from __future__ import division, print_function
from keras.models import Model
from keras.layers import Input, Conv1D, Dense, add, Flatten, Dropout,MaxPooling1D, Activation, BatchNormalization, Lambda
from keras import backend as K
from keras.optimizers import Adam
from keras.saving import register_keras_serializable
import tensorflow as tf
@register_keras_serializable(package="custom")
def zeropad(x):
"""
zeropad and zeropad_output_shapes are from
https://github.com/awni/ecg/blob/master/ecg/network.py
"""
y = tf.zeros_like(x)
return tf.concat([x, y], axis=2)
@register_keras_serializable(package="custom")
def zeropad_output_shape(input_shape):
shape = list(input_shape)
assert len(shape) == 3
shape[2] *= 2
return tuple(shape)
def ECG_model(config):
"""
implementation of the model in https://www.nature.com/articles/s41591-018-0268-3
also have reference to codes at
https://github.com/awni/ecg/blob/master/ecg/network.py
and
https://github.com/fernandoandreotti/cinc-challenge2017/blob/master/deeplearn-approach/train_model.py
"""
def first_conv_block(inputs, config):
layer = Conv1D(filters=config.filter_length,
kernel_size=config.kernel_size,
padding='same',
strides=1,
kernel_initializer='he_normal')(inputs)
layer = BatchNormalization()(layer)
layer = Activation('relu')(layer)
shortcut = MaxPooling1D(pool_size=1,
strides=1)(layer)
layer = Conv1D(filters=config.filter_length,
kernel_size=config.kernel_size,
padding='same',
strides=1,
kernel_initializer='he_normal')(layer)
layer = BatchNormalization()(layer)
layer = Activation('relu')(layer)
layer = Dropout(config.drop_rate)(layer)
layer = Conv1D(filters=config.filter_length,
kernel_size=config.kernel_size,
padding='same',
strides=1,
kernel_initializer='he_normal')(layer)
return add([shortcut, layer])
def main_loop_blocks(layer, config):
filter_length = config.filter_length
n_blocks = 15
for block_index in range(n_blocks):
subsample_length = 2 if block_index % 2 == 0 else 1
shortcut = MaxPooling1D(pool_size=subsample_length)(layer)
# 5 is chosen instead of 4 from the original model
if block_index % 4 == 0 and block_index > 0 :
# double size of the network and match the shapes of both branches
shortcut = Lambda(zeropad, output_shape=zeropad_output_shape)(shortcut)
filter_length *= 2
layer = BatchNormalization()(layer)
layer = Activation('relu')(layer)
layer = Conv1D(filters= filter_length,
kernel_size=config.kernel_size,
padding='same',
strides=subsample_length,
kernel_initializer='he_normal')(layer)
layer = BatchNormalization()(layer)
layer = Activation('relu')(layer)
layer = Dropout(config.drop_rate)(layer)
layer = Conv1D(filters= filter_length,
kernel_size=config.kernel_size,
padding='same',
strides= 1,
kernel_initializer='he_normal')(layer)
layer = add([shortcut, layer])
return layer
def output_block(layer, config):
layer = BatchNormalization()(layer)
layer = Activation('relu')(layer)
layer = Flatten()(layer)
outputs = Dense(len_classes, activation='softmax')(layer)
model = Model(inputs=inputs, outputs=outputs)
adam = Adam(learning_rate=0.1, beta_1=0.9, beta_2=0.999, epsilon=1e-7, amsgrad=False)
model.compile(optimizer= adam,
loss='categorical_crossentropy',
metrics=['accuracy'])
model.summary()
return model
classes = ['N','V','/','A','F','~']#,'L','R',f','j','E','a']#,'J','Q','e','S'] are too few or not in the trainset, so excluded out
len_classes = len(classes)
inputs = Input(shape=(config.input_size, 1), name='input')
layer = first_conv_block(inputs, config)
layer = main_loop_blocks(layer, config)
return output_block(layer, config)
Datasets
Models
