def inceptionB(input_layer, name, nfilt):
l1 = conv2dbn(
input_layer, name='%s_inceptB_1_3x3' % name,
num_filters=nfilt[0][0], filter_size=3, stride=2
)
l2 = conv2dbn(
input_layer, name='%s_inceptB_2_1x1' % name,
num_filters=nfilt[1][0], filter_size=1
)
l2 = conv2dbn(
l2, name='%s_inceptB_2_3x3_1' % name,
num_filters=nfilt[1][1], filter_size=3, pad=1
)
l2 = conv2dbn(
l2, name='%s_inceptB_2_3x3_2' % name,
num_filters=nfilt[1][2], filter_size=3, stride=2
)
l3 = avg_pool(
input_layer, name='%s_inceptB_3p' % name,
pool_size=3, stride=2,
)
return nn.layers.ConcatLayer(
[l1, l2, l3], name='%s_inceptB_concat' % name
)
def inceptionB(input_layer, name, nfilt):
l1 = conv2dbn(input_layer,
name='%s_inceptB_1_3x3' % name,
num_filters=nfilt[0][0],
filter_size=3,
stride=2)
l2 = conv2dbn(input_layer,
name='%s_inceptB_2_1x1' % name,
num_filters=nfilt[1][0],
filter_size=1)
l2 = conv2dbn(l2,
name='%s_inceptB_2_3x3_1' % name,
num_filters=nfilt[1][1],
filter_size=3,
pad=1)
l2 = conv2dbn(l2,
name='%s_inceptB_2_3x3_2' % name,
num_filters=nfilt[1][2],
filter_size=3,
stride=2)
l3 = avg_pool(
input_layer,
name='%s_inceptB_3p' % name,
pool_size=3,
stride=2,
)
return nn.layers.ConcatLayer([l1, l2, l3], name='%s_inceptB_concat' % name)
def inceptionC(input_layer, name, nfilt):
l1 = conv2dbn(input_layer,
name='%s_inceptC_1_3x3' % name,
num_filters=nfilt[0][0],
filter_size=1)
l2 = conv2dbn(input_layer,
name='%s_inceptC_2_3x3' % name,
num_filters=nfilt[1][0],
filter_size=1)
l2 = conv2dbn(l2,
name='%s_inceptC_2_1x7' % name,
num_filters=nfilt[1][1],
filter_size=(1, 7),
pad=(0, 3))
l2 = conv2dbn(l2,
name='%s_inceptC_2_7x1' % name,
num_filters=nfilt[1][2],
filter_size=(7, 1),
pad=(3, 0))
l3 = conv2dbn(input_layer,
name='%s_inceptC_3_1x1' % name,
num_filters=nfilt[2][0],
filter_size=1)
l3 = conv2dbn(l3,
name='%s_inceptC_3_7x1_1' % name,
num_filters=nfilt[2][1],
filter_size=(7, 1),
pad=(3, 0))
l3 = conv2dbn(l3,
name='%s_inceptC_3_1x7_1' % name,
num_filters=nfilt[2][2],
filter_size=(1, 7),
pad=(0, 3))
l3 = conv2dbn(l3,
name='%s_inceptC_3_7x1_2' % name,
num_filters=nfilt[2][3],
filter_size=(7, 1),
pad=(3, 0))
l3 = conv2dbn(l3,
name='%s_inceptC_3_1x7_2' % name,
num_filters=nfilt[2][4],
filter_size=(1, 7),
pad=(0, 3))
l4 = avg_pool(input_layer,
name='%s_inceptC_4p' % name,
pool_size=3,
stride=1,
pad=1)
l4 = conv2dbn(l4,
name='%s_inceptC_4_1x1' % name,
num_filters=nfilt[3][0],
filter_size=1)
return nn.layers.ConcatLayer([l1, l2, l3, l4],
name='%s_inceptC_concat' % name)
def inceptionC(input_layer, name, nfilt):
l1 = conv2dbn(
input_layer, name='%s_inceptC_1_3x3' % name,
num_filters=nfilt[0][0], filter_size=1
)
l2 = conv2dbn(
input_layer, name='%s_inceptC_2_3x3' % name,
