def get_test_symbol(self, num_classes, cfg):
"""
get symbol for training
:param num_classes: num of classes
:return: the symbol for training
"""
# shared convolutional symbols
data_shape = (cfg.TEST.BATCH_IMAGES, 3, cfg.SCALES[0][0],
cfg.SCALES[0][1])
data = mx.sym.Variable('data', shape=data_shape)
seg_cls_gt = mx.symbol.Variable(name='label')
# shared convolutional symbols
P_feature = resnet38_v5_dcn.get_conv_feature(
data,
is_train=False,
workspace=self.workspace,
fix_gamma=self.fix_gamma,
use_global_stats=self.use_global_stats,
eps=self.eps)
rnn_feature_list = []
scale_name = ['a', 'b', 'c']
if cfg.network.scale_list == [1, 2, 4]:
scale_name = ['', '', '']
if cfg.network.use_weight:
self.num_hidden = 512
for idx, i in enumerate(cfg.network.scale_list):
if cfg.network.use_weight:
num_hidden = self.num_hidden / i
else:
num_hidden = self.num_hidden
print "num_hidden", num_hidden
rnn_feature = RnnMap_cudnn_v2(P_feature[3],
name="RNNRelation_" + str(i) +
scale_name[idx],
type=self.rnn_type,
PatchSize=self.patch_size,
num_hidden=num_hidden,
use_memory=True,
old_type=False,
skip_step=i,
parallel=True)
rnn_feature_list.append(rnn_feature)
rnn_feature = mx.symbol.concat(*rnn_feature_list)
fcn_fusion = self.get_fcn_top(rnn_feature,
seg_cls_gt,
'FUSION',
num_classes=num_classes,
cfg=cfg,
is_train=False)
return mx.symbol.Group([fcn_fusion, rnn_feature])
def get_train_symbol(self, num_classes, cfg):
"""
get symbol for training
:param num_classes: num of classes
:return: the symbol for training
"""
if cfg.TRAIN.enable_crop:
data_shape = (cfg.TRAIN.BATCH_IMAGES, 3, cfg.TRAIN.crop_size[0], cfg.TRAIN.crop_size[1])
else:
data_shape = (cfg.TRAIN.BATCH_IMAGES, 3, cfg.SCALES[0][0], cfg.SCALES[0][1])
data = mx.sym.Variable('data', shape=data_shape)
seg_cls_gt = mx.symbol.Variable(name='label')
num_instance = data_shape[0] * data_shape[2] * data_shape[3]
# shared convolutional symbols
array = cfg.network.use_mult_label_weight
P_feature = resnet38_v5_dcn.get_conv_feature(data, is_train=True, workspace=self.workspace,
fix_gamma=self.fix_gamma, use_global_stats=self.use_global_stats,
eps=self.eps)
rnn_feature_list = []
metric_top_list = []
scale_name = ['a','b','c']
if cfg.network.scale_list == [1,2,4]:
scale_name=['','','']
metric_grad_scale = float(array[1]) / num_instance / 9
print "metric grad scale", metric_grad_scale
if cfg.network.use_weight:
self.num_hidden= 512
for idx, i in enumerate(cfg.network.scale_list):
if cfg.network.use_weight:
num_hidden = self.num_hidden/i
else:
num_hidden = self.num_hidden
print "num_hidden", num_hidden
rnn_feature = RnnMap_cudnn_v2(P_feature[3], name="RNNRelation_"+str(i)+scale_name[idx], type=self.rnn_type, PatchSize=self.patch_size,
num_hidden=num_hidden, use_memory=True,old_type=False,skip_step=i)
metric_gt = mx.symbol.Variable(name='metric_label_'+str(i)+scale_name[idx])
metric_top_list.append(self.get_metric_top(conv_feature=rnn_feature, metric_label=metric_gt,
grad_scale=metric_grad_scale,skip_step=i,scale_name=scale_name[idx]))
rnn_feature_list.append(rnn_feature)
rnn_feature = mx.symbol.concat(*rnn_feature_list)
if cfg.network.use_origin:
rnn_feature = mx.symbol.concat(rnn_feature,P_feature[3])
fcn_fusion = self.get_fcn_top(rnn_feature, seg_cls_gt, 'FUSION', num_classes=num_classes, cfg=cfg,is_train=True,grad_scale=array[0])
loss = [fcn_fusion, metric_top_list[0],metric_top_list[1],metric_top_list[2]]
sym = mx.symbol.Group(loss)
return sym
def get_train_symbol(self, num_classes, cfg):
"""
get symbol for training
:param num_classes: num of classes
:return: the symbol for training
"""
if cfg.TRAIN.enable_crop:
data_shape = (cfg.TRAIN.BATCH_IMAGES, 3, cfg.TRAIN.crop_size[0],
cfg.TRAIN.crop_size[1])
else:
data_shape = (cfg.TRAIN.BATCH_IMAGES, 3, cfg.SCALES[0][0],
cfg.SCALES[0][1])
data = mx.sym.Variable('data', shape=data_shape)
seg_cls_gt = mx.symbol.Variable(name='label')
metric_gt = mx.symbol.Variable(name='metric_label')
array = cfg.network.use_mult_label_weight
num_instance = data_shape[0] * data_shape[2] * data_shape[3]
metric_grad_scale = float(array[1]) / num_instance / 9
# shared convolutional symbols
P_feature = resnet38_v5_dcn.get_conv_feature(
data,
is_train=True,
workspace=self.workspace,
fix_gamma=self.fix_gamma,
use_global_stats=self.use_global_stats,
eps=self.eps)
rnn_feature = RnnMap_cudnn(P_feature[3],
name="RNNRelation",
type=self.rnn_type,
PatchSize=self.patch_size,
num_hidden=self.num_hidden,
use_memory=True)
array = cfg.network.use_mult_label_weight
fcn_fusion, metric_feature = self.get_fcn_top(rnn_feature,
seg_cls_gt,
'FUSION',
num_classes=num_classes,
cfg=cfg,
is_train=True,
grad_scale=array[0])
metric_top = self.get_metric_top(conv_feature=metric_feature,
metric_label=metric_gt,
grad_scale=metric_grad_scale)
sym = mx.symbol.Group([fcn_fusion, metric_top])
return sym
def get_test_symbol(self, num_classes, cfg):
"""
get symbol for training
:param num_classes: num of classes
:return: the symbol for training
"""
# shared convolutional symbols
data_shape = (cfg.TEST.BATCH_IMAGES, 3, cfg.SCALES[0][0],
cfg.SCALES[0][1])
data = mx.sym.Variable('data', shape=data_shape)
seg_cls_gt = mx.symbol.Variable(name='label')
# shared convolutional symbols
P_feature = resnet38_v5_dcn.get_conv_feature(
data,
is_train=True,
workspace=self.workspace,
fix_gamma=self.fix_gamma,
use_global_stats=self.use_global_stats,
eps=self.eps)
rnn_feature = RnnMap_cudnn(P_feature[3],
name="RNNRelation",
type=self.rnn_type,
PatchSize=self.patch_size,
num_hidden=self.num_hidden,
use_memory=True)
fcn_fusion, _ = self.get_fcn_top(rnn_feature,
seg_cls_gt,
'FUSION',
num_classes=num_classes,
cfg=cfg,
is_train=False)
return mx.sym.Group([fcn_fusion, P_feature[3], rnn_feature])
