def comResnet(input, d):
activation = d.act
inputShape = (100, 100, 3)
channelAxis = 1 if K.image_data_format() == 'channels_first' else -1
convArgs = {
"padding": "same",
"use_bias": False,
"kernel_regularizer": l2(0.0001)
}
bnArgs = {"axis": channelAxis, "momentum": 0.9, "epsilon": 1e-04}
convArgs.update({
"spectral_parametrization": d.spectral_param,
"kernel_initializer": d.comp_init
})
O = learnConcatRealImagBlock(input, (1, 1), (3, 3), 0, '0', convArgs,
bnArgs, d)
#O = tf.concat([input, O], 1)
O = Concatenate(channelAxis)([input, O])
O = ComplexConv2D(filters=64, kernel_size=9, name='conv1', **convArgs)(O)
O = ComplexBN(name='bn conv1 2a', **bnArgs)(O)
O = tf.nn.relu(O)
# residual
O = getResidualBlock(O, (3, 3), [64, 64], 2, '0', 'regular', convArgs,
bnArgs, d)
O = getResidualBlock(O, (3, 3), [64, 64], 2, '0', 'regular', convArgs,
bnArgs, d)
O = getResidualBlock(O, (3, 3), [64, 64], 2, '0', 'regular', convArgs,
bnArgs, d)
O = getResidualBlock(O, (3, 3), [64, 64], 2, '0', 'regular', convArgs,
bnArgs, d)
O = ComplexConv2D(filters=64, kernel_size=9, name='conv2', **convArgs)(O)
O = ComplexBN(name='bn conv2 1a', **bnArgs)(O)
O = tf.nn.relu(O)
O = ComplexConv2D(filters=64, kernel_size=9, name='conv3', **convArgs)(O)
O = ComplexBN(name='bn conv3 1a', **bnArgs)(O)
O = tf.nn.relu(O)
O = ComplexConv2D(filters=3, kernel_size=9, name='conv4', **convArgs)(O)
O = ComplexBN(name='bn conv4 1a', **bnArgs)(O)
O = tf.nn.tanh(O)
return O
def getResidualBlock(I, filter_size, featmaps, stage, block, shortcut,
convArgs, bnArgs, d):
"""Get residual block."""
activation = d.act
drop_prob = d.dropout
nb_fmaps1, nb_fmaps2 = featmaps
conv_name_base = 'res' + str(stage) + block + '_branch'
bn_name_base = 'bn' + str(stage) + block + '_branch'
if K.image_data_format() == 'channels_first' and K.ndim(I) != 3:
channel_axis = 1
else:
channel_axis = -1
if d.model == "real":
O = BatchNormalization(name=bn_name_base + '_2a', **bnArgs)(I)
elif d.model == "complex":
O = ComplexBN(name=bn_name_base + '_2a', **bnArgs)(I)
O = Activation(activation)(O)
if shortcut == 'regular' or d.spectral_pool_scheme == "nodownsample":
if d.model == "real":
O = Conv2D(nb_fmaps1,
filter_size,
name=conv_name_base + '2a',
**convArgs)(O)
elif d.model == "complex":
O = ComplexConv2D(nb_fmaps1,
filter_size,
name=conv_name_base + '2a',
**convArgs)(O)
elif shortcut == 'projection':
if d.spectral_pool_scheme == "proj":
O = applySpectralPooling(O, d)
if d.model == "real":
O = Conv2D(nb_fmaps1,
filter_size,
name=conv_name_base + '2a',
strides=(2, 2),
**convArgs)(O)
elif d.model == "complex":
O = ComplexConv2D(nb_fmaps1,
filter_size,
name=conv_name_base + '2a',
strides=(2, 2),
**convArgs)(O)
if d.model == "real":
O = BatchNormalization(name=bn_name_base + '_2b', **bnArgs)(O)
O = Activation(activation)(O)
O = Conv2D(nb_fmaps2,
filter_size,
name=conv_name_base + '2b',
**convArgs)(O)
elif d.model == "complex":
O = ComplexBN(name=bn_name_base + '_2b', **bnArgs)(O)
O = Activation(activation)(O)
O = ComplexConv2D(nb_fmaps2,
filter_size,
name=conv_name_base + '2b',
**convArgs)(O)
if shortcut == 'regular':
O = Add()([O, I])
