def evaluateBest(ind):
try:
layers = ind[0]
par = ind[1]
dfns = []
for i in range(len(layers)):
inpt = None
if len(dfns) == 0:
inpt = params["inputShape"]
else:
inpt = dfns[-1].get_output_shape()
dfn = defineLayer(inpt, layers[i])
if dfn != None:
dfns.append(dfn)
net = NNet(dfns, params["batchSize"])
fitness = [[-1],[-1]]
fitness = net.train(training_data, 100, params["batchSize"], par["learningRate"],
validation_data, test_data, lmbda=par["l2"])
print("Individual fitness " + str(numpy.average(fitness[1])))
print(fitness)
return fitness
except Exception as e:
print("error in evaluation")
print(e)
return [[-1],[-1]]
import theano
import sys
import numpy
from NN.NN import NNet
from NN.Layers.FullyConnected import FullyConnectedLayer
from NN.Layers.Softmax import SoftmaxLayer
from NN.MNISTLoader import load_data
# CONFIG
try: theano.config.device = 'gpu'
except: pass
theano.config.floatX = 'float32'
numpy.random.seed(0)
# HYPER PARAMETERS
mini_batch_size = 100
epochs = 100
learning_rate = 1
net = NNet([
FullyConnectedLayer(n_in=784, n_out=30),
FullyConnectedLayer(n_in=30, n_out=10),
SoftmaxLayer(n_in=10, n_out=10)],
mini_batch_size)
training_data, validation_data, test_data = load_data()
net.train(training_data, epochs, mini_batch_size, learning_rate,
validation_data, test_data)