num_filters=nfilt[1][0], filter_size=1
)
l2 = conv2dbn(
l2, name='%s_inceptC_2_1x7' % name,
num_filters=nfilt[1][1], filter_size=(1, 7), pad=(0, 3)
)
l2 = conv2dbn(
l2, name='%s_inceptC_2_7x1' % name,
num_filters=nfilt[1][2], filter_size=(7, 1), pad=(3, 0)
)
l3 = conv2dbn(
input_layer, name='%s_inceptC_3_1x1' % name,
num_filters=nfilt[2][0], filter_size=1
)
l3 = conv2dbn(
l3, name='%s_inceptC_3_7x1_1' % name,
num_filters=nfilt[2][1], filter_size=(7, 1), pad=(3, 0)
)
l3 = conv2dbn(
l3, name='%s_inceptC_3_1x7_1' % name,
num_filters=nfilt[2][2], filter_size=(1, 7), pad=(0, 3)
)
l3 = conv2dbn(
l3, name='%s_inceptC_3_7x1_2' % name,
num_filters=nfilt[2][3], filter_size=(7, 1), pad=(3, 0)
)
l3 = conv2dbn(
l3, name='%s_inceptC_3_1x7_2' % name,
num_filters=nfilt[2][4], filter_size=(1, 7), pad=(0, 3)
)
l4 = avg_pool(
input_layer, name='%s_inceptC_4p' % name,
pool_size=3, stride=1, pad=1
)
l4 = conv2dbn(
l4, name='%s_inceptC_4_1x1' % name,
num_filters=nfilt[3][0], filter_size=1
)
return nn.layers.ConcatLayer(
[l1, l2, l3, l4], name='%s_inceptC_concat' % name
)
def inceptionE(input_layer, name, nfilt, pool_type):
l1 = conv2dbn(
input_layer, name='%s_inceptE_1_1x1' % name,
num_filters=nfilt[0][0], filter_size=1
)
l2 = conv2dbn(
input_layer, name='%s_inceptE_2_1x1' % name,
num_filters=nfilt[1][0], filter_size=1
)
l2a = conv2dbn(
l2, name='%s_inceptE_2a_1x3' % name,
num_filters=nfilt[1][1], filter_size=(1, 3), pad=(0, 1)
)
l2b = conv2dbn(
l2, name='%s_inceptE_2b_3x1' % name,
num_filters=nfilt[1][2], filter_size=(3, 1), pad=(1, 0)
)
l3 = conv2dbn(
input_layer, name='%s_inceptE_3_1x1_1' % name,
num_filters=nfilt[2][0], filter_size=1
)
l3 = conv2dbn(
l3, name='%s_inceptE_3_1x1_2' % name,
num_filters=nfilt[2][1], filter_size=3, pad=1
)
l3a = conv2dbn(
l3, name='%s_inceptE_3a_1x3' % name,
num_filters=nfilt[2][2], filter_size=(1, 3), pad=(0, 1)
)
l3b = conv2dbn(
l3, name='%s_inceptE_3b_3x1' % name,
num_filters=nfilt[2][3], filter_size=(3, 1), pad=(1, 0)
)
if pool_type == 'avg':
l4 = avg_pool(
input_layer, name='%s_inceptE_4p' % name,
pool_size=3, stride=1, pad=1
)
elif pool_type == 'max':
l4 = nn.layers.dnn.Pool2DDNNLayer(
input_layer, name='%s_inceptE_4p' % name,
pool_size=3, stride=1, pad=1, mode='max'
)
else:
raise ValueError('unrecognized pool_type')
l4 = conv2dbn(
l4, name='%s_inceptE_4_1x1' % name,
num_filters=nfilt[3][0], filter_size=1
)
return nn.layers.ConcatLayer(
[l1, l2a, l2b, l3a, l3b, l4], name='%s_inceptE_concat' % name
)
def inceptionD(input_layer, name, nfilt):
l1 = conv2dbn(
input_layer, name='%s_inceptD_1_1x1' % name,
num_filters=nfilt[0][0], filter_size=1
)
l1 = conv2dbn(
l1, name='%s_inceptD_1_3x3' % name,
