def train(net, uris, epochs, learning_rate, validation_percentage, save_file = False):
"""
@param net network to be trained
@param uris uris that the data is parsed from
@param epochs maximum number of iterations
@param learning_rate learing rate
@param validation_percentage percentage of training_set that should be used for validation instead
@param save_file optional filepath to save weights and biases
@postcondition: trains network with accurate weights and biases
with given arguments (network not returned, object is just modified)
@return percentage of correct test/verification data (see if it trained correctly)
"""
global net
global mean
global std
pitches = parser.parse(uris, stand)
inputs = np.array([ [pitch[0], pitch[1]] for pitch in pitches ])
outputs = np.array([ [pitch[2]] for pitch in pitches ])
mean = np.mean(inputs, axis=0)
std = np.std(inputs, axis=0)
inputs -= mean # zeroing data
inputs /= std # normalizing data
training_set = [ (i,o) for i, o in zip(inputs, outputs) ] # data structure for data according to neuralpy
random.shuffle(training_set) # randomize the training set so not training same things in same order
neuralpy.output("len: " + str(len(training_set)))
cutoff = int(validation_percentage * len(training_set)) # determine the cutoff index
test_set = training_set[:cutoff] # fraction of all data that is test set
training_set = training_set[cutoff:] # training set being cut down to other fraction
batch_length = int(.6 * len(training_set))
net.train(training_set, epochs, learning_rate, batch_length=batch_length, monitor_cost=True)
# count = test(net, training_set) # getting number of correct examples in training set
count = test(net, test_set) # number of correct examples in test_set
# if there is a save file specified, save the weights
# and biases to the file in json format.
if save_file:
save(save_file, net)
network = net # setting the global variable for reuse
return float(count)/len(test_set)
import neuralpy net = neuralpy.Network(2, 10, 8, 1) neuralpy.output(net.feedforward([1,1])) neuralpy.output(net.feedforward([0,1])) neuralpy.output(net.feedforward([1,0])) neuralpy.output(net.feedforward([0,0])) datum_1 = ([1, 1], [0]) datum_2 = ([1, 0], [1]) datum_3 = ([0, 1], [1]) datum_4 = ([0, 0], [0]) training_data = [datum_1, datum_2, datum_3, datum_4] epochs = 300 learning_rate = 1 net.train(training_data, epochs, learning_rate, monitor_cost=True) neuralpy.output() neuralpy.output(net.feedforward([1,1])) neuralpy.output(net.feedforward([0,1])) neuralpy.output(net.feedforward([1,0])) neuralpy.output(net.feedforward([0,0])) net.show_costs()
import neuralpy net = neuralpy.Network(2, 10, 8, 1) neuralpy.output(net.feedforward([1, 1])) neuralpy.output(net.feedforward([0, 1])) neuralpy.output(net.feedforward([1, 0])) neuralpy.output(net.feedforward([0, 0])) datum_1 = ([1, 1], [0]) datum_2 = ([1, 0], [1]) datum_3 = ([0, 1], [1]) datum_4 = ([0, 0], [0]) training_data = [datum_1, datum_2, datum_3, datum_4] epochs = 300 learning_rate = 3 net.train(training_data, epochs, learning_rate, monitor_cost=True) neuralpy.output() neuralpy.output(net.feedforward([1, 1])) neuralpy.output(net.feedforward([0, 1])) neuralpy.output(net.feedforward([1, 0])) neuralpy.output(net.feedforward([0, 0])) net.show_costs()
# | F | F | F | # |_______|_______|___________| # # In our network, 1 will represent True and # 0 will represent False # import neuralpy # set up a basic neural network with a 2-node input layer, # one 3-neuron hidden layer, and one 1-neuron output layer net = neuralpy.Network(2, 3, 1) # here is some arbitrary input that we will use to test x = [1, 1] out = net.feedforward(x) neuralpy.output(out) # here is our training_data the reflects the truth table # in the header of this file datum_1 = ([1, 1], [1]) datum_2 = ([1, 0], [1]) datum_3 = ([0, 1], [1]) datum_4 = ([0, 0], [0]) training_data = [datum_1, datum_2, datum_3, datum_4] # we set our other hyperparameter and the number of # epochs that we want to train the network for learning_rate = 1 epochs = 100
# | F | F | F | # |_______|_______|___________| # # In our network, 1 will represent True and # 0 will represent False # import neuralpy # set up a basic neural network with a 2-node input layer, # one 3-neuron hidden layer, and one 1-neuron output layer net = neuralpy.Network(2, 3, 1) # here is some arbitrary input that we will use to test x = [1, 1] out = net.feedforward(x) neuralpy.output(out) # here is our training_data the reflects the truth table # in the header of this file datum_1 = ([1, 1], [1]) datum_2 = ([1, 0], [1]) datum_3 = ([0, 1], [1]) datum_4 = ([0, 0], [0]) training_data = [datum_1, datum_2, datum_3, datum_4] # we set our other hyperparameter and the number of # epochs that we want to train the network for learning_rate = 1 epochs = 100
import classifier
import neuralpy
import grapher
net = neuralpy.Network(2, 8, 1)
uris = [ "miller_xml/" + str(i) + ".xml" for i in range(1,13) ]
epochs = 200
learning_rate = 0.05
validation_percentage = .32
ps = []
classifier.stand = "L"
for i in range(0, 10):
net.randomize_parameters()
p = classifier.train(net, uris, epochs, learning_rate, validation_percentage, save_file='results/miller_' + str(i) + '.txt')
neuralpy.output(p)
ps.append(p)
i = ps.index(max(ps))
neuralpy.output("\n\n" + str(max(ps)) + " at " + str(i))
grapher.graph(filepath='results/miller_' + str(i) + '.txt')
# grapher.graph(filepath='results/miller_4.txt')