deepmie/scnets.py

from keras import backend as K
from keras.models import Sequential, Model
from keras.layers import Dense, Dropout
from keras.layers import Reshape, UpSampling1D, Conv1D
from keras.layers import Flatten, Activation
from keras.utils import np_utils, multi_gpu_model
from keras.regularizers import l2
from keras.wrappers.scikit_learn import KerasRegressor
from keras.optimizers import Adam
import numpy as np
import matplotlib.pyplot as plt

#function to test performance on testset
def calc_mre(y_true, y_pred):
    y_err = 100*np.abs(y_true - y_pred)/y_true
    return np.mean(y_err)

#function to test performance on testset
def calc_mre_K(y_true, y_pred):
    y_err = 100*np.abs(y_true - y_pred)/y_true
    return K.mean(y_err)

#naive percentage loss
def relerr_loss(y_true, y_pred):
    y_err = np.abs(y_true - y_pred)/y_true
    y_err_f = K.flatten(y_err)
    return K.sum(y_err_f)

def conv1dmodel(in_size=8, out_size=256, ker_size=3):
    # create model
    model = Sequential()

    model.add(Dense(out_size, input_dim=in_size, 
        kernel_initializer='normal',
        name='first' ))
    model.add(Activation('relu'))
    
    model.add(Reshape((4, 64), name='Reshape1'))
    model.add(UpSampling1D(size=2, name='Up1'))

    model.add(Conv1D(filters=64, 
        kernel_size=ker_size, strides=1, padding='same', 
        dilation_rate=1, name='Conv1', 
        kernel_initializer='normal'))
    model.add(Activation('relu'))

    model.add(Conv1D(filters=32, 
        kernel_size=ker_size, strides=1, padding='same', 
        dilation_rate=1, name='Conv2',
        kernel_initializer='normal'))
    model.add(Activation('relu'))

    model.add(Flatten()) 
    
    model.compile(loss=relerr_loss, optimizer='adam', metrics=['accuracy', calc_mre_K])
    return model

def fullycon( in_size=8, out_size=250, N_hidden=3, N_neurons=250, N_gpus=1):
    """
    Returns a fully-connected model which will take a normalized size vector and return a
    spectrum
    in_size: length of the size vector
    out_size: length of the spectrum vector
    N_hidden: number of hidden layers
    N_neurons: number of neurons in each of the hidden layers
    """
    model = Sequential()
    model.add(Dense(out_size, input_dim=in_size, kernel_initializer='normal', activation='relu',
        name='first' ))
    for h in np.arange(N_hidden):
        lname = "H"+str(h)
        model.add(Dense(out_size, kernel_initializer='normal', activation='relu', name=lname ))

    model.add(Dense(out_size, kernel_initializer='normal', name='last'))

    # Compile model
    if N_gpus == 1:
        model.compile(loss=relerr_loss, optimizer='adam', metrics=['accuracy'])
    else:
        gpu_list = ["gpu(%d)" % i for i in range(N_gpus)]
        model.compile(loss=relerr_loss, optimizer='adam', metrics=['accuracy'], context = gpu_list)
    return model

#staging area for new models
def plot_training_history(history, red_factor):
    loss, val_loss = history.history['loss'], history.history['val_loss']
    loss = np.asarray(loss)/red_factor
    val_loss = np.asarray(val_loss)/red_factor
    epochs = len(loss)

    fig, axs = plt.subplots(1,1, figsize=(5,5))
    axs.semilogy(np.arange(1, epochs + 1), loss, label='train error')
    axs.semilogy(np.arange(1, epochs + 1), val_loss, label='validation error')
    axs.set_xlabel('Epoch number')
    axs.set_ylabel('Mean Relative Error (MRE) (%)')
    axs.legend(loc="best")