elif shortcut == 'projection':
if d.spectral_pool_scheme == "proj":
I = applySpectralPooling(I, d)
if d.model == "real":
X = Conv2D(
nb_fmaps2, (1, 1),
name=conv_name_base + '1',
strides=(2, 2) if d.spectral_pool_scheme != "nodownsample" else
(1, 1),
**convArgs)(I)
O = Concatenate(channel_axis)([X, O])
elif d.model == "complex":
X = ComplexConv2D(
nb_fmaps2, (1, 1),
name=conv_name_base + '1',
strides=(2, 2) if d.spectral_pool_scheme != "nodownsample" else
(1, 1),
**convArgs)(I)
O_real = Concatenate(channel_axis)([GetReal()(X), GetReal()(O)])
O_imag = Concatenate(channel_axis)([GetImag()(X), GetImag()(O)])
O = Concatenate(1)([O_real, O_imag])
return O
def getResnetModel(d):
n = d.num_blocks
sf = d.start_filter
dataset = d.dataset
activation = d.act
advanced_act = d.aact
drop_prob = d.dropout
inputShape = (3, 32, 32) if K.image_dim_ordering() == "th" else (32, 32, 3)
channelAxis = 1 if K.image_data_format() == 'channels_first' else -1
filsize = (3, 3)
convArgs = {
"padding": "same",
"use_bias": False,
"kernel_regularizer": l2(0.0001),
}
bnArgs = {"axis": channelAxis, "momentum": 0.9, "epsilon": 1e-04}
if d.model == "real":
sf *= 2
convArgs.update({"kernel_initializer": Orthogonal(float(np.sqrt(2)))})
elif d.model == "complex":
convArgs.update({
"spectral_parametrization": d.spectral_param,
"kernel_initializer": d.comp_init
})
#
# Input Layer
#
I = Input(shape=inputShape)
#
# Stage 1
#
O = learnConcatRealImagBlock(I, (1, 1), (3, 3), 0, '0', convArgs, bnArgs,
d)
O = Concatenate(channelAxis)([I, O])
if d.model == "real":
O = Conv2D(sf, filsize, name='conv1', **convArgs)(O)
O = BatchNormalization(name="bn_conv1_2a", **bnArgs)(O)
else:
O = ComplexConv2D(sf, filsize, name='conv1', **convArgs)(O)
O = ComplexBN(name="bn_conv1_2a", **bnArgs)(O)
O = Activation(activation)(O)
#
# Stage 2
#
for i in range(n):
O = getResidualBlock(O, filsize, [sf, sf], 2, str(i), 'regular',
convArgs, bnArgs, d)
if i == n // 2 and d.spectral_pool_scheme == "stagemiddle":
O = applySpectralPooling(O, d)
#
# Stage 3
#
O = getResidualBlock(O, filsize, [sf, sf], 3, '0', 'projection', convArgs,
bnArgs, d)
if d.spectral_pool_scheme == "nodownsample":
O = applySpectralPooling(O, d)
for i in range(n - 1):
O = getResidualBlock(O, filsize, [sf * 2, sf * 2], 3, str(i + 1),
'regular', convArgs, bnArgs, d)
if i == n // 2 and d.spectral_pool_scheme == "stagemiddle":
O = applySpectralPooling(O, d)
#
# Stage 4
#
O = getResidualBlock(O, filsize, [sf * 2, sf * 2], 4, '0', 'projection',
convArgs, bnArgs, d)
if d.spectral_pool_scheme == "nodownsample":
O = applySpectralPooling(O, d)
for i in range(n - 1):
O = getResidualBlock(O, filsize, [sf * 4, sf * 4], 4, str(i + 1),
'regular', convArgs, bnArgs, d)
if i == n // 2 and d.spectral_pool_scheme == "stagemiddle":
O = applySpectralPooling(O, d)
#
# Pooling
#
if d.spectral_pool_scheme == "nodownsample":
O = applySpectralPooling(O, d)
O = AveragePooling2D(pool_size=(32, 32))(O)
else:
O = AveragePooling2D(pool_size=(8, 8))(O)
#
# Flatten
#
O = Flatten()(O)
#
# Dense
#
if dataset == 'cifar10':
O = Dense(10, activation='softmax', kernel_regularizer=l2(0.0001))(O)