num_filters=nfilt[0][1], filter_size=3, stride=2
)
l2 = conv2dbn(
input_layer, name='%s_inceptD_2_1x1' % name,
num_filters=nfilt[1][0], filter_size=1
)
l2 = conv2dbn(
l2, name='%s_inceptD_2_1x7' % name,
num_filters=nfilt[1][1], filter_size=(1, 7), pad=(0, 3)
)
l2 = conv2dbn(
l2, name='%s_inceptD_2_7x1' % name,
num_filters=nfilt[1][2], filter_size=(7, 1), pad=(3, 0)
)
l2 = conv2dbn(
l2, name='%s_inceptD_2_3x3' % name,
num_filters=nfilt[1][3], filter_size=3, stride=2
)
l3 = nn.layers.dnn.Pool2DDNNLayer(
input_layer, name='%s_inceptD_3p' % name,
pool_size=3, stride=2, mode='max'
)
return nn.layers.ConcatLayer(
[l1, l2, l3], name='%s_inceptD_concat' % name
)
def inceptionA(input_layer, name, nfilt):
l1 = conv2dbn(input_layer,
name='%s_inceptA_1_1x1' % name,
num_filters=nfilt[0][0],
filter_size=1)
l2 = conv2dbn(input_layer,
name='%s_inceptA_2_1x1' % name,
num_filters=nfilt[1][0],
filter_size=1)
l2 = conv2dbn(l2,
name='%s_inceptA_2_5x5' % name,
num_filters=nfilt[1][1],
filter_size=5,
pad=2)
l3 = conv2dbn(input_layer,
name='%s_inceptA_3_1x1' % name,
num_filters=nfilt[2][0],
filter_size=1)
l3 = conv2dbn(l3,
name='%s_inceptA_3_3x3_1' % name,
num_filters=nfilt[2][1],
filter_size=3,
pad=1)
l3 = conv2dbn(l3,
name='%s_inceptA_3_3x3_2' % name,
num_filters=nfilt[2][2],
filter_size=3,
pad=1)
l4 = avg_pool(
input_layer,
name='%s_inceptA_4p' % name,
pool_size=3,
stride=1,
pad=1,
)
l4 = conv2dbn(l4,
name='%s_inceptA_4_1x1' % name,
num_filters=nfilt[3][0],
filter_size=1)
return nn.layers.ConcatLayer([l1, l2, l3, l4],
name='%s_inceptA_concat' % name)
def inceptionA(input_layer, name, nfilt):
l1 = conv2dbn(
input_layer, name='%s_inceptA_1_1x1' % name,
num_filters=nfilt[0][0], filter_size=1
)
l2 = conv2dbn(
input_layer, name='%s_inceptA_2_1x1' % name,
num_filters=nfilt[1][0], filter_size=1
)
l2 = conv2dbn(
l2, name='%s_inceptA_2_5x5' % name,
num_filters=nfilt[1][1], filter_size=5, pad=2
)
l3 = conv2dbn(
input_layer, name='%s_inceptA_3_1x1' % name,
num_filters=nfilt[2][0], filter_size=1
)
l3 = conv2dbn(
l3, name='%s_inceptA_3_3x3_1' % name,
num_filters=nfilt[2][1], filter_size=3, pad=1
)
l3 = conv2dbn(
l3, name='%s_inceptA_3_3x3_2' % name,
num_filters=nfilt[2][2], filter_size=3, pad=1
)
l4 = avg_pool(
input_layer, name='%s_inceptA_4p' % name,
pool_size=3, stride=1, pad=1,
)
l4 = conv2dbn(
l4, name='%s_inceptA_4_1x1' % name,
num_filters=nfilt[3][0], filter_size=1
)
return nn.layers.ConcatLayer(
[l1, l2, l3, l4], name='%s_inceptA_concat' % name
)
def inceptionD(input_layer, name, nfilt):
l1 = conv2dbn(input_layer,
name='%s_inceptD_1_1x1' % name,
num_filters=nfilt[0][0],
filter_size=1)
l1 = conv2dbn(l1,
name='%s_inceptD_1_3x3' % name,
