def run_network():
# Generate standard layered network architecture and create network
conec = mlgraph((14,28,1))
net = ffnet(conec)
df = pd.read_csv('data/copacabana.csv', sep=';')
variables = [
'Posicao', 'Quartos', 'Vagas', 'DistIpanema', 'DistPraia',
'DistFavela', 'RendaMedia', 'RendaMovel', 'RendaMovelRua',
'Vu2009', 'Mes', 'Idade', 'Tipologia', 'AreaConstruida'
]
input = df[variables]
target = df[['VAL_UNIT']]
# Train network
#first find good starting point with genetic algorithm (not necessary, but may be helpful)
print "FINDING STARTING WEIGHTS WITH GENETIC ALGORITHM..."
net.train_genetic(input, target, individuals=20, generations=500)
#then train with scipy tnc optimizer
print "TRAINING NETWORK..."
net.train_tnc(input, target, maxfun = 1000, messages=1)
print "TESTING NETWORK..."
output, regression = net.test(input, target, iprint=0)
# Save/load/export network
print "Network is saved..."
savenet(net, "data/capacabana.net")
return output, regression
def train(data):
#trains neural network based on passed training data.
#training data is a list of [input,output] lists
print "amount of training data:" + str(len(data))
inputsize = 30 * 30
outsize = 10
nodes = 350 #((inputsize + outsize) * 2) / 3
inp = [i for i,t in data]
trg = [t for i,t in data]
print "creating neural network, hidden nodes:" + str(nodes)
conec = mlgraph((inputsize,nodes,outsize))
print "initializing ffnet"
net = ffnet(conec)
#print "loading ffnet"
#net = loadnet("ffnet.net")
# print "assigning random weights"
# net.randomweights()
# Train process
print "training network"
net.train_tnc(inp, trg, messages=1,nproc=4)
print "saving trained network"
savenet(net, "ffnet.net")
print "testing network"
net.test(inp, trg, iprint = 1)
def create_then_save_network_trained_on(name, pattern, train_method):
input, output = load_snns(pattern)
net = create_bp(input, output)
print("Training the net with {0} algorithm...".format(train_method))
training = getattr(net, "train_" + train_method)
training(input, output)
filename = "{0}_{1}.net".format(name, train_method)
savenet(net, filename)
print("saved as: {0}".format(filename))
def export_network(net, ANNname, Nlayers, Nneurons, pspace):
from ffnet import savenet
import os
try:
f = '{0:}_{1:}_{2:}'.format(ANNname, Nlayers, Nneurons)
savenet(net, f)
export_ANN_to_nc(net, ANNname + '.nc', Nlayers, Nneurons, pspace)
except Exception, e:
print 'Error when writing ffnet network file', e
raise (e)
def learn_net(filename, word, **args):
(network, people,
people_num) = load_speaker_recognition_newtork(filename, True)
print "Loading voice samples..."
samples = [voice_sample.VoicePersonSamples(s) for s in people]
rs = args.get("rs", None)
re = args.get("re", None)
if rs is None or re is None:
ans = raw_input("Select sample range (1-10): ")
(rs, re) = map(lambda x: int(x), ans.strip().split("-"))
(inp, out) = get_inputs_and_outputs(samples, word, rs, re)
train_method = args.get("method", "momentum")
training = getattr(network, "train_" + train_method)
training(inp, out)
_save = args.get("save", True)
_save and savenet(network, filename)
return (network, samples)
def learn_net(filename, word, **args):
(network, people, people_num) = load_speaker_recognition_newtork(filename, True);
print "Loading voice samples..."
samples = [ voice_sample.VoicePersonSamples(s) for s in people]
rs = args.get("rs",None)
re = args.get("re", None)
if rs is None or re is None:
ans = raw_input("Select sample range (1-10): ")
(rs, re) = map(lambda x: int(x), ans.strip().split("-"))
(inp, out) = get_inputs_and_outputs(samples, word, rs, re);
train_method = args.get("method", "momentum")
training = getattr(network, "train_" + train_method)
training(inp, out)
_save = args.get("save", True)
_save and savenet(network, filename)
return (network, samples)
def save_network(self):
# Save/load/export network
from ffnet import savenet, loadnet, exportnet
print "Network is saved..."
savenet(self.net, "xor.net")
print "Network is reloaded..."
net = loadnet("xor.net")
print "Network is tested again, but nothing is printed..."
output, regression = net.test(input, target, iprint=0)
print
print "Exporting trained network to the fortran source..."
exportnet(net, "xor.f")
print "Done..."
print "Look at the generated xor.f file."