elif dataset == 'cifar100':
O = Dense(100, activation='softmax', kernel_regularizer=l2(0.0001))(O)
elif dataset == 'svhn':
O = Dense(10, activation='softmax', kernel_regularizer=l2(0.0001))(O)
else:
raise ValueError("Unknown dataset " + d.dataset)
# Return the model
return Model(I, O)
# model.add(Conv2D(32, kernel_size=(3, 3),
# activation='relu',
# kernel_initializer='he_normal',
# input_shape=input_shape))
model.add(
Conv2D(32,
kernel_size=(3, 3),
kernel_initializer='he_normal',
input_shape=input_shape))
model.add(BatchNormalization(**bnArgs))
model.add(Activation('relu'))
model.add(MaxPooling2D((2, 2)))
model.add(Dropout(0.25))
if COMPLEX_MODE:
# model.add(ComplexConv2D(32, (3, 3)))
model.add(ComplexConv2D(32, (3, 3), **convArgs))
model.add(ComplexBN(**bnArgs))
else:
# model.add(Conv2D(64, (3, 3), activation='relu'))
# model.add(Conv2D(32*2, (3, 3)))
model.add(Conv2D(32 * 2, (3, 3), **convArgs))
model.add(BatchNormalization(**bnArgs))
model.add(Activation('relu'))
# O = ComplexConv2D(sf, filsize, name='conv1', **convArgs)(O)
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Dropout(0.25))
if COMPLEX_MODE:
# model.add(ComplexConv2D(64, (3, 3)))
model.add(ComplexConv2D(64, (3, 3), **convArgs))
model.add(ComplexBN(**bnArgs))
else:
def getResidualBlock(I, filter_size, featmaps, stage, block, shortcut,
convArgs, convArgs_real, bnArgs, d):
"""Get residual block."""
activation = d.act
drop_prob = d.dropout
nb_fmaps1, nb_fmaps2 = featmaps
conv_name_base = 'res' + str(stage) + block + '_branch'
bn_name_base = 'bn' + str(stage) + block + '_branch'
if K.image_data_format() == 'channels_first' and K.ndim(I) != 3:
channel_axis = 1
else:
channel_axis = -1
# if d.model == "real":
if "real" in d.model:
O = BatchNormalization(name=bn_name_base + '_2a', **bnArgs)(I)
# elif d.model == "complex":
elif "complex" in d.model:
O = ComplexBN(name=bn_name_base + '_2a', **bnArgs)(I)
if d.aact == "complex_joint_relu":
O = Lambda(ComplexJointReLU)(O)
else:
O = Activation(activation)(O)
if shortcut == 'regular' or d.spectral_pool_scheme == "nodownsample":
if d.model == "real":
O = Conv2D(nb_fmaps1,
filter_size,
name=conv_name_base + '2a',
**convArgs)(O)
elif d.model == "real_dws":
O = SeparableConv2D(nb_fmaps1,
filter_size,
name=conv_name_base + '2a',
**convArgs)(O)
elif (d.model
== "real_group") or (d.model == "real_group_pwc_full") or (
d.model == "real_group_pwc_group"):
O_g0 = Lambda(lambda O: O[:, :(O.shape[1] // 2), :, :])(O)
O_g1 = Lambda(lambda O: O[:, (O.shape[1] // 2):, :, :])(O)
O_g0 = Conv2D(nb_fmaps1 // 2,
filter_size,
name=conv_name_base + '2a_g0',
**convArgs)(O_g0)
O_g1 = Conv2D(nb_fmaps1 // 2,
filter_size,
name=conv_name_base + '2a_g1',
**convArgs)(O_g1)
O_g00 = Lambda(lambda O_g0: O_g0[:, :(O_g0.shape[1] // 2), :, :])(
O_g0)
O_g01 = Lambda(lambda O_g0: O_g0[:, (O_g0.shape[1] // 2):, :, :])(
O_g0)
O_g10 = Lambda(lambda O_g1: O_g1[:, :(O_g1.shape[1] // 2), :, :])(
O_g1)
O_g11 = Lambda(lambda O_g1: O_g1[:, (O_g1.shape[1] // 2):, :, :])(
O_g1)
O = Concatenate(axis=1)(
[O_g00, O_g11, O_g01, O_g10]
) # This ordering allows permutation of odd-numbered outputs (O_g0, O_g1).
