def get_model(dhw=[48,48,48],weights_decay = 0.,kernel_initializer='he_uniform',weights=None, deeper=False, activation=lambda: Activation('relu')):
shape = dhw+[1]
inputs = Input(shape = shape)
conv1 = Conv3D(32, kernel_size=(3, 3, 3), padding='same', use_bias=False, kernel_initializer=kernel_initializer,kernel_regularizer=l2_penality(weights_decay))(inputs)
id1 = identity_resnet_A(conv1, filters=32, weights_decay=weights_decay, kernel_initializer=kernel_initializer,
use_bias=False, bottleneck=2,activation=activation)
down1 = Reduction_A(id1, filters=64, weights_decay=weights_decay, kernel_initializer=kernel_initializer,
use_bias=False, bottleneck=2,activation=activation)
id2 = identity_resnet_A(down1, filters=64, weights_decay=weights_decay, kernel_initializer=kernel_initializer,
use_bias=False, bottleneck=2,activation=activation)
down2 = Reduction_A(id2, filters=128, weights_decay=weights_decay, kernel_initializer=kernel_initializer,
use_bias=False, bottleneck=2,activation=activation)
id3 = identity_resnet_A(down2, filters=128, weights_decay=weights_decay, kernel_initializer=kernel_initializer,
use_bias=False, bottleneck=2,activation=activation)
down3 = Reduction_A(id3, filters=256, weights_decay=weights_decay, kernel_initializer=kernel_initializer,
use_bias=False, bottleneck=2,activation=activation)
pool = GlobalAvgPool3D()(down3)
outputs = Dense(1, kernel_regularizer=l2_penality(weights_decay), kernel_initializer=kernel_initializer, activation='sigmoid')(pool)
model = Model(inputs, outputs)
if weights is not None:
model.load_weights(weights, by_name=True)
return model
def transmit_block(x, compression, activation, bn_scale, kernel_initializer,
weights_decay):
x = BatchNormalization(scale=bn_scale, axis=-1)(x)
x = activation()(x)
if (compression is not None) and (compression > 1):
*_, f = x.get_shape().as_list()
x = Conv3D(f // compression,
kernel_size=(1, 1, 1),
padding='same',
use_bias=True,
kernel_initializer=kernel_initializer,
kernel_regularizer=l2_penalty(weights_decay))(x)
x = AveragePooling3D((2, 2, 2), padding='valid')(x)
else:
x = GlobalAvgPool3D()(x)
return x
def _transmit_block(x, is_last):
bn_scale = PARAMS['bn_scale']
activation = PARAMS['activation']
kernel_initializer = PARAMS['kernel_initializer']
weight_decay = PARAMS['weight_decay']
compression = PARAMS['compression']
x = BatchNormalization(scale=bn_scale, axis=-1)(x)
x = activation()(x)
if is_last:
x = GlobalAvgPool3D()(x)
else:
*_, f = x.get_shape().as_list()
x = Conv3D(f // compression, kernel_size=(1, 1, 1), padding='same', use_bias=True,
kernel_initializer=kernel_initializer,
kernel_regularizer=l2_penalty(weight_decay))(x)
x = AveragePooling3D((2, 2, 2), padding='valid')(x)
return x
def conv_layer(inp,
f,
k=4,
s=2,
p='same',
act='relu',
bn=True,
transpose=False,
se=False,
se_ratio=16):
initializer = act if act is not None else ''
initializer = 'he_uniform' if initializer.find(
'relu') != -1 else 'glorot_uniform'
fun = Conv3DTranspose if transpose else Conv3D
out = fun(f,
k,
strides=s,
padding=p,
use_bias=False,
kernel_initializer=initializer)(inp)
if bn: out = BatchNormalization()(out)
if act == 'lrelu':
out = LeakyReLU(alpha=0.2)(out)
elif act is not None:
out = Activation(act)(out)
# squeeze and excite
if se:
out_se = GlobalAvgPool3D()(out)
r = f // se_ratio if (f // se_ratio) > 0 else 1
out_se = Reshape((1, 1, f))(out_se)
out_se = Dense(r,
use_bias=False,
kernel_initializer='he_uniform',
activation='relu')(out_se)
out_se = Dense(f, use_bias=False, activation='sigmoid')(out_se)
out = Multiply()([out, out_se])
return out
convModel = BatchNormalization(axis=-1, center=True, scale=False)(convModel)
convModel = AveragePooling2D(pool_size=2, strides=2)(convModel)
# convModel = GlobalAvgPool2D()(convModel)
convModel = Flatten()(convModel)
scz_sm = Dense(units=512, activation='relu',
kernel_regularizer=reg1)(convModel)
scz_sm = Dense(units=512, activation='relu', kernel_regularizer=reg1)(scz_sm)
scz_sm = Dropout(rate=DROPOUT)(scz_sm)
scz_sm = Dense(units=64, activation='relu')(scz_sm)
# Second, global pooling that uses the same snp emb data as input
snp_emb = Embedding(input_dim=inputDim,
output_dim=outputDim,
embeddings_regularizer=reg2,
name='SNP_input')(sczInput)
snp_gap = GlobalAvgPool3D()(snp_emb)
snp_gap = Dense(units=64, activation='relu')(snp_gap)
# combine snp conv 2D with snp embedding
concat = concatenate([scz_sm, snp_gap])
dropout = Dropout(rate=DROPOUT)(concat)
combined_output = Dense(units=numClasses, activation='softmax')(dropout)
classifier = Model(inputs=sczInput, outputs=combined_output)
# summarize layers
print("Model summary: \n", classifier.summary())
# compile the model
classifier.compile(
optimizer=adam,
# loss=[focal_loss],
kernel_regularizer=reg2,
activation='relu',
dilation_rate=(2, 2),
padding='same')(exp_conv)
exp_conv = BatchNormalization(axis=-1, center=True, scale=False)(exp_conv)
exp_conv = AveragePooling2D(pool_size=1, strides=1)(exp_conv)
exp_conv = Flatten()(exp_conv)
exp_conv = Dense(units=256, activation='relu')(exp_conv)
exp_conv = Dense(units=64, activation='relu')(exp_conv)
# Second, exp model which uses the sam exp data as input
exp_emb = Embedding(input_dim=inputDim,
output_dim=outputDim,
embeddings_regularizer=reg2,
name='EXP_input2')(exp_input)
exp_gap = GlobalAvgPool3D()(exp_emb)
exp_gap = Dense(units=64, activation='relu')(exp_gap)
# combine the SNP layers
combined1 = add([exp_conv, exp_gap])
combined2 = multiply([exp_conv, exp_gap])
combined = concatenate([combined1, combined2])
combined = Dense(units=64, activation='relu')(combined)
combined = Dropout(rate=DROPOUT)(combined)
combined_output = Dense(units=numClasses, activation='softmax')(combined)
classifier = Model(inputs=exp_input, outputs=combined_output)
# summarize layers
print("Model summary: \n", classifier.summary())
# compile the model