num_filters=nfilt[0][1],
filter_size=3,
stride=2)
l2 = conv2dbn(input_layer,
name='%s_inceptD_2_1x1' % name,
num_filters=nfilt[1][0],
filter_size=1)
l2 = conv2dbn(l2,
name='%s_inceptD_2_1x7' % name,
num_filters=nfilt[1][1],
filter_size=(1, 7),
pad=(0, 3))
l2 = conv2dbn(l2,
name='%s_inceptD_2_7x1' % name,
num_filters=nfilt[1][2],
filter_size=(7, 1),
pad=(3, 0))
l2 = conv2dbn(l2,
name='%s_inceptD_2_3x3' % name,
num_filters=nfilt[1][3],
filter_size=3,
stride=2)
l3 = nn.layers.dnn.Pool2DDNNLayer(input_layer,
name='%s_inceptD_3p' % name,
pool_size=3,
stride=2,
mode='max')
return nn.layers.ConcatLayer([l1, l2, l3], name='%s_inceptD_concat' % name)
conv_kwargs = dict(
pad='same',
nonlinearity=nn.nonlinearities.very_leaky_rectify
)
pool_kwargs = dict(
pool_size=2,
)
# 256
l = nn.layers.InputLayer(name='in', shape=(None, 3, image_size, image_size))
l = conv2dbn(l, name='l1c1', num_filters=16, filter_size=(7, 7), **conv_kwargs)
l = nn.layers.dnn.MaxPool2DDNNLayer(l, name='l1p', **pool_kwargs)
# 128
l = conv2dbn(l, name='l2c1', num_filters=32, filter_size=(3, 3), **conv_kwargs)
l = conv2dbn(l, name='l2c2', num_filters=32, filter_size=(3, 3), **conv_kwargs)
l = conv2dbn(l, name='l2c3', num_filters=32, filter_size=(3, 3), **conv_kwargs)
# 64
l = conv2dbn(l, name='l3c1', num_filters=64, filter_size=(3, 3), stride=2, **conv_kwargs)
l = conv2dbn(l, name='l3c2', num_filters=64, filter_size=(3, 3), **conv_kwargs)
l = conv2dbn(l, name='l3c3', num_filters=64, filter_size=(3, 3), **conv_kwargs)
# 32
l = conv2dbn(l, name='l4c1', num_filters=128, filter_size=(3, 3), stride=2, **conv_kwargs)
l = conv2dbn(l, name='l4c2', num_filters=128, filter_size=(3, 3), **conv_kwargs)
save_weights = SaveWeights(model_fname, only_best=True, pickle=False)
save_training_history = SaveTrainingHistory(model_history_fname)
plot_training_history = PlotTrainingHistory(model_graph_fname)
early_stopping = EarlyStopping(patience=100)
conv_kwargs = dict(pad='same',
nonlinearity=nn.nonlinearities.very_leaky_rectify)
pool_kwargs = dict(pool_size=2, )
l = nn.layers.InputLayer(name='in', shape=(None, 3, image_size, image_size))
# 256
l = conv2dbn(l,
name='l1c1',
num_filters=32,
filter_size=(7, 7),
stride=2,
**conv_kwargs)
l = conv2dbn(l, name='l2c1', num_filters=48, filter_size=(3, 3), **conv_kwargs)
l = conv2dbn(l, name='l2c2', num_filters=48, filter_size=(3, 3), **conv_kwargs)
l = conv2dbn(l, name='l2c3', num_filters=48, filter_size=(3, 3), **conv_kwargs)
# 128
l = conv2dbn(l,
name='l3c1',
num_filters=64,
filter_size=(3, 3),
stride=2,
**conv_kwargs)
l = conv2dbn(l, name='l3c2', num_filters=64, filter_size=(3, 3), **conv_kwargs)