if d.model == "real_group_pwc_full":
O = Conv2D(nb_fmaps1, (1, 1),
name=conv_name_base + '2a_pwc',
**convArgs)(O)
elif d.model == "real_group_pwc_group":
O_g0 = Lambda(lambda O: O[:, :(O.shape[1] // 2), :, :])(O)
O_g1 = Lambda(lambda O: O[:, (O.shape[1] // 2):, :, :])(O)
O_g0 = Conv2D(int(O.shape[1] // 2), (1, 1),
name=conv_name_base + '2a_pwc_g0',
**convArgs_real)(O_g0)
O_g1 = Conv2D(int(O.shape[1] // 2), (1, 1),
name=conv_name_base + '2a_pwc_g1',
**convArgs_real)(O_g1)
O = Concatenate(axis=1)([O_g0, O_g1])
elif d.model == "complex":
O = ComplexConv2D(nb_fmaps1,
filter_size,
name=conv_name_base + '2a',
**convArgs)(O)
elif (d.model == "complex_concat") or (d.model
== "complex_concat_pwc_group"):
O = ComplexConvConcat2D(nb_fmaps1 // 2,
filter_size,
name=conv_name_base + '2a',
**convArgs)(O)
if d.model == "complex_concat_pwc_group":
O_g0 = Lambda(lambda O: O[:, :(O.shape[1] // 2), :, :])(O)
O_g1 = Lambda(lambda O: O[:, (O.shape[1] // 2):, :, :])(O)
O_g0 = Conv2D(int(O.shape[1] // 2), (1, 1),
name=conv_name_base + '2a_pwc_g0',
**convArgs_real)(O_g0)
O_g1 = Conv2D(int(O.shape[1] // 2), (1, 1),
name=conv_name_base + '2a_pwc_g1',
**convArgs_real)(O_g1)
O = Concatenate(axis=1)([O_g0, O_g1])
else:
print("Error: unknown model type")
exit(-1)
elif shortcut == 'projection':
if d.spectral_pool_scheme == "proj":
O = applySpectralPooling(O, d)
if d.model == "real":
O = Conv2D(nb_fmaps1,
filter_size,
name=conv_name_base + '2a',
strides=(2, 2),
**convArgs)(O)
elif d.model == "real_dws":
O = SeparableConv2D(nb_fmaps1,
filter_size,
name=conv_name_base + '2a',
strides=(2, 2),
**convArgs)(O)
elif (d.model
== "real_group") or (d.model == "real_group_pwc_full") or (
d.model == "real_group_pwc_group"):
O_g0 = Lambda(lambda O: O[:, :(O.shape[1] // 2), :, :])(O)
O_g1 = Lambda(lambda O: O[:, (O.shape[1] // 2):, :, :])(O)
O_g0 = Conv2D(nb_fmaps1 // 2,
filter_size,
name=conv_name_base + '2a_g0',
strides=(2, 2),
**convArgs)(O_g0)
O_g1 = Conv2D(nb_fmaps1 // 2,
filter_size,
name=conv_name_base + '2a_g1',
strides=(2, 2),
**convArgs)(O_g1)
O_g00 = Lambda(lambda O_g0: O_g0[:, :(O_g0.shape[1] // 2), :, :])(
O_g0)
O_g01 = Lambda(lambda O_g0: O_g0[:, (O_g0.shape[1] // 2):, :, :])(
O_g0)
O_g10 = Lambda(lambda O_g1: O_g1[:, :(O_g1.shape[1] // 2), :, :])(
O_g1)
O_g11 = Lambda(lambda O_g1: O_g1[:, (O_g1.shape[1] // 2):, :, :])(
O_g1)
O = Concatenate(axis=1)([O_g00, O_g11, O_g01, O_g10])
if d.model == "real_group_pwc_full":
O = Conv2D(nb_fmaps1, (1, 1),
name=conv_name_base + '2a_pwc',
**convArgs)(O)
elif d.model == "real_group_pwc_group":
O_g0 = Lambda(lambda O: O[:, :(O.shape[1] // 2), :, :])(O)
O_g1 = Lambda(lambda O: O[:, (O.shape[1] // 2):, :, :])(O)
O_g0 = Conv2D(int(O.shape[1] // 2), (1, 1),
name=conv_name_base + '2a_pwc_g0',
**convArgs_real)(O_g0)
O_g1 = Conv2D(int(O.shape[1] // 2), (1, 1),
name=conv_name_base + '2a_pwc_g1',
**convArgs_real)(O_g1)
O = Concatenate(axis=1)([O_g0, O_g1])
elif d.model == "complex":
O = ComplexConv2D(nb_fmaps1,
filter_size,
name=conv_name_base + '2a',
strides=(2, 2),
