==> 19 x 19
# conv6_2 ==> 10 x 10
# conv7_2 ==> 5 x 5
# conv8_2 ==> 3 x 3
# conv9_2 ==> 1 x 1
mbox_source_layers = [
'conv4_3', 'fc7', 'conv6_2', 'conv7_2', 'conv8_2', 'conv9_2'
]
# in percent %
min_ratio = 20
max_ratio = 90
step = int(
math.floor(
(max_ratio - min_ratio) / (len(mbox_source_layers) - 2)))
min_sizes = []
max_sizes = []
for ratio in xrange(min_ratio, max_ratio + 1, step):
min_sizes.append(min_dim * ratio / 100.)
max_sizes.append(min_dim * (ratio + step) / 100.)
min_sizes = [min_dim * 10 / 100.] + min_sizes
max_sizes = [min_dim * 20 / 100.] + max_sizes
steps = [8, 16, 32, 64, 100, 320]
aspect_ratios = [[2], [2, 3], [2, 3], [2, 3], [2], [2]]
# L2 normalize conv4_3.
normalizations = [20, -1, -1, -1, -1, -1]
# variance used to encode/decode prior bboxes.
if code_type == P.PriorBox.CENTER_SIZE:
prior_variance = [0.1, 0.1, 0.2, 0.2]
else:
prior_variance = [0.1]
flip = True
clip = False
# parameters for generating detection output.
det_out_param = {
'num_classes': num_classes,
'share_location': share_location,
'background_label_id': background_label_id,
'nms_param': {
'nms_threshold': 0.45,
'top_k': 400
},
'keep_top_k': 200,
'confidence_threshold': 0.01,
'code_type': code_type,
}
mbox_layers = CreateMultiBoxHead(net,
data_layer='data',
from_layers=mbox_source_layers,
use_batchnorm=use_batchnorm,
min_sizes=min_sizes,
max_sizes=max_sizes,
aspect_ratios=aspect_ratios,
steps=steps,
normalizations=normalizations,
num_classes=num_classes,
share_location=share_location,
flip=flip,
clip=clip,
prior_variance=prior_variance,
kernel_size=3,
pad=1,
lr_mult=lr_mult)
conf_name = "mbox_conf"
if multibox_loss_param["conf_loss_type"] == P.MultiBoxLoss.SOFTMAX:
reshape_name = "{}_reshape".format(conf_name)
net[reshape_name] = L.Reshape(net[conf_name],
shape=dict(dim=[0, -1, num_classes]))
softmax_name = "{}_softmax".format(conf_name)
net[softmax_name] = L.Softmax(net[reshape_name], axis=2)
flatten_name = "{}_flatten".format(conf_name)
net[flatten_name] = L.Flatten(net[softmax_name], axis=1)
mbox_layers[1] = net[flatten_name]
elif multibox_loss_param["conf_loss_type"] == P.MultiBoxLoss.LOGISTIC:
sigmoid_name = "{}_sigmoid".format(conf_name)
net[sigmoid_name] = L.Sigmoid(net[conf_name])
mbox_layers[1] = net[sigmoid_name]
net.detection_out = L.DetectionOutput(
*mbox_layers,
detection_output_param=det_out_param,
include=dict(phase=caffe_pb2.Phase.Value('TEST')))
if not is_cls:
if not is_deploy:
net.__setattr__('cls_silence', L.Silence(net.cls, ntop=0))
else:
# class vector embedding deconvolution net for class-specific semantic segmentation
net.cls_reshape = L.Reshape(net.cls, shape=dict(dim=[0, 0, 1, 1]))
# add top-down deconvolution net
# mbox_source_layers = ['conv4_3', 'fc7', 'conv6_2', 'conv7_2', 'conv8_2', 'conv9_2']
AddExtraTopDownLayers(net,
use_batchnorm=True,
lr_mult=1,
crop_layers=crop_layers,
is_cls=is_cls,
is_crop_all=is_crop_all,
is_crop_cls=is_crop_cls)
DeVGGNetBody(net,
from_layer='deconv6_1',
fully_conv=True,
reduced=True,
dilated=True,
dropout=False,
pool_mask=True,
extra_crop_layers=crop_layers,
is_crop_all=is_crop_all,
is_crop_cls=is_crop_cls,
is_crop_merge_feature=is_crop_merge_feature)
dekwargs = {
'weight_filler': dict(type='xavier'),
'bias_filler': dict(type='constant', value=0)
}
deparam = {
'param':
[dict(lr_mult=1, decay_mult=1),
dict(lr_mult=2, decay_mult=0)]
}
net.seg_score = L.Deconvolution(net.derelu1_1,
convolution_param=dict(num_output=2,
pad=1,
kernel_size=3,
**dekwargs),
**deparam)
if is_crop_last:
out_layer = "seg_score_crop"
net[out_layer] = L.CropBBox(net["seg_score"],
net["cls_specific_bbox"],
is_crop_score_map=True)
else:
out_layer = "seg_score"
if is_deploy:
net.seg_prob = L.Softmax(net[out_layer])
else:
net.seg_loss = L.SoftmaxWithLoss(net[out_layer],
net.binary_mask,
loss_param=dict(ignore_label=255))
return net.to_proto()