print "Note: you must compile xor.f along with the ffnet.f"
print "file which can be found in ffnet sources."
def write(self, desc, stats):
self.lock.acquire()
self.counter += 1
#print "write: "+repr(stats)
fid = open(self.filename, 'a')
r2 = np.array([t[2]**2 for t in stats.regression_ve])
print "%3d: hh=%8s, iters=%7d, r2_ve=%0.3f" % (
self.counter, repr(desc.hh), stats.iters, r2.mean()),
if 'regression_te' in stats:
r2 = np.array([t[2]**2 for t in stats.regression_te])
print ", r2_te=%0.3f" % r2.mean(),
print
#print desc,stats
#print "write: 1"
nname = "%s_%03d_%s.ffnet" % (self.project, self.counter, repr(
desc.hh))
#print "write: 2: "+os.path.join(self.outfolder,nname)
savenet(stats.net, os.path.join(self.outfolder, nname))
#print "write: 3"
fid.write(
'%d\t%s\t%d' %
(self.counter, '_'.join([str(h)
for h in desc.hh]), len(stats.net.conec)))
#print "write: 4"
#tt=net.test(ens.input_data(2),ens.target_data(2),iprint=0)
#print "1"
fid.write('\t' + '\t'.join(['%0.3f' % rr for rr in r2]))
#print "2"
fid.write('\t' + '%0.3f' % r2.mean())
fid.write('\n')
#print "3"
#fid.flush()
#print "4"
fid.close()
#print "Gotovo zapisivanje. ..."
self.num_in_qeue -= 1
self.lock.release()
def train_ANN(inputs_array, target_array, iterations, node_architecture,
**configs_dict):
# Same first dimension?
if not inputs_array.shape[0] == target_array.shape[0]:
raise Exception('Input and target arrays must have same first ' \
'dimension!')
# Specified number of input nodes matches second dim of input array?
n_input_nodes = node_architecture[0]
if len(inputs_array.shape) == 1:
sec_dim_inputs = 1
else:
sec_dim_inputs = inputs_array.shape[1]
if not n_input_nodes == sec_dim_inputs:
raise Exception('Specified input node architecture (n = %s) ' \
'incompatible with passed input arrays... Returning!'
%str(n_input_nodes))
# Specified number of target nodes matches second dim of target array?
n_target_nodes = node_architecture[-1]
if len(target_array.shape) == 1:
sec_dim_target = 1
else:
sec_dim_target = target_array.shape[1]
if not n_target_nodes == sec_dim_target:
raise Exception('Specified target node architecture (n = %s) ' \
'incompatible with passed input arrays... Returning!'
%str(n_target_nodes))
# Missing data in inputs array? (Warning only)
if np.isnan(inputs_array).any():
missing_inputs_flag = True
warnings.warn('Specified ANN training input variables contain missing ' \
'data. NaNs will be inserted into prediction series!')
else:
missing_inputs_flag = False
# Missing data in target array? (Warning only)
if np.isnan(target_array).any():
missing_target_flag = True
warnings.warn('Specified ANN training target variables contain missing ' \
'data. These will be removed for training!')
else:
missing_target_flag = False
# Check if saving trained network
save_flag = False
if 'save_network' in configs_dict.keys():
if configs_dict['save_network']:
save_flag = True
if not 'network_filepath' in configs_dict.keys():
raise Exception('You must specify a file path if you wish to ' \
'save a new network!')
else:
split_pathname_list = os.path.split(
configs_dict['network_filepath'])
if not os.path.isdir(split_pathname_list[0]):
raise Exception('The specified file path is not valid!')
if split_pathname_list[1] == '':
print 'Filename not supplied - using this_net.ann!'