**convArgs)(O)
elif (d.model == "complex_concat") or (d.model
== "complex_concat_pwc_group"):
O = ComplexConvConcat2D(nb_fmaps1 // 2,
filter_size,
name=conv_name_base + '2a',
strides=(2, 2),
**convArgs)(O)
if d.model == "complex_concat_pwc_group":
O_g0 = Lambda(lambda O: O[:, :(O.shape[1] // 2), :, :])(O)
O_g1 = Lambda(lambda O: O[:, (O.shape[1] // 2):, :, :])(O)
O_g0 = Conv2D(int(O.shape[1] // 2), (1, 1),
name=conv_name_base + '2a_pwc_g0',
**convArgs_real)(O_g0)
O_g1 = Conv2D(int(O.shape[1] // 2), (1, 1),
name=conv_name_base + '2a_pwc_g1',
**convArgs_real)(O_g1)
O = Concatenate(axis=1)([O_g0, O_g1])
else:
print("Error: unknown model type")
exit(-1)
if d.model == "real":
O = BatchNormalization(name=bn_name_base + '_2b', **bnArgs)(O)
O = Activation(activation)(O)
O = Conv2D(nb_fmaps2,
filter_size,
name=conv_name_base + '2b',
**convArgs)(O)
elif d.model == "real_dws":
O = BatchNormalization(name=bn_name_base + '_2b', **bnArgs)(O)
O = Activation(activation)(O)
O = SeparableConv2D(nb_fmaps2,
filter_size,
name=conv_name_base + '2b',
**convArgs)(O)
elif (d.model == "real_group") or (d.model == "real_group_pwc_full") or (
d.model == "real_group_pwc_group"):
O = BatchNormalization(name=bn_name_base + '_2b', **bnArgs)(O)
O = Activation(activation)(O)
O_g0 = Lambda(lambda O: O[:, :(O.shape[1] // 2), :, :])(O)
O_g1 = Lambda(lambda O: O[:, (O.shape[1] // 2):, :, :])(O)
O_g0 = Conv2D(nb_fmaps2 // 2,
filter_size,
name=conv_name_base + '2b_g0',
**convArgs)(O_g0)
O_g1 = Conv2D(nb_fmaps2 // 2,
filter_size,
name=conv_name_base + '2b_g1',
**convArgs)(O_g1)
O_g00 = Lambda(lambda O_g0: O_g0[:, :(O_g0.shape[1] // 2), :, :])(O_g0)
O_g01 = Lambda(lambda O_g0: O_g0[:, (O_g0.shape[1] // 2):, :, :])(O_g0)
O_g10 = Lambda(lambda O_g1: O_g1[:, :(O_g1.shape[1] // 2), :, :])(O_g1)
O_g11 = Lambda(lambda O_g1: O_g1[:, (O_g1.shape[1] // 2):, :, :])(O_g1)
O = Concatenate(axis=1)([O_g00, O_g11, O_g01, O_g10])
if d.model == "real_group_pwc_full":
O = Conv2D(nb_fmaps2, (1, 1),
name=conv_name_base + '2b_pwc',
**convArgs)(O)
elif d.model == "real_group_pwc_group":
O_g0 = Lambda(lambda O: O[:, :(O.shape[1] // 2), :, :])(O)
O_g1 = Lambda(lambda O: O[:, (O.shape[1] // 2):, :, :])(O)
O_g0 = Conv2D(int(O.shape[1] // 2), (1, 1),
name=conv_name_base + '2b_pwc_g0',
**convArgs_real)(O_g0)
O_g1 = Conv2D(int(O.shape[1] // 2), (1, 1),
name=conv_name_base + '2b_pwc_g1',
**convArgs_real)(O_g1)
O = Concatenate(axis=1)([O_g0, O_g1])
elif d.model == "complex":
O = ComplexBN(name=bn_name_base + '_2b', **bnArgs)(O)
if d.aact == "complex_joint_relu":
O = Lambda(ComplexJointReLU)(O)
else:
O = Activation(activation)(O)
O = ComplexConv2D(nb_fmaps2,
filter_size,
name=conv_name_base + '2b',
**convArgs)(O)
elif (d.model == "complex_concat") or (d.model
== "complex_concat_pwc_group"):
O = ComplexBN(name=bn_name_base + '_2b', **bnArgs)(O)
if d.aact == "complex_joint_relu":
O = Lambda(ComplexJointReLU)(O)
else:
O = Activation(activation)(O)
O = ComplexConvConcat2D(nb_fmaps2 // 2,
filter_size,
name=conv_name_base + '2b',
**convArgs)(O)
if d.model == "complex_concat_pwc_group":
