def fcn_top_1(feat, classifier1, fc_name, classifier2, fc_name2, classes, bootstrapping=False):
crop_size = 224
top = feat
for j, layer in enumerate(classifier1[:-1]):
# This naming (conv6) is derived from the ResNets (with five levels),
# which is not accurate for our networks (with seven levels).
top = conv_stage_v1(top, 'conv6{}'.format(chr(j+97)),
layer.channels,
kernel=layer.kernel,
dilate=layer.dilate,
dropout_rate=0.)
layer = classifier1[-1]
scores = conv(top, fc_name,
layer.channels,
kernel=layer.kernel,
dilate=layer.dilate)
print 'Scores'
print scores.infer_shape(data=(64, 3, crop_size, crop_size))[1]
top2 = feat
for j, layer in enumerate(classifier2[:-1]):
# This naming (conv6) is derived from the ResNets (with five levels),
# which is not accurate for our networks (with seven levels).
top2 = conv_stage_v1(top2, 'conv6{}'.format(chr(j+97)),
layer.channels,
kernel=layer.kernel,
dilate=layer.dilate,
dropout_rate=0.)
layer2 = classifier2[-1]
scores2 = conv(top2, fc_name2,
layer2.channels,
kernel=layer2.kernel,
dilate=layer2.dilate)
print 'Scores'
#print scores.infer_shape(data=(64, 3, crop_size, crop_size))[1]
classifier21 = rn_top_1(scores2, 'linear{}'.format(classes), classes)
if not bootstrapping:
return softmax_out(scores, multi_output=True), softmax_out(scores)
else:
from layer import OhemSoftmax, OhemSoftmaxProp
return mx.sym.Custom(data=scores, name='softmax',
op_type='ohem_softmax',
ignore_label=255,
# ignore_label=65,
thresh=0.6,
min_kept=256,
margin=-1), softmax_out(scores)
def fcn_top(feat, classifier, fc_name, bootstrapping=False):
crop_size = 224
top = feat
for j, layer in enumerate(classifier[:-1]):
# This naming (conv6) is derived from the ResNets (with five levels),
# which is not accurate for our networks (with seven levels).
top = conv_stage_v1(top, 'conv6{}'.format(chr(j+97)),
layer.channels,
kernel=layer.kernel,
dilate=layer.dilate,
dropout_rate=0.)
layer = classifier[-1]
scores = conv(top, fc_name,
layer.channels,
kernel=layer.kernel,
dilate=layer.dilate)
print 'Scores'
print scores.infer_shape(data=(64, 3, crop_size, crop_size))[1]
if not bootstrapping:
return softmax_out(scores, multi_output=True)
else:
from layer import OhemSoftmax, OhemSoftmaxProp
return mx.sym.Custom(data=scores, name='softmax',
op_type='ohem_softmax',
ignore_label=255,
thresh=0.6,
min_kept=256,
margin=-1)
def fcn_top(feat, classifier, fc_name):
top = feat
for j, layer in enumerate(classifier[:-1]):
top = conv_state_v1(top,
'conv6{}'.format(chr(j + 97)),
layer.channels,
kernel=layer.kernel,
dilate=layer.dilate,
dropout_rate=0.)
layer = classifier[-1]
scores = conv(top,
fc_name,
layer.channels,
kernel=layer.kernel,
dilate=layer.dilate)
return softmax_out(scores, multi_output=True)
def fcn_top(feat, classifier, fc_name):
top = feat
for j, layer in enumerate(classifier[:-1]):
# This naming (conv6) is derived from the ResNets (with five levels),
# which is not accurate for our networks (with seven levels).
top = conv_state_v1(top,
'conv6{}'.format(chr(j + 97)),
layer.channels,
kernel=layer.kernel,
dilate=layer.dilate,
dropout_rate=0.)
layer = classifier[-1]
scores = conv(top,
fc_name,
layer.channels,
kernel=layer.kernel,
dilate=layer.dilate)
return softmax_out(scores, multi_output=True)
def rn_top(feat, fc_name, classes):
pool7 = pool(feat, 'pool7', pool_type='avg', global_pool=True)
scores = fc(pool7, fc_name, classes)
return softmax_out(scores)
def rn_top(feat, fc_name, classes):
pool7 = pool(feat, 'pool7', pool_type='avg', global_pool=True)
scores = fc(pool7, fc_name, classes)
return softmax_out(scores)
def fcn_top_1(feat,
classifier1,
fc_name,
classifier2,
fc_name2,
classes,
bootstrapping=False):
crop_size = 224
top = feat
for j, layer in enumerate(classifier1[:-1]):
# This naming (conv6) is derived from the ResNets (with five levels),
# which is not accurate for our networks (with seven levels).
top = conv_stage_v1(top,
'conv6{}'.format(chr(j + 97)),
layer.channels,
kernel=layer.kernel,
dilate=layer.dilate,
dropout_rate=0.)
layer = classifier1[-1]
scores = conv(top,
fc_name,
layer.channels,
kernel=layer.kernel,
dilate=layer.dilate)
print 'Scores'
print scores.infer_shape(data=(64, 3, crop_size, crop_size))[1]
top2 = feat
for j, layer2 in enumerate(classifier2[:-1]):
# This naming (conv6) is derived from the ResNets (with five levels),
# which is not accurate for our networks (with seven levels).
top2 = conv_stage_v1(top2,
'conv61{}'.format(chr(j + 97)),
layer2.channels,
kernel=layer2.kernel,
dilate=layer2.dilate,
dropout_rate=0.)
layer2 = classifier2[-1]
scores2 = conv(top2,
fc_name2,
layer2.channels,
kernel=layer2.kernel,
dilate=layer2.dilate)
print 'Scores'
print scores2.infer_shape(data=(64, 3, crop_size, crop_size))[1]
#print ('I am here')
classifier21 = rn_top_1(scores2, 'bCls0', classes)
print classifier21.infer_shape(data=(64, 3, crop_size, crop_size))[1]
#mx.symbol.Group([sm1, sm2])
#print 'I am here'
print bootstrapping, classes, classifier21
if not bootstrapping:
#print 'I am here'
return mx.symbol.Group(
[softmax_out(scores, multi_output=True), classifier21])
else:
print 'I am here'
from layer import OhemSoftmax, OhemSoftmaxProp
return mx.symbol.Group([
mx.sym.Custom(
data=scores,
name='softmax',
op_type='ohem_softmax',
ignore_label=255,
# ignore_label=65,
thresh=0.6,
min_kept=256,
margin=-1),
classifier21
])
#def fcrna_model_a1_1(classes, inv_resolution=8, bootstrapping=False):
'''FCRNA Model A1_1'''
'''feat = rna_feat_a1(inv_resolution, dropout=True)