configs_dict['network_filepath'] = os.path.join(
split_pathname_list[0], 'this_net.ann')
# Check if doing testing
test_flag = False
if 'test' in configs_dict:
if configs_dict['test']:
test_flag = True
# Create a second series with nans dropped
if missing_inputs_flag or missing_target_flag:
new_array = np.empty(
[inputs_array.shape[0], sec_dim_inputs + sec_dim_target])
new_array[:, :sec_dim_target] = target_array
new_array[:, sec_dim_target:] = inputs_array
new_array = new_array[~np.isnan(new_array).any(axis=1)]
clean_target_array = new_array[:, :sec_dim_target]
clean_inputs_array = new_array[:, sec_dim_target:]
# Generate network and train
conec = tmlgraph(node_architecture)
net = ffnet(conec)
net.train_tnc(clean_inputs_array,
clean_target_array,
maxfun=iterations,
messages=1)
# Save network if requested
if save_flag:
ffnet_class.savenet(net, configs_dict['network_filepath'])
# Generate full series from inputs
predict_array = net.call(inputs_array)
# Do testing if requested
if test_flag:
vars_list = [
'slope', 'intercept', 'r-value', 'p-value', 'slope stderr',
'estim. stderr'
]
valid_predict_array, stats_list = net.test(clean_inputs_array,
clean_target_array)
stats_dict = {var: stats_list[0][i] for i, var in enumerate(vars_list)}
return predict_array, stats_dict
else:
return predict_array
binary_solution = m.get_binary_from_code(m.get_morse_code_from_text())
m.read_wav_file("%s" % filename)
input_data, certainty = get_nnet_input_data(m)
output_data = np.array(binary_solution, dtype=np.float32)
if X_train is None:
X_train = input_data
Y_train = output_data
else:
X_train = np.concatenate((X_train, input_data), axis=0)
Y_train = np.concatenate((Y_train, output_data), axis=0)
net.train_tnc(X_train, Y_train, nproc=4, maxfun=200000, messages=2)
savenet(net, "morse.net")
else:
net = loadnet("morse.net")
## ===============================
## GENERATE SUBMISSION
## ===============================
with open('sampleSubmission.csv') as f:
trainingset = f.read().split("\n")
files = [(('000' + x.split(",")[0])[-3:], x.split(",")[1])
for x in trainingset[1:] if "," in x]
f = open('submission.csv', 'w')
f.write("ID,Prediction\n")
def train_ANN(inputs_array, target_array, iterations, node_architecture,
**configs_dict):
# Same first dimension?
if not inputs_array.shape[0] == target_array.shape[0]:
raise Exception('Input and target arrays must have same first ' \
'dimension!')
# Specified number of input nodes matches second dim of input array?
n_input_nodes = node_architecture[0]
if len(inputs_array.shape) == 1:
sec_dim_inputs = 1
else:
sec_dim_inputs = inputs_array.shape[1]
if not n_input_nodes == sec_dim_inputs:
raise Exception('Specified input node architecture (n = %s) ' \
'incompatible with passed input arrays... Returning!'
%str(n_input_nodes))
# Specified number of target nodes matches second dim of target array?
n_target_nodes = node_architecture[-1]
if len(target_array.shape) == 1:
sec_dim_target = 1
else:
sec_dim_target = target_array.shape[1]
if not n_target_nodes == sec_dim_target:
raise Exception('Specified target node architecture (n = %s) ' \
'incompatible with passed input arrays... Returning!'
%str(n_target_nodes))
# Missing data in inputs array? (Warning only)
if np.isnan(inputs_array).any():
missing_inputs_flag = True
warnings.warn('Specified ANN training input variables contain missing ' \
'data. NaNs will be inserted into prediction series!')
else:
missing_inputs_flag = False
# Missing data in target array? (Warning only)
if np.isnan(target_array).any():
missing_target_flag = True
warnings.warn('Specified ANN training target variables contain missing ' \
'data. These will be removed for training!')
else:
missing_target_flag = False
# Check if saving trained network
save_flag = False
if 'save_network' in configs_dict.keys():
if configs_dict['save_network']:
save_flag = True
if not 'network_filepath' in configs_dict.keys():
raise Exception('You must specify a file path if you wish to ' \
'save a new network!')
else:
split_pathname_list = os.path.split(configs_dict['network_filepath'])
if not os.path.isdir(split_pathname_list[0]):
raise Exception('The specified file path is not valid!')
if split_pathname_list[1] == '':
print 'Filename not supplied - using this_net.ann!'
configs_dict['network_filepath'] = os.path.join(split_pathname_list[0],
'this_net.ann')
# Check if doing testing
test_flag = False
if 'test' in configs_dict:
if configs_dict['test']:
test_flag = True
# Create a second series with nans dropped
if missing_inputs_flag or missing_target_flag:
new_array = np.empty([inputs_array.shape[0],
sec_dim_inputs + sec_dim_target])
new_array[:, :sec_dim_target] = target_array
new_array[:, sec_dim_target:] = inputs_array
new_array = new_array[~np.isnan(new_array).any(axis = 1)]
clean_target_array = new_array[:, :sec_dim_target]
clean_inputs_array = new_array[:, sec_dim_target:]
# Generate network and train
conec = tmlgraph(node_architecture)
net = ffnet(conec)
net.train_tnc(clean_inputs_array, clean_target_array,
maxfun = iterations, messages = 1)
# Save network if requested
if save_flag:
ffnet_class.savenet(net, configs_dict['network_filepath'])
# Generate full series from inputs
predict_array = net.call(inputs_array)
# Do testing if requested
if test_flag:
vars_list = ['slope', 'intercept', 'r-value', 'p-value',
'slope stderr', 'estim. stderr']
valid_predict_array, stats_list = net.test(clean_inputs_array,
clean_target_array)
stats_dict = {var: stats_list[0][i] for i, var in enumerate(vars_list)}
return predict_array, stats_dict
else:
return predict_array
def save_net(self):
savenet((self.net, self.epoch), "netdata.dat")
def save(self, filename):
savenet(self.network, filename)
# Train network
#first find good starting point with genetic algorithm (not necessary, but may be helpful)
print "FINDING STARTING WEIGHTS WITH GENETIC ALGORITHM..."