save_weights = SaveWeights(model_fname, only_best=True, pickle=False)
save_training_history = SaveTrainingHistory(model_history_fname)
plot_training_history = PlotTrainingHistory(model_graph_fname)
early_stopping = EarlyStopping(patience=100)
conv_kwargs = dict(
pad='same',
nonlinearity=nn.nonlinearities.very_leaky_rectify,
)
l = nn.layers.InputLayer(name='in', shape=(None, 3, image_size, image_size))
# 256x256
l = conv2dbn(l,
name='l1c1',
num_filters=32,
filter_size=(7, 7),
stride=2,
**conv_kwargs)
# 128x128
l = nn.layers.dnn.MaxPool2DDNNLayer(l, name='l1p', pool_size=(3, 3), stride=2)
# 64x64
for i in range(3):
l = residual_block(
l,
name='2c%s' % i,
# bottleneck=True, bottleneck_factor=4,
num_filters=48,
filter_size=(3, 3),
num_layers=2,
save_weights = SaveWeights(model_fname, only_best=True, pickle=False)
save_training_history = SaveTrainingHistory(model_history_fname)
plot_training_history = PlotTrainingHistory(model_graph_fname)
early_stopping = EarlyStopping(patience=100)
conv_kwargs = dict(
pad='same',
nonlinearity=nn.nonlinearities.very_leaky_rectify,
)
l = nn.layers.InputLayer(name='in', shape=(None, 3, image_size, image_size))
# 256x256
l = conv2dbn(
l, name='l1c1', num_filters=32, filter_size=(7, 7), stride=2,
**conv_kwargs
)
# 128x128
l = nn.layers.dnn.MaxPool2DDNNLayer(l, name='l1p', pool_size=(3, 3), stride=2)
# 64x64
for i in range(3):
l = residual_block(
l, name='2c%s' % i,
# bottleneck=True, bottleneck_factor=4,
num_filters=48, filter_size=(3, 3),
num_layers=2,
**conv_kwargs
)
# 64x64
test_iterator = TestIterator(**test_iterator_kwargs)
save_weights = SaveWeights(model_fname, only_best=True, pickle=False)
save_training_history = SaveTrainingHistory(model_history_fname)
plot_training_history = PlotTrainingHistory(model_graph_fname)
early_stopping = EarlyStopping(patience=100)
conv_kwargs = dict(
pad='same',
nonlinearity=nn.nonlinearities.very_leaky_rectify,
W=nn.init.GlorotNormal(gain=1 / 3.0),
)
l = nn.layers.InputLayer(name='in', shape=(None, 3, image_size, image_size))
l = conv2dbn(l, name='1c1', num_filters=32, filter_size=3, stride=2)
l = conv2dbn(l, name='1c2', num_filters=32, filter_size=3)
l = conv2dbn(l, name='1c3', num_filters=64, filter_size=3, pad=1)
l = nn.layers.dnn.MaxPool2DDNNLayer(l, name='1p', pool_size=3, stride=2)
l = nn.layers.DropoutLayer(l, name='1cdrop', p=0.1)
l = conv2dbn(l, name='2c1', num_filters=80, filter_size=1)
l = conv2dbn(l, name='2c2', num_filters=192, filter_size=3)
l = nn.layers.dnn.MaxPool2DDNNLayer(l, name='2p', pool_size=3, stride=2)
l = nn.layers.DropoutLayer(l, name='2cdrop', p=0.1)
l = inceptionA(l, name='3', nfilt=((64, ), (48, 64), (64, 96, 96), (32, )))
