chairgen.py

# Create first network with Keras
from keras.models import Sequential
from keras.layers import *
from keras import backend as K
from keras.engine.topology import Layer
import numpy as np
import codecs, json, math

# fix random seed for reproducibility
seed = 7
np.random.seed(seed)

def jsonFrom3D(a, file):
    vx = []
    s = a.shape
    a0 = s[0]
    a1 = s[1]
    a2 = s[2]

    for x in range(a0):
        for y in range(a1):
            for z in range(a2):
                if a[x][y][z] > 0.5:
                    c = {'x':x,'y':y,'z':z}
                    vx.append(c)

    outdata = {  'dimension' : [{'width': 128, 'height': 128, 'depth': 256}], 'voxels' : vx }

    with open('/tmp/'+file, 'w') as outfile:
        json.dump(outdata, outfile)

class RepeatVector4D(Layer):

    def __init__(self, n, **kwargs):
        self.n = n
        self.input_spec = [InputSpec(ndim=3)]
        super(RepeatVector4D, self).__init__(**kwargs)

    def get_output_shape_for(self, input_shape):
        return (input_shape[0], self.n, input_shape[1], input_shape[2])

    def call(self, x, mask=None):
        x = K.expand_dims(x, 1)
        pattern = K.stack([1, self.n, 1, 1])
        out = K.tile(x, pattern)
        # jsonFrom3D(out[0], 'RV4in.json')
        return out

def loadY(file):
  # not included hehe
  
def loadX(file):
  # not included hehe


X = loadX("../index.tsv") # I spare you the content of these
Y = loadY("../index.tsv") #
print("INPUTS:  " + str(X.shape)) # > (16,9)
print("OUTPUTS: " + str(Y.shape)) # > (16,128,128,256)

# save training data to json to validate data preprocessing (spoiler: it works)
jsonFrom3D(Y[0], 'validation_0.json')


# create model
model = Sequential()
# input.shape = (*, 9)
model.add(Dense(128, input_dim=9, init='uniform', activation='relu')) # shape = (*, 128)
model.add(RepeatVector(128)) # shape = (*, 128, 128) 
model.add(RepeatVector4D(256)) # shape = (*, 256, 128, 128) 
model.add(Dense(128))
model.add(GaussianNoise(0.5))
model.add(Dense(128))
model.add(Dropout(0.2))
model.add(Dense(128))
model.add(Dense(128, activation='hard_sigmoid'))
model.add(Reshape((128,128,256)))

# Compile model
model.compile(loss='binary_crossentropy', optimizer='adam', metrics=['accuracy'])
# Fit the model
# 
model.fit(X, Y, nb_epoch=150, batch_size=16,  verbose=2)

# calculate predictions (not yet there)
hopefully_a_chair = model.predict(np.array([[2015,40,3,9,400,6,4,1,0]]))[0]

jsonFrom3D(hopefully_a_chair, 'hopefully_a_chair.json')