#!/usr/bin/env python2.7
# Li Ding
# Mar. 2018
from __future__ import division
from keras.models import Model
from keras.layers import *
from keras.layers.core import *
from keras.layers.convolutional import *
from keras import backend as K
from keras.optimizers import rmsprop
def dice_coeff(y_true, y_pred):
smooth = 0.
y_true_f = K.flatten(y_true)
y_pred_f = K.flatten(y_pred)
intersection = K.sum(y_true_f * y_pred_f)
score = (2. * intersection + smooth) / (K.sum(y_true_f) + K.sum(y_pred_f) + smooth)
return score
def IOU(y_true, y_pred):
y_true_f = K.flatten(y_true)
y_pred_f = K.flatten(y_pred)
intersection = K.sum(y_true_f * y_pred_f)
score = intersection / (K.sum(y_true_f) + K.sum(y_pred_f) - intersection)
return score
def dice_loss(y_true, y_pred):
loss = 1 - dice_coeff(y_true, y_pred)
return loss
def sigmoid_cross_entropy(y_true, y_pred):
z = K.flatten(y_true)
x = K.flatten(y_pred)
q = 10
l = (1 + (q - 1) * z)
loss = (K.sum((1 - z) * x) + K.sum(l * (K.log(1 + K.exp(- K.abs(x))) + K.max(-x, 0)))) / 500
return loss
def GRU64(n_nodes, conv_len, n_classes, n_feat, in_len,
optimizer=rmsprop(lr=1e-3), return_param_str=False):
n_layers = len(n_nodes)
inputs = Input(shape=(in_len, n_feat))
model = inputs
model = CuDNNGRU(64, return_sequences=True)(model)
model = SpatialDropout1D(0.5)(model)
model.set_shape((None, in_len, 64))
model = TimeDistributed(Dense(n_classes, name='fc', activation='softmax'))(model)
model = Model(inputs=inputs, outputs=model)
model.compile(optimizer=optimizer, loss='categorical_crossentropy', sample_weight_mode="temporal")
if return_param_str:
param_str = "GRU_C{}_L{}".format(conv_len, n_layers)
return model, param_str
else:
return model
def TCFPN(n_nodes, conv_len, n_classes, n_feat, in_len,
optimizer=rmsprop(lr=1e-4), return_param_str=False):
n_layers = len(n_nodes)
inputs = Input(shape=(in_len, n_feat))
model = inputs
lyup = []
lydown = []
# ---- Encoder ----
for i in range(n_layers):
model = Conv1D(n_nodes[i], conv_len, padding='same', use_bias=False)(model)
model = BatchNormalization()(model)
model = SpatialDropout1D(0.1)(model)
model = Activation('relu')(model)
model = MaxPooling1D(2, padding='same')(model)
lyup.append(model)
# ---- Decoder ----
model = Conv1D(n_nodes[0], 1, padding='same', use_bias=False)(model)
modelout = SpatialDropout1D(0.1)(model)
modelout = TimeDistributed(Dense(n_classes, name='fc', activation='softmax'))(modelout)
modelout = UpSampling1D(8)(modelout)
lydown.append(modelout)
model = UpSampling1D(2)(model)
res = Conv1D(n_nodes[0], 1, padding='same', use_bias=False)(lyup[-2])
model = add([model, res])
model = Conv1D(n_nodes[0], conv_len, padding='same', use_bias=False)(model)
modelout = SpatialDropout1D(0.1)(model)
modelout = TimeDistributed(Dense(n_classes, name='fc', activation='softmax'))(modelout)
modelout = UpSampling1D(4)(modelout)
lydown.append(modelout)
model = UpSampling1D(2)(model)
res = Conv1D(n_nodes[0], 1, padding='same', use_bias=False)(lyup[-3])
model = add([model, res])
model = Conv1D(n_nodes[0], conv_len, padding='same', use_bias=False)(model)
modelout = SpatialDropout1D(0.1)(model)
modelout = TimeDistributed(Dense(n_classes, name='fc', activation='softmax'))(modelout)
modelout = UpSampling1D(2)(modelout)
lydown.append(modelout)
model = UpSampling1D(2)(model)
res = Conv1D(n_nodes[0], 1, padding='same', use_bias=False)(inputs)
model = add([model, res])
model = Conv1D(n_nodes[0], conv_len, padding='same', use_bias=False)(model)
modelout = SpatialDropout1D(0.1)(model)
modelout = TimeDistributed(Dense(n_classes, name='fc', activation='softmax'))(modelout)
lydown.append(modelout)
model = average(lydown)
model = Model(inputs=inputs, outputs=model)
model.compile(optimizer=optimizer, loss='categorical_crossentropy',
sample_weight_mode="temporal")
if return_param_str:
param_str = "TCFPN_C{}_L{}".format(conv_len, n_layers)
return model, param_str
else:
return model
