def testSingleImprovement(self):
P, T = loadsyn3(100)
test, validation = get_validation_set(P, T)
net1 = build_feedforward(2, 6, 1)
epochs = 100
P, T = test
Y = net1.sim(P)
[num_correct_first, num_correct_second, initial_test_performance, num_first, num_second, missed] = stat(Y, T) #@UnusedVariable
P, T = validation
Y = net1.sim(P)
[num_correct_first, num_correct_second, initial_val_performance, num_first, num_second, missed] = stat(Y, T) #@UnusedVariable
best1 = train_evolutionary(net1, test, validation, epochs, random_range = 5)
P, T = test
Y = best1.sim(P)
[num_correct_first, num_correct_second, genetic_test_performance, num_first, num_second, missed] = stat(Y, T) #@UnusedVariable
P, T = validation
Y = best1.sim(P)
[num_correct_first, num_correct_second, genetic_val_performance, num_first, num_second, missed] = stat(Y, T) #@UnusedVariable
#Test sets
print(initial_test_performance, genetic_test_performance)
assert(initial_test_performance < genetic_test_performance)
#Validation sets
#print(initial_val_performance, genetic_val_performance)
#assert(initial_val_performance < genetic_val_performance)
net2 = build_feedforward(2, 6, 1)
epochs = 100
P, T = test
Y = net2.sim(P)
[num_correct_first, num_correct_second, initial_test_performance, num_first, num_second, missed] = stat(Y, T) #@UnusedVariable
P, T = validation
Y = net2.sim(P)
[num_correct_first, num_correct_second, initial_val_performance, num_first, num_second, missed] = stat(Y, T) #@UnusedVariable
best2 = traingd(net2, test, validation, epochs, block_size = 10)
P, T = test
Y = best2.sim(P)
[num_correct_first, num_correct_second, gd_test_performance, num_first, num_second, missed] = stat(Y, T) #@UnusedVariable
P, T = validation
Y = best2.sim(P)
[num_correct_first, num_correct_second, gd_val_performance, num_first, num_second, missed] = stat(Y, T) #@UnusedVariable
#Assert that an improvement has occurred in each step
#Test sets
print(initial_test_performance, genetic_test_performance, gd_test_performance)
assert(initial_test_performance < gd_test_performance)
def test3Bias(self):
net = build_feedforward(input_number = 2, hidden_number = 2, output_number = 1, hidden_function = "linear", output_function = "linear")
#Bias node should always give an output of 1
bias = net.bias_node
assert(1.0 == bias.output([1,1]))
assert(1.0 == bias.output("Crazy string argument"))
assert(1.0 == bias.output(1.0))
assert(1.0 == bias.output(-1))
assert(1.0 == bias.output(None))
inputs = [1.0,1.0]
#Set all weights to 1
for node in net.hidden_nodes:
for key, val in node.weights.iteritems():
node.weights[key] = 1.0
for key, val in net.output_nodes[0].weights.iteritems():
net.output_nodes[0].weights[key] = 1.0
one_bias = net.update(inputs)
print("Bias weight is 1: {0}".format(one_bias))
assert(one_bias == 7.0)
#Set bias weights to 0
for node in net.hidden_nodes:
node.weights[bias] = 0.0
net.output_nodes[0].weights[bias] = 0.0
zero_bias = net.update(inputs)
print("Bias weight is 0: {0}".format(zero_bias))
assert(one_bias > zero_bias)
assert(zero_bias == 4.0)
#Set bias weights to -99
for node in net.hidden_nodes:
node.weights[bias] = -99.0
net.output_nodes[0].weights[bias] = -99.0
neg_bias = net.update(inputs)
print("Bias weight is -99: {0}".format(neg_bias))
assert(zero_bias > neg_bias)
assert(neg_bias == ((1.0*2 - 99.0)*2 - 99.0))
def test1Simple(self):
net = build_feedforward(input_number = 2, hidden_number = 3, output_number = 1)
results = net.update([1, 2])
print(results)
results = net.sim([[1, 2], [2, 3]])
print(results)
com = build_feedforward_committee(input_number = 2, hidden_number = 3, output_number = 1)
results = com.update([1, 2])
print(results)
results = com.sim([[1, 2], [2, 3]])
print(results)
print("Simple done")
def test6NetworkGDTrain(self):
import numpy as np
#Train on XOR
indata = np.array([[0, 0],
[0, 1],
[1, 0],
[1, 1]])
outdata = np.array([[0],
[1],
[1],
[0]])
net = build_feedforward(2, 4, 1)
net.learn(indata, outdata, epochs=500)
print('network training xor results:')
for result, target in zip(net.sim(indata), outdata):
print(result[0], target[0])
assert(round(result[0]) == target[0])
def test2Multiplication(self):
net = build_feedforward(input_number = 2, hidden_number = 3, output_number = 1)
first_sum = 0
for node in net.get_all_nodes():
for weight in node.weights.values():
first_sum += weight
a = -11.0
net = net * a
second_sum = 0
for node in net.get_all_nodes():
for weight in node.weights.values():
second_sum += weight
net / a
third_sum = 0
for node in net.get_all_nodes():
for weight in node.weights.values():
third_sum += weight
print(first_sum, second_sum, third_sum)
assert(round(a * first_sum, 10) == round(second_sum, 10))
assert(round(first_sum, 10) == round(second_sum / a, 10))
assert(round(first_sum, 10) == round(third_sum, 10))
print("mul done")
def test5NormalNodesLinear(self):
net = build_feedforward(input_number = 2, hidden_number = 2, output_number = 1, hidden_function = "linear", output_function = "linear")
assert(5 == len(net))
#Bias node should always give an output of 1
#bias = net.bias_node
inputs = [1.0, 1.0]
#Set all weights to 1
for node in net.hidden_nodes:
for key, val in node.weights.iteritems():
node.weights[key] = 1.0
for key, val in net.output_nodes[0].weights.iteritems():
net.output_nodes[0].weights[key] = 1.0
one = net.update(inputs)
print("Weights are 1: {0}".format(one))
#Assert each individually
for node in net.hidden_nodes:
out_val = node.output(inputs)
print("Node output is: {0}".format(out_val))
assert(3.0 == out_val)
#Each hidden should have an output of 3
#So net output should be 2*3 + 1 = 7
assert(one == 7.0)
def testFilehandling(self):
print ("Testing network saving/loading")
net = build_feedforward()
results1 = net.update([1, 2])
print (results1)
filename = path.join(path.expanduser("~"), "test.ann")
print ("saving and reloading")
save_network(net, filename)
net = load_network(filename)
results2 = net.update([1, 2])
print (results2)
print (abs(results1[0] - results2[0]))
assert abs(results1[0] - results2[0]) < 0.0001 # float doesn't handle absolutes so well
print ("Good, now testing committee...")
