def example_NN(hf=True):
p, inputs, s, costs = simple_NN((2, 50, 40, 30, 1))
xor_dataset = [[], []]
for i in range(50000):
x = numpy.random.randint(0, 2, (50, 2))
t = (x[:, 0:1] ^ x[:, 1:2]).astype(theano.config.floatX)
x = x.astype(theano.config.floatX)
xor_dataset[0].append(x)
xor_dataset[1].append(t)
training_examples = len(xor_dataset[0]) * 3 / 4
train = [
xor_dataset[0][:training_examples], xor_dataset[1][:training_examples]
]
valid = [
xor_dataset[0][training_examples:], xor_dataset[1][training_examples:]
]
gradient_dataset = SequenceDataset(train,
batch_size=None,
number_batches=10000)
cg_dataset = SequenceDataset(train, batch_size=None, number_batches=5000)
valid_dataset = SequenceDataset(valid,
batch_size=None,
number_batches=5000)
if hf:
hf_optimizer(p, inputs, s, costs).train(gradient_dataset,
cg_dataset,
initial_lambda=1.0,
preconditioner=True,
validation=valid_dataset)
else:
sgd_optimizer(p, inputs, costs, gradient_dataset, lr=1e-3)
def example_RNN(hf=True):
p, inputs, s, costs, h, ha = simple_RNN(100)
# memorize the first unit for 100 time-steps with binary noise
memorization_dataset = [[]]
for i in range(100000):
memorization_dataset[0].append(
numpy.random.randint(2,
size=(100, 1)).astype(theano.config.floatX))
train = [memorization_dataset[0][:-1000]]
valid = [memorization_dataset[0][-1000:]]
gradient_dataset = SequenceDataset(train,
batch_size=None,
number_batches=5000)
cg_dataset = SequenceDataset(train, batch_size=None, number_batches=1000)
valid_dataset = SequenceDataset(valid,
batch_size=None,
number_batches=1000)
if hf:
hf_optimizer(p, inputs, s, costs, 0.5 * (h + 1),
ha).train(gradient_dataset,
cg_dataset,
initial_lambda=0.5,
mu=1.0,
preconditioner=False,
validation=valid_dataset)
else:
sgd_optimizer(p, inputs, costs, gradient_dataset, lr=5e-5)
def __init__(self, network, **kwargs):
import os, tempfile, urllib
sys.path.append(tempfile.gettempdir())
try:
import hf
except:
# if hf failed to import, try downloading it and saving it locally.
logging.error('hf import failed, attempting to download %s', HF.URL)
path = os.path.join(tempfile.gettempdir(), 'hf.py')
urllib.urlretrieve(HF.URL, path)
logging.error('downloaded hf code to %s', path)
import hf
self.params = network.params(**kwargs)
self.opt = hf.hf_optimizer(
self.params,
network.inputs,
network.y,
[network.J(**kwargs)] + network.monitors,
network.hiddens[-1] if isinstance(network, recurrent.Network) else None)
logging.info('HF: %d named parameters to learn', len(self.params))
# fix mapping from kwargs into a dict to send to the hf optimizer
kwargs['validation_frequency'] = kwargs.pop('validate', sys.maxint)
for k in set(kwargs) - set(self.opt.train.im_func.func_code.co_varnames[1:]):
kwargs.pop(k)
self.kwargs = kwargs
def test_Oscillator(n_updates=50):
""" Test RNN with real-valued outputs. """
n_hidden = 250
n_in = 1
n_out = 2
n_steps = 200
n_seq = 1
fracZero = 2
np.random.seed(np.random.randint(200)+52)
baseFreq = np.linspace(0,numpy.pi*4,n_steps)
seq = np.zeros((n_seq,n_steps,n_in))
#sigDecay = np.zeros((n_steps/2))
#sigDecay[:(n_steps/fracZero)/2] = np.linspace(1,0,(n_steps/fracZero)/2)
#seqSig = sigDecay * np.cos(baseFreq)/2 + 1
#seq[:,n_steps/4:n_steps*3/4,0] = np.tile(seqSig,(n_seq,1))
#seq[:,n_steps/4:3*n_steps/4,0]]
seq[:,:n_steps/fracZero,0] = np.tile(np.linspace(1,0,num=n_steps/fracZero) * np.cos(baseFreq[:n_steps/fracZero]),(n_seq,1))
#seq = np.random.randn(n_seq, n_steps, n_in)
targets = np.zeros((n_seq, n_steps, n_out))
targets[:, :, 0] = np.tile(np.cos(baseFreq),(n_seq,1))