l = nn.layers.DropoutLayer(l, name='3cdrop', p=0.1)
l = inceptionA(l, name='4', nfilt=((64, ), (48, 64), (64, 96, 96), (64, )))
l = nn.layers.DropoutLayer(l, name='4cdrop', p=0.1)
test_iterator = TestIterator(**test_iterator_kwargs)
save_weights = SaveWeights(model_fname, only_best=True, pickle=False)
save_training_history = SaveTrainingHistory(model_history_fname)
plot_training_history = PlotTrainingHistory(model_graph_fname)
early_stopping = EarlyStopping(patience=100)
conv_kwargs = dict(
pad='same',
nonlinearity=nn.nonlinearities.very_leaky_rectify,
W=nn.init.GlorotNormal(gain=1 / 3.0),
)
l = nn.layers.InputLayer(name='in', shape=(None, 3, image_size, image_size))
l = conv2dbn(l, name='1c1', num_filters=32, filter_size=3, stride=2)
l = conv2dbn(l, name='1c2', num_filters=32, filter_size=3)
l = conv2dbn(l, name='1c3', num_filters=64, filter_size=3, pad=1)
l = nn.layers.dnn.MaxPool2DDNNLayer(l, name='1p', pool_size=3, stride=2)
l = nn.layers.DropoutLayer(l, name='1cdrop', p=0.1)
l = conv2dbn(l, name='2c1', num_filters=80, filter_size=1)
l = conv2dbn(l, name='2c2', num_filters=192, filter_size=3)
l = nn.layers.dnn.MaxPool2DDNNLayer(l, name='2p', pool_size=3, stride=2)
l = nn.layers.DropoutLayer(l, name='2cdrop', p=0.1)
l = inceptionA(
l, name='3', nfilt=(
(64,),
(48, 64),
(64, 96, 96),
def inceptionE(input_layer, name, nfilt, pool_type):
l1 = conv2dbn(input_layer,
name='%s_inceptE_1_1x1' % name,
num_filters=nfilt[0][0],
filter_size=1)
l2 = conv2dbn(input_layer,
name='%s_inceptE_2_1x1' % name,
num_filters=nfilt[1][0],
filter_size=1)
l2a = conv2dbn(l2,
name='%s_inceptE_2a_1x3' % name,
num_filters=nfilt[1][1],
filter_size=(1, 3),
pad=(0, 1))
l2b = conv2dbn(l2,
name='%s_inceptE_2b_3x1' % name,
num_filters=nfilt[1][2],
filter_size=(3, 1),
pad=(1, 0))
l3 = conv2dbn(input_layer,
name='%s_inceptE_3_1x1_1' % name,
num_filters=nfilt[2][0],
filter_size=1)
l3 = conv2dbn(l3,
name='%s_inceptE_3_1x1_2' % name,
num_filters=nfilt[2][1],
filter_size=3,
pad=1)
l3a = conv2dbn(l3,
name='%s_inceptE_3a_1x3' % name,
num_filters=nfilt[2][2],
filter_size=(1, 3),
pad=(0, 1))
l3b = conv2dbn(l3,
name='%s_inceptE_3b_3x1' % name,
num_filters=nfilt[2][3],
filter_size=(3, 1),
pad=(1, 0))
if pool_type == 'avg':
l4 = avg_pool(input_layer,
name='%s_inceptE_4p' % name,
pool_size=3,
stride=1,
pad=1)
elif pool_type == 'max':
l4 = nn.layers.dnn.Pool2DDNNLayer(input_layer,
name='%s_inceptE_4p' % name,
pool_size=3,
stride=1,
pad=1,
mode='max')
else:
raise ValueError('unrecognized pool_type')
l4 = conv2dbn(l4,
name='%s_inceptE_4_1x1' % name,
num_filters=nfilt[3][0],
filter_size=1)
return nn.layers.ConcatLayer([l1, l2a, l2b, l3a, l3b, l4],
name='%s_inceptE_concat' % name)