O_g0 = Lambda(lambda O: O[:, :(O.shape[1] // 2), :, :])(O)
O_g1 = Lambda(lambda O: O[:, (O.shape[1] // 2):, :, :])(O)
O_g0 = Conv2D(int(O.shape[1] // 2), (1, 1),
name=conv_name_base + '2b_pwc_g0',
**convArgs_real)(O_g0)
O_g1 = Conv2D(int(O.shape[1] // 2), (1, 1),
name=conv_name_base + '2b_pwc_g1',
**convArgs_real)(O_g1)
O = Concatenate(axis=1)([O_g0, O_g1])
else:
print("Error: unknown model type")
exit(-1)
if shortcut == 'regular':
O = Add()([O, I])
elif shortcut == 'projection':
if d.spectral_pool_scheme == "proj":
I = applySpectralPooling(I, d)
if d.model == "real":
X = Conv2D(
nb_fmaps2, (1, 1),
name=conv_name_base + '1',
strides=(2, 2) if d.spectral_pool_scheme != "nodownsample" else
(1, 1),
**convArgs)(I)
O = Concatenate(channel_axis)([X, O])
elif d.model == "real_dws":
X = SeparableConv2D(
nb_fmaps2, (1, 1),
name=conv_name_base + '1',
strides=(2, 2) if d.spectral_pool_scheme != "nodownsample" else
(1, 1),
**convArgs)(I)
O = Concatenate(channel_axis)([X, O])
elif (d.model
== "real_group") or (d.model == "real_group_pwc_full") or (
d.model == "real_group_pwc_group"):
I_g0 = Lambda(lambda I: I[:, :(I.shape[1] // 2), :, :])(I)
I_g1 = Lambda(lambda I: I[:, (I.shape[1] // 2):, :, :])(I)
X_g0 = Conv2D(
nb_fmaps2 // 2, (1, 1),
name=conv_name_base + '1_g0',
strides=(2, 2) if d.spectral_pool_scheme != "nodownsample" else
(1, 1),
**convArgs)(I_g0)
X_g1 = Conv2D(
nb_fmaps2 // 2, (1, 1),
name=conv_name_base + '1_g1',
strides=(2, 2) if d.spectral_pool_scheme != "nodownsample" else
(1, 1),
**convArgs)(I_g1)
X_g00 = Lambda(lambda X_g0: X_g0[:, :(X_g0.shape[1] // 2), :, :])(
X_g0)
X_g01 = Lambda(lambda X_g0: X_g0[:, (X_g0.shape[1] // 2):, :, :])(
X_g0)
X_g10 = Lambda(lambda X_g1: X_g1[:, :(X_g1.shape[1] // 2), :, :])(
X_g1)
X_g11 = Lambda(lambda X_g1: X_g1[:, (X_g1.shape[1] // 2):, :, :])(
X_g1)
X = Concatenate(axis=1)([X_g00, X_g11, X_g01, X_g10])
if d.model == "real_group_pwc_full":
X = Conv2D(nb_fmaps2, (1, 1),
name=conv_name_base + '1_pwc',
**convArgs)(X)
elif d.model == "real_group_pwc_group":
X_g0 = Lambda(lambda X: X[:, :(X.shape[1] // 2), :, :])(X)
X_g1 = Lambda(lambda X: X[:, (X.shape[1] // 2):, :, :])(X)
X_g0 = Conv2D(int(X.shape[1] // 2), (1, 1),
name=conv_name_base + '1_pwc_g0',
**convArgs_real)(X_g0)
X_g1 = Conv2D(int(X.shape[1] // 2), (1, 1),
name=conv_name_base + '1_pwc_g1',
**convArgs_real)(X_g1)
X = Concatenate(axis=1)([X_g0, X_g1])
O = Concatenate(channel_axis)([X, O])
elif d.model == "complex":
X = ComplexConv2D(
nb_fmaps2, (1, 1),
name=conv_name_base + '1',
strides=(2, 2) if d.spectral_pool_scheme != "nodownsample" else
(1, 1),
**convArgs)(I)
O_real = Concatenate(channel_axis)([GetReal()(X), GetReal()(O)])
O_imag = Concatenate(channel_axis)([GetImag()(X), GetImag()(O)])
O = Concatenate(1)([O_real, O_imag])
elif (d.model == "complex_concat") or (d.model
== "complex_concat_pwc_group"):
X = ComplexConvConcat2D(
nb_fmaps2 // 2, (1, 1),
name=conv_name_base + '1',
strides=(2, 2) if d.spectral_pool_scheme != "nodownsample" else
(1, 1),
**convArgs)(I)
if d.model == "complex_concat_pwc_group":