com = build_feedforward_committee()
results1 = com.update([1, 2])
print (results1)
filename = path.join(path.expanduser("~"), "test.anncom")
print ("saving and reloading")
save_committee(com, filename)
com = load_committee(filename)
results2 = com.update([1, 2])
print (results2)
assert abs(results1[0] - results2[0]) < 0.0001 # float doesn't handle absolutes so well
print ("Results are good. Testing input parsing....")
filename = path.join(path.expanduser("~"), "ann_input_data_test_file.txt")
print ("First, split the file into a test set(80%) and validation set(20%)...")
inputs, targets = parse_file(filename, targetcols=5, ignorecols=[0, 1, 4], ignorerows=[])
test, validation = get_validation_set(inputs, targets, validation_size=0.5)
print (len(test[0]))
print (len(test[1]))
print (len(validation[0]))
print (len(validation[1]))
assert len(test) == 2
assert len(test[0]) > 0
assert len(test[1]) > 0
assert len(validation) == 2
assert len(validation[0]) > 0
assert len(validation[1]) > 0
print ("Went well, now expecting a zero size validation set...")
test, validation = get_validation_set(inputs, targets, validation_size=0)
print (len(test[0]))
print (len(test[1]))
print (len(validation[0]))
print (len(validation[1]))
assert len(test) == 2
assert len(test[0]) > 0
assert len(test[1]) > 0
assert len(validation) == 2
assert len(validation[0]) == 0
assert len(validation[1]) == 0
print ("As expected. Now a 100% validation set...")
test, validation = get_validation_set(inputs, targets, validation_size=1)
print (len(test[0]))
print (len(test[1]))
print (len(validation[0]))
print (len(validation[1]))
assert len(test) == 2
assert len(test[0]) == 0
assert len(test[1]) == 0
assert len(validation) == 2
assert len(validation[0]) > 0
assert len(validation[1]) > 0
print ("Now we test a stratified set...")
test, validation = get_validation_set(inputs, targets, validation_size=0.5, binary_column=0)
print (len(test[0]))
print (len(test[1]))
print (len(validation[0]))
print (len(validation[1]))
assert len(test) == 2
assert len(test[0]) > 0
assert len(test[1]) > 0
assert len(validation) == 2
assert len(validation[0]) > 0
assert len(validation[1]) > 0
print ("Test with no targets, the no inputs")
inputs, targets = parse_file(filename, ignorecols=[0, 1, 4], ignorerows=[])
assert (targets.size) == 0
assert (inputs.size) > 0
inputs, targets = parse_file(filename, targetcols=3, ignorecols=[0, 1, 2, 4, 5, 6, 7, 8, 9], ignorerows=[])
assert (targets.size) > 0
assert (inputs.size) == 0
#node.weights[keynode] += uniform(-random_range, random_range)
_node.weights[keynode] += choice([-1, 1]) * numpy.random.exponential(random_mean)
def train_evolutionary(net, (input_array, output_array), (validation_inputs, validation_targets), epochs = 1, population_size = 100, mutation_chance = 0.25, random_range = 1.0, random_mean = 0.2, top_number = 25, cross_over_chance = 0.5, error_function = sumsquare_total, loglevel = None, *args, **kwargs): #@UnusedVariable
"""Creates more networks and evolves the best it can.
Does NOT use any validation set..."""
if top_number > population_size:
raise ValueError('Top_number({0}) can not be larger than population size({1})!'.format(top_number, population_size))
if loglevel is not None:
logging.basicConfig(level = loglevel)
#Create a population of 50 networks
best = None
best_error = None
best_val_error = None
population = [build_feedforward(net.num_of_inputs, len(net.hidden_nodes), len(net.output_nodes)) for each in xrange(int(population_size))]
#For each generation
for generation in xrange(int(epochs)):
try: # I want to catch keyboard interrupt if the user wants to cancel training
error = {} #reset errors
top_networks = [None for each in xrange(int(top_number))] #reset top five
#For all networks, simulate, measure their error, and save the best network so far
#sim_results = mp_nets_sim(population, input_array)
sim_results = [net.sim(input_array) for net in population]
for member, sim_result in zip(population, sim_results):
error[member] = error_function(output_array, sim_result) / len(output_array)
#compare with best
if not best or error[member] < best_error:
best = member
best_error = error[member]
if validation_inputs is not None and len(validation_inputs) > 0:
