combo_test.py

#!/usr/bin/env python
### Script to run a complete experiment

## Import libraries
import datatools
import os
reload(datatools)
from datatools import load_dataset, plot_images, plot_compare, load_sequence
from keras.models import Sequential  
from keras.layers.core import Dense, Activation, Flatten, Reshape
from keras.layers.recurrent import SimpleRNN, LSTM
from keras.layers.extra import TimeDistributedFlatten,TimeDistributedConvolution2D
from keras.layers.wrappers import TimeDistributed
from keras.layers.convolutional import Convolution2D

import numpy as np

import matplotlib.pyplot as plt

plt.close("all")

# Data preparation
n_prev = 12
hor = 6
Xtmp = load_dataset(linnum=[44,55,66,77,88,99,110,121])
X = np.sum(Xtmp,axis=1)

#normalisation
mu = np.mean(X.flatten())
sigma = np.std(X.flatten())

X = (X - mu)/sigma
nt,nx,ny = X.shape
#X = X.reshape([nt,nx*ny])
Xapp,yapp = load_sequence(X,n_prev=n_prev,hor=hor)

Xapp = Xapp.reshape([Xapp.shape[0],n_prev,1,nx,ny])
yapp = yapp.reshape([yapp.shape[0],nx*ny])


# Model Definition
in_out_neurons = nx*ny
n_feat = 32
filter_size=3
nhid = 25

model = Sequential()

model.add(TimeDistributed(Convolution2D(n_feat,filter_size,filter_size,border_mode='same'),input_shape=(n_prev,1,nx,ny)))
model.add(Activation("linear"))
model.add(TimeDistributed(Flatten()))
model.add(TimeDistributed(Dense(nhid)))
model.add(Activation("relu"))
model.add(LSTM(output_dim=nhid,return_sequences=False))
#model.add(SimpleRNN(input_dim=in_out_neurons,output_dim=nhid,return_sequences=False))
# parent args
#keras.layers.recurrent.Recurrent(weights=None, return_sequences=False, go_backwards=False, stateful=False, unroll=False, consume_less='cpu', input_dim=None, input_length=None)
# heir args
#keras.layers.recurrent.SimpleRNN(output_dim, init='glorot_uniform', inner_init='orthogonal', activation='tanh', W_regularizer=None, U_regularizer=None, b_regularizer=None, dropout_W=0.0, dropout_U=0.0)
model.add(Dense(input_dim=nhid,output_dim=n_feat*nx*ny))
model.add(Activation("relu"))
model.add(Reshape((n_feat,nx,ny)))
model.add(Convolution2D(1,filter_size,filter_size,border_mode='same'))
model.add(Activation("linear"))
model.add(Flatten())

model.compile(loss="mean_squared_error",optimizer="rmsprop")

model.fit(Xapp,yapp,batch_size=32,nb_epoch=50,validation_split=0.05)
y_predict = model.predict(Xapp)

#plot_compare(yapp.reshape([-1,7,7]),y_predict.reshape([-1,7,7]),[50,1000,5500,7000,11000])

# Save nn
outdir = '../data/nn_rec2'
modelname = 'rnn.json'
weights = 'weights.h5'
data =  'data.npz'

json_string = model.to_json()
open(os.path.join(outdir,modelname),'w').write(json_string)
model.save_weights(os.path.join(outdir,weights),overwrite=True)
np.savez(os.path.join(outdir,data),Xapp=Xapp,yapp=yapp)