targets[:, :, 1] = np.tile(np.sin(baseFreq),(n_seq,1))
# SequenceDataset wants a list of sequences
# this allows them to be different lengths, but here they're not
seq = [i for i in seq]
targets = [i for i in targets]
gradient_dataset = SequenceDataset([seq, targets], batch_size=None,
number_batches=2)
cg_dataset = SequenceDataset([seq, targets], batch_size=None,
number_batches=2)
model = MetaRNN(n_in=n_in, n_hidden=n_hidden, n_out=n_out,
activation='cappedrelu')
opt = hf_optimizer(p=model.rnn.params, inputs=[model.x, model.y],
s=model.rnn.y_pred,
costs=[model.rnn.loss(model.y)], h=model.rnn.h)
opt.train(gradient_dataset, cg_dataset, num_updates=n_updates)
plt.close('all')
fig = plt.figure()
ax1 = plt.subplot(211)
plt.plot(seq[0])
ax1.set_title('input')
ax2 = plt.subplot(212)
true_targets = plt.plot(targets[0])
guess = model.predict(seq[0])
guessed_targets = plt.plot(guess, linestyle='--')
for i, x in enumerate(guessed_targets):
x.set_color(true_targets[i].get_color())
ax2.set_title('solid: true output, dashed: model output')
return model
def test_real(n_updates=100):
""" Test RNN with real-valued outputs. """
n_hidden = 10
n_in = 5
n_out = 3
n_steps = 10
n_seq = 1000
np.random.seed(0)
# simple lag test
seq = np.random.randn(n_seq, n_steps, n_in)
print 'seq1'
print seq
targets = np.zeros((n_seq, n_steps, n_out))
targets[:, 1:, 0] = seq[:, :-1, 3] # delayed 1
targets[:, 1:, 1] = seq[:, :-1, 2] # delayed 1
targets[:, 2:, 2] = seq[:, :-2, 0] # delayed 2
targets += 0.01 * np.random.standard_normal(targets.shape)
print 'targets'
print targets
# SequenceDataset wants a list of sequences
# this allows them to be different lengths, but here they're not
seq = [i for i in seq]
print 'seq2'
print seq
targets = [i for i in targets]
gradient_dataset = SequenceDataset([seq, targets], batch_size=None,
number_batches=100)
cg_dataset = SequenceDataset([seq, targets], batch_size=None,
number_batches=20)
model = MetaRNN(n_in=n_in, n_hidden=n_hidden, n_out=n_out,
activation='tanh')
opt = hf_optimizer(p=model.rnn.params, inputs=[model.x, model.y],
s=model.rnn.y_pred,
costs=[model.rnn.loss(model.y)], h=model.rnn.h)
print model
opt.train(gradient_dataset, cg_dataset, num_updates=n_updates)
plt.close('all')
fig = plt.figure()
ax1 = plt.subplot(211)
plt.plot(seq[0])
ax1.set_title('input')
ax2 = plt.subplot(212)
true_targets = plt.plot(targets[0])
guess = model.predict(seq[0])
guessed_targets = plt.plot(guess, linestyle='--')
for i, x in enumerate(guessed_targets):
x.set_color(true_targets[i].get_color())
ax2.set_title('solid: true output, dashed: model output')
def test_real(n_updates=100):
""" Test RNN with real-valued outputs. """
n_hidden = 10
n_in = 5
n_out = 3
n_steps = 10
n_seq = 1000
np.random.seed(0)
# simple lag test
seq = np.random.randn(n_seq, n_steps, n_in)
targets = np.zeros((n_seq, n_steps, n_out))
targets[:, 1:, 0] = seq[:, :-1, 3] # delayed 1
targets[:, 1:, 1] = seq[:, :-1, 2] # delayed 1
targets[:, 2:, 2] = seq[:, :-2, 0] # delayed 2
targets += 0.01 * np.random.standard_normal(targets.shape)
# SequenceDataset wants a list of sequences
# this allows them to be different lengths, but here they're not
seq = [i for i in seq]
targets = [i for i in targets]
gradient_dataset = SequenceDataset([seq, targets],
batch_size=None,
number_batches=100)
cg_dataset = SequenceDataset([seq, targets],
batch_size=None,
number_batches=20)
model = MetaRNN(n_in=n_in,
n_hidden=n_hidden,
n_out=n_out,
activation='tanh')
opt = hf_optimizer(p=model.rnn.params,
inputs=[model.x, model.y],
s=model.rnn.y_pred,