X_g0 = Lambda(lambda X: X[:, :(X.shape[1] // 2), :, :])(X)
X_g1 = Lambda(lambda X: X[:, (X.shape[1] // 2):, :, :])(X)
X_g0 = Conv2D(int(X.shape[1] // 2), (1, 1),
name=conv_name_base + '1_pwc_g0',
**convArgs_real)(X_g0)
X_g1 = Conv2D(int(X.shape[1] // 2), (1, 1),
name=conv_name_base + '1_pwc_g1',
**convArgs_real)(X_g1)
X = Concatenate(axis=1)([X_g0, X_g1])
O_real = Concatenate(channel_axis)([GetReal()(X), GetReal()(O)])
O_imag = Concatenate(channel_axis)([GetImag()(X), GetImag()(O)])
O = Concatenate(1)([O_real, O_imag])
else:
print("Error: unknown model type")
exit(-1)
return O
def getResnetModel(d):
n = d.num_blocks
sf = d.start_filter
dataset = d.dataset
activation = d.act
advanced_act = d.aact
drop_prob = d.dropout
if "mnist" in dataset:
inputShape = (1, 28, 28) if K.image_dim_ordering() == "th" else (28,
28, 1)
else:
inputShape = (3, 32, 32) if K.image_dim_ordering() == "th" else (32,
32, 3)
channelAxis = 1 if K.image_data_format() == 'channels_first' else -1
filsize = (3, 3)
convArgs = {
"padding": "same",
"use_bias": False,
"kernel_regularizer": l2(0.0001),
}
bnArgs = {"axis": channelAxis, "momentum": 0.9, "epsilon": 1e-04}
import copy
convArgs_real = copy.deepcopy(convArgs)
# if d.model == "real":
if "real" in d.model:
sf *= 2
convArgs.update({"kernel_initializer": Orthogonal(float(np.sqrt(2)))})
convArgs_real = convArgs
# elif d.model == "complex":
elif "complex" in d.model:
convArgs.update({
"spectral_parametrization": d.spectral_param,
"kernel_initializer": d.comp_init
})
#
# Input Layer
#
I = Input(shape=inputShape)
#
# Stage 1
#
O = learnConcatRealImagBlock(I, (1, 1), (3, 3), 0, '0', convArgs, bnArgs,
d)
O = Concatenate(channelAxis)([I, O])
if d.model == "real":
O = Conv2D(sf, filsize, name='conv1', **convArgs)(O)
O = BatchNormalization(name="bn_conv1_2a", **bnArgs)(O)
elif d.model == "real_dws":
O = SeparableConv2D(sf, filsize, name='conv1', **convArgs)(O)
O = BatchNormalization(name="bn_conv1_2a", **bnArgs)(O)
elif (d.model == "real_group") or (d.model == "real_group_pwc_full") or (
d.model == "real_group_pwc_group"):
O_g0 = Lambda(lambda O: O[:, :(O.shape[1] // 2), :, :])(O)
O_g1 = Lambda(lambda O: O[:, (O.shape[1] // 2):, :, :])(O)
O_g0 = Conv2D(sf // 2, filsize, name='conv1_g0', **convArgs)(O_g0)
O_g1 = Conv2D(sf // 2, filsize, name='conv1_g1', **convArgs)(O_g1)
O_g00 = Lambda(lambda O_g0: O_g0[:, :(O_g0.shape[1] // 2), :, :])(O_g0)
O_g01 = Lambda(lambda O_g0: O_g0[:, (O_g0.shape[1] // 2):, :, :])(O_g0)
O_g10 = Lambda(lambda O_g1: O_g1[:, :(O_g1.shape[1] // 2), :, :])(O_g1)
O_g11 = Lambda(lambda O_g1: O_g1[:, (O_g1.shape[1] // 2):, :, :])(O_g1)
O = Concatenate(axis=1)([O_g00, O_g11, O_g01, O_g10])
if d.model == "real_group_pwc_full":
O = Conv2D(sf, (1, 1), name='conv1_pwc', **convArgs)(O)
elif d.model == "real_group_pwc_group":
O_g0 = Lambda(lambda O: O[:, :(O.shape[1] // 2), :, :])(O)
O_g1 = Lambda(lambda O: O[:, (O.shape[1] // 2):, :, :])(O)
O_g0 = Conv2D(int(O.shape[1] // 2), (1, 1),
name='conv1_pwc_g0',
**convArgs_real)(O_g0)
O_g1 = Conv2D(int(O.shape[1] // 2), (1, 1),
name='conv1_pwc_g1',
**convArgs_real)(O_g1)
O = Concatenate(axis=1)([O_g0, O_g1])
O = BatchNormalization(name="bn_conv1_2a", **bnArgs)(O)
elif d.model == "complex":
O = ComplexConv2D(sf, filsize, name='conv1', **convArgs)(O)
O = ComplexBN(name="bn_conv1_2a", **bnArgs)(O)
elif (d.model == "complex_concat") or (d.model
== "complex_concat_pwc_group"):
O = ComplexConvConcat2D(sf // 2, filsize, name='conv1', **convArgs)(O)
O = ComplexBN(name="bn_conv1_2a", **bnArgs)(O)
if d.model == "complex_concat_pwc_group":
O_g0 = Lambda(lambda O: O[:, :(O.shape[1] // 2), :, :])(O)
O_g1 = Lambda(lambda O: O[:, (O.shape[1] // 2):, :, :])(O)
O_g0 = Conv2D(int(O.shape[1] // 2), (1, 1),
name='conv1_pwc_g0',
**convArgs_real)(O_g0)
O_g1 = Conv2D(int(O.shape[1] // 2), (1, 1),
name='conv1_pwc_g1',
**convArgs_real)(O_g1)
O = Concatenate(axis=1)([O_g0, O_g1])
else:
print("Error: unknown model type")
exit(-1)
if d.aact == "complex_joint_relu":
O = Lambda(ComplexJointReLU)(O)
else:
O = Activation(activation)(O)
#
# Stage 2
#
for i in xrange(n):
O = getResidualBlock(O, filsize, [sf, sf], 2, str(i), 'regular',
convArgs, convArgs_real, bnArgs, d)
if i == n // 2 and d.spectral_pool_scheme == "stagemiddle":
O = applySpectralPooling(O, d)
#
# Stage 3
#
O = getResidualBlock(O, filsize, [sf, sf], 3, '0', 'projection', convArgs,
convArgs_real, bnArgs, d)
if d.spectral_pool_scheme == "nodownsample":
O = applySpectralPooling(O, d)
for i in xrange(n - 1):
O = getResidualBlock(O, filsize, [sf * 2, sf * 2], 3, str(i + 1),
'regular', convArgs, convArgs_real, bnArgs, d)
if i == n // 2 and d.spectral_pool_scheme == "stagemiddle":
O = applySpectralPooling(O, d)
#
# Stage 4
#
O = getResidualBlock(O, filsize, [sf * 2, sf * 2], 4, '0', 'projection',
convArgs, convArgs_real, bnArgs, d)
if d.spectral_pool_scheme == "nodownsample":
O = applySpectralPooling(O, d)
for i in xrange(n - 1):
O = getResidualBlock(O, filsize, [sf * 4, sf * 4], 4, str(i + 1),
'regular', convArgs, convArgs_real, bnArgs, d)
if i == n // 2 and d.spectral_pool_scheme == "stagemiddle":
O = applySpectralPooling(O, d)
#
# Pooling
#
if d.spectral_pool_scheme == "nodownsample":
O = applySpectralPooling(O, d)
if "mnist" in dataset:
O = AveragePooling2D(pool_size=(28, 28))(O)
else:
O = AveragePooling2D(pool_size=(32, 32))(O)
else:
if "mnist" in dataset:
O = AveragePooling2D(pool_size=(7, 7))(O)
else:
O = AveragePooling2D(pool_size=(8, 8))(O)
#
# Flatten
#
O = Flatten()(O)
#
# Dense
#
if dataset == 'cifar10':
O = Dense(10, activation='softmax', kernel_regularizer=l2(0.0001))(O)
elif dataset == 'cifar100':
O = Dense(100, activation='softmax', kernel_regularizer=l2(0.0001))(O)
elif dataset == 'svhn':
O = Dense(10, activation='softmax', kernel_regularizer=l2(0.0001))(O)
elif dataset == 'mnist':
O = Dense(10, activation='softmax', kernel_regularizer=l2(0.0001))(O)
elif dataset == 'fashion_mnist':
O = Dense(10, activation='softmax', kernel_regularizer=l2(0.0001))(O)
else:
raise ValueError("Unknown dataset " + d.dataset)
# Return the model
return Model(I, O)