net.train_genetic(input, target, individuals=20, generations=500)
#then train with scipy tnc optimizer
print "TRAINING NETWORK..."
net.train_tnc(input, target, maxfun=1000, messages=1)
# Test network
print
print "TESTING NETWORK..."
output, regression = net.test(input, target, iprint=2)
# Save/load/export network
from ffnet import savenet, loadnet, exportnet
print "Network is saved..."
savenet(net, "xor.net")
print "Network is reloaded..."
net = loadnet("xor.net")
print "Network is tested again, but nothing is printed..."
output, regression = net.test(input, target, iprint=0)
print
print "Exporting trained network to the fortran source..."
exportnet(net, "xor.f")
print "Done..."
print "Look at the generated xor.f file."
print "Note: you must compile xor.f along with the ffnet.f"
print "file which can be found in ffnet sources."
coeff_type=args.coeff_type,
ANN_setup=setup,
test_percentage=args.test_perc,
num_of_proc=args.nproc,
basename=args.basename,
netcdf_output=netcdf_output,
err_increase=args.err_inc)
# train the network for the first time
net.train_tnc(src_train, trg_train, nproc=nproc)
err = Test_ANN(net, [src, trg], [src_train, trg_train], [src_test, trg_test])
print 'err:\terr_train:\terr_test:\trel_err [%]:\trel_err_train [%]:\trel_err_test [%]:'
print '{}\t{}\t{}\t{}\t{}\t{}'.format(*err)
# save network as ffnet and in netCDF4-format:
ff.savenet(net, args.basename + '_1_.net')
train_num = 2
if args.coeff_type == 'diffuse' or args.coeff_type == 'diff2diff':
ANN_to_NetCDF(net,
netcdf_output,
iprint=False,
dz=LUT[0]['dz'],
kabs=LUT[0]['kabs'],
ksca=LUT[0]['ksca'],
g=LUT[0]['g'])
else:
ANN_to_NetCDF(net,
netcdf_output,
iprint=False,
dz=LUT[0]['dz'],
kabs=LUT[0]['kabs'],
pa_data = pa_data[:len(pa_data)//N*N] limit = len(hr_data) hr_input = [sum(hr_data[x:x+N])/N for x in range(0, limit-N, N)] hr_target = [sum(hr_data[x:x+N])/N for x in range(N, limit, N)] pa_diffs = [abs(pa_data[x]-pa_data[x+1]) for x in range(limit-N)] #acc_diffs = [abs(fst-snd) for fst, snd in zip(pa_data[:-1], pa_data[1:])] pa_input = [sum(pa_diffs[x:x+N])/N for x in range(0, limit-N, N)] input = map(list, zip(hr_input, pa_input)) target = [[n] for n in hr_target] print input[:10] print target[:10] net.train_tnc(input, target) #output, regression = net.test(input, hr_target, iprint=3) output, regression = net.test(input, target, iprint=2) #print #print output #print regression savenet(net, 'hrnet.net')
binary_solution = m.get_binary_from_code(m.get_morse_code_from_text())
m.read_wav_file("%s" % filename)
input_data, certainty = get_nnet_input_data(m)
output_data = np.array(binary_solution, dtype=np.float32)
if X_train is None:
X_train = input_data
Y_train = output_data
else:
X_train = np.concatenate((X_train, input_data), axis=0)
Y_train = np.concatenate((Y_train, output_data), axis=0)
net.train_tnc(X_train, Y_train, nproc=4, maxfun=200000, messages=2)
savenet(net, "morse.net")
else:
net = loadnet("morse.net")
## ===============================
## GENERATE SUBMISSION
## ===============================
with open('sampleSubmission.csv') as f:
trainingset = f.read().split("\n")
files = [(('000' + x.split(",")[0])[-3:], x.split(",")[1]) for x in trainingset[1:] if "," in x]
f = open('submission.csv','w')
f.write("ID,Prediction\n")
target = [[1.], [0.], [0.], [1.]]
# Train network
#first find good starting point with genetic algorithm (not necessary, but may be helpful)
print "FINDING STARTING WEIGHTS WITH GENETIC ALGORITHM..."
net.train_genetic(input, target, individuals=20, generations=500)
#then train with scipy tnc optimizer
print "TRAINING NETWORK..."
net.train_tnc(input, target, maxfun = 1000, messages=1)
# Test network
print
print "TESTING NETWORK..."
output, regression = net.test(input, target, iprint = 2)
# Save/load/export network
from ffnet import savenet, loadnet, exportnet
print "Network is saved..."
savenet(net, "xor.net")
print "Network is reloaded..."
net = loadnet("xor.net")
print "Network is tested again, but nothing is printed..."
output, regression = net.test(input, target, iprint = 0)
print
print "Exporting trained network to the fortran source..."
exportnet(net, "xor.f")
print "Done..."
print "Look at the generated xor.f file."
print "Note: you must compile xor.f along with the ffnet.f"
print "file which can be found in ffnet sources."
def save(self, name):
savenet(self.net, name)
def save_net(self):
savenet((self.net, self.epoch), "netdata.dat")
def build_ffnet(sorted_data,training_set_size):
logging.info('starting new run! -----------------------------')
print 'defining network'
from ffnet import ffnet, imlgraph, mlgraph, loadnet, savenet
from time import time
from multiprocessing import cpu_count
import networkx
import pylab
#data_in_training2 = sorted_data[:training_set_size,10:-2].astype(float).tolist()
data_target_training2 = [[i] for i in sorted_data[:training_set_size,0].astype(float)]
new_data_in = sorted_data[:training_set_size,col_training_set[0]]
for i in col_training_set[1:]:
new_data_in = numpy.column_stack((new_data_in, sorted_data[:training_set_size,i]))
data_in_training2 = new_data_in.astype(float).tolist()
# Define net (large one)
conec = mlgraph(network_config, biases=False) #skipping first 11 cols
net = ffnet(conec)
print 'saving initialized net'
savenet(net, 'starting_net.n')
#net = loadnet('starting_net.n') # this way we can init a complex net just once
#print 'draw network'
#networkx.draw_graphviz(net.graph, prog='dot')
#pylab.show()
graph_weekly(net, sorted_data,training_set_size) # just saving a pic
logging.info('network built as: ' + str(network_config) )
print "TRAINING NETWORK..."
# that are many different training algos
#net.train_rprop(data_in_training2, data_target_training2, a=1.9, b=0.1, mimin=1e-06, mimax=15.0, xmi=0.5, maxiter=max_functions, disp=1)
###net.train_momentum(data_in_training2, data_target_training2, eta=0.2, momentum=0.1, maxiter=max_functions, disp=1)
stats = []
smallest_error = 1000
total = 0
try:
for i in xrange(min_loops,max_loops):
total += max_functions+i
if total>max_total:
break
print 'training for:',max_functions+i, "total is:", total
net.train_tnc(data_in_training2, data_target_training2, maxfun = max_functions+i, messages=1)
#net.train_rprop(data_in_training2, data_target_training2, a=1.2, b=0.5, mimin=1e-06, mimax=50.0, xmi=0.1, maxiter=max_functions*20, disp=1)
graph_weekly(net, sorted_data,training_set_size) # just saving a pic
in0, out0, s1, s2, mape_weekly_all = calc_stats(net,sorted_data,training_set_size)
stats.append((in0, out0,total, s1, s2, mape_weekly_all))
#if out0<=(biggest/1.4) and in0>.7:
#if out0<=(smallest_error/4) and in0>overfitting_threshold:
# print 'we hit overfitting threshold - breaking out early'
# break
if mape_weekly_all < smallest_error: # found a new best
smallest_error = mape_weekly_all
savenet(net, 'best_net.n')
except KeyboardInterrupt: # this way command-c just breaks out of this loop
pass
#net.train_cg(data_in_training2, data_target_training2, maxiter=max_functions, disp=1)
#net.train_genetic(data_in_training2, data_target_training2, individuals=max_population, generations=max_functions)
#net.train_bfgs(data_in_training2, data_target_training2, maxfun = max_functions, disp=1)
stats = sorted(stats, reverse=True, key=lambda x: x[1])
for i in stats:
temp_string = ''
for x in i:
temp_string += str(x) + ','
print temp_string
net = loadnet('best_net.n')
return net