costs=[model.rnn.loss(model.y)],
h=model.rnn.h)
opt.train(gradient_dataset, cg_dataset, num_updates=n_updates)
plt.close('all')
fig = plt.figure()
ax1 = plt.subplot(211)
plt.plot(seq[0])
ax1.set_title('input')
ax2 = plt.subplot(212)
true_targets = plt.plot(targets[0])
guess = model.predict(seq[0])
guessed_targets = plt.plot(guess, linestyle='--')
for i, x in enumerate(guessed_targets):
x.set_color(true_targets[i].get_color())
ax2.set_title('solid: true output, dashed: model output')
def example_RNN(hf=True):
p, inputs, s, costs, h = simple_RNN(100)
memorization_dataset = [[]] # memorize the first unit for 100 time-steps with binary noise
for i in xrange(100000):
memorization_dataset[0].append(numpy.random.randint(2, size=(100, 1)).astype(theano.config.floatX))
train = [memorization_dataset[0][:-1000]]
valid = [memorization_dataset[0][-1000:]]
gradient_dataset = SequenceDataset(train, batch_size=None, number_batches=5000)
cg_dataset = SequenceDataset(train, batch_size=None, number_batches=1000)
valid_dataset = SequenceDataset(valid, batch_size=None, number_batches=1000)
if hf:
hf_optimizer(p, inputs, s, costs, h).train(gradient_dataset, cg_dataset, initial_lambda=0.5, mu=1.0, preconditioner=False, validation=valid_dataset)
else:
sgd_optimizer(p, inputs, costs, gradient_dataset, lr=5e-5)
def example_NN(hf=True):
p, inputs, s, costs = simple_NN((2, 50, 40, 30, 1))
xor_dataset = [[], []]
for i in xrange(50000):
x = numpy.random.randint(0, 2, (50, 2))
t = (x[:, 0:1] ^ x[:, 1:2]).astype(theano.config.floatX)
x = x.astype(theano.config.floatX)
xor_dataset[0].append(x)
xor_dataset[1].append(t)
training_examples = len(xor_dataset[0]) * 3/4
train = [xor_dataset[0][:training_examples], xor_dataset[1][:training_examples]]
valid = [xor_dataset[0][training_examples:], xor_dataset[1][training_examples:]]
gradient_dataset = SequenceDataset(train, batch_size=None, number_batches=10000)
cg_dataset = SequenceDataset(train, batch_size=None, number_batches=5000)
valid_dataset = SequenceDataset(valid, batch_size=None, number_batches=5000)
if hf:
hf_optimizer(p, inputs, s, costs).train(gradient_dataset, cg_dataset, initial_lambda=1.0, preconditioner=True, validation=valid_dataset)
else:
sgd_optimizer(p, inputs, costs, gradient_dataset, lr=1e-3)
def HFTest(seq, targets, t_seq, t_targets, n_hidden=10, n_updates=250):
""" Test RNN with hessian free optimization """
n_in = 2
n_out = 2
n_classes = 10
# SequenceDataset wants a list of sequences
# this allows them to be different lengths, but here they're not
seq = [i for i in seq]
targets = [i for i in targets]
gradient_dataset = SequenceDataset([seq, targets], batch_size=None,
number_batches=500)
cg_dataset = SequenceDataset([seq, targets], batch_size=None,
number_batches=100)
model = MetaRNN(n_in=n_in, n_hidden=n_hidden, n_out=n_out,
activation='tanh', output_type='softmax',
use_symbolic_softmax=True)
# optimizes negative log likelihood
# but also reports zero-one error
opt = hf_optimizer(p=model.rnn.params, inputs=[model.x, model.y],
s=model.rnn.y_pred,
costs=[model.rnn.loss(model.y),
model.rnn.errors(model.y)], h=model.rnn.h)
mse_updates = []
for i in range(n_updates):
opt.train(gradient_dataset, cg_dataset, num_updates=1)
mse = 0
for t in range(len(t_seq)):
guess = model.predict_proba(t_seq[t])
if guess != t_target:
mse += 1
mse_updates.append(mse)
print i
return (mse_updates, model)
def __init__(self, network, **kwargs):
import os, tempfile
try:
import urllib.request
except: # Python 2.x
import urllib
sys.path.append(tempfile.gettempdir())
try:
import hf
except:
# if hf failed to import, try downloading it and saving it locally.
logging.error('hf import failed, attempting to download %s', HF.URL)
path = os.path.join(tempfile.gettempdir(), 'hf.py')
try:
urllib.request.urlretrieve(HF.URL, path)
except: # Python 2.x
urllib.urlretrieve(HF.URL, path)
logging.info('downloaded hf code to %s', path)
import hf
loss, monitors, _ = network.loss(**kwargs)
self.params = network.params
self.opt = hf.hf_optimizer(
self.params,
network.inputs,
network.outputs[0],
[loss] + [mon for _, mon in monitors],
None)
# fix mapping from kwargs into a dict to send to the hf optimizer
kwargs['validate_every'] = kwargs.pop('validate', 1 << 60)
try:
func = self.opt.train.__func__.__code__
except: # Python 2.x
func = self.opt.train.im_func.func_code
for k in set(kwargs) - set(func.co_varnames[1:]):
kwargs.pop(k)
self.kwargs = kwargs
def __init__(self, network, **kwargs):
import os, tempfile
try:
import urllib.request
except: # Python 2.x
import urllib
sys.path.append(tempfile.gettempdir())
try:
import hf
except:
# if hf failed to import, try downloading it and saving it locally.
logging.error('hf import failed, attempting to download %s',
HF.URL)
path = os.path.join(tempfile.gettempdir(), 'hf.py')
try:
urllib.request.urlretrieve(HF.URL, path)
except: # Python 2.x
urllib.urlretrieve(HF.URL, path)
logging.info('downloaded hf code to %s', path)
import hf
self.params = network.params(**kwargs)
self.opt = hf.hf_optimizer(
self.params, network.inputs, network.y,
[network.J(**kwargs)] + [mon for _, mon in network.monitors],
network.hiddens[-1]
if isinstance(network, recurrent.Network) else None)
# fix mapping from kwargs into a dict to send to the hf optimizer
kwargs['validation_frequency'] = kwargs.pop('validate', 1 << 60)
try:
func = self.opt.train.__func__.__code__
except: # Python 2.x
func = self.opt.train.im_func.func_code
for k in set(kwargs) - set(func.co_varnames[1:]):
kwargs.pop(k)
self.kwargs = kwargs
def test_real(n_updates=100):
""" Test RNN with real-valued outputs. """
train, valid, test = process_data.load_data()
tseq, ttargets = train
vseq, vtargets = valid
test_seq, test_targets = test
length = len(tseq)
n_hidden = 6
n_in = 48
n_out = 12
n_steps = 1
n_seq = length
seq = [[i] for i in tseq]
targets = [[i] for i in ttargets]
gradient_dataset = SequenceDataset([seq, targets], batch_size=None, number_batches=100)
cg_dataset = SequenceDataset([seq, targets], batch_size=None,
number_batches=20)
model = MetaRNN(n_in=n_in, n_hidden=n_hidden, n_out=n_out,
learning_rate=0.001, learning_rate_decay=0.999,
n_epochs=500, activation='relu')
opt = hf_optimizer(p=model.rnn.params, inputs=[model.x, model.y],
s=model.rnn.y_pred,
costs=[model.rnn.loss(model.y)], h=model.rnn.h)
opt.train(gradient_dataset, cg_dataset, num_updates=n_updates)
test_seq = [[i] for i in test_seq]
test_targets = [[i] for i in test_targets]
plt.close("all")
for idx in xrange(len(test_seq)):
guess = model.predict(test_seq[idx])
plot_predictions(test_seq[idx][0], test_targets[idx][0], guess[0])
def test_cg(n=500): '''Attempt to solve a linear system using the CG function in hf_optimizer.''' A = numpy.random.uniform(-1, 1, (n, n)) A = numpy.dot(A.T, A) val, vec = numpy.linalg.eig(A) val = numpy.random.uniform(1, 5000, (n, 1)) A = numpy.dot(vec.T, val*vec) # hack into a fake hf_optimizer object x = theano.shared(0.0) s = 2.0*x hf = hf_optimizer([x], [], s, [s**2]) hf.quick_cost = lambda *args, **kwargs: 0.0 hf.global_backtracking = False hf.preconditioner = False hf.max_cg_iterations = 300 hf.batch_Gv = lambda v: numpy.dot(A, v) b = numpy.random.random(n) c, x, j, i = hf.cg(b) print print 'error on b =', abs(numpy.dot(A, x) - b).mean() print 'error on x =', abs(numpy.linalg.solve(A, b) - x).mean()
def test_cg(n=500): '''Attempt to solve a linear system using the CG function in hf_optimizer.''' A = numpy.random.uniform(-1, 1, (n, n)) A = numpy.dot(A.T, A) val, vec = numpy.linalg.eig(A) val = numpy.random.uniform(1, 5000, (n, 1)) A = numpy.dot(vec.T, val*vec) # hack into a fake hf_optimizer object x = theano.shared(0.0) s = 2.0*x hf = hf_optimizer([x], [], s, [s**2]) hf.quick_cost = lambda *args, **kwargs: 0.0 hf.global_backtracking = False hf.preconditioner = False hf.max_cg_iterations = 300 hf.batch_Gv = lambda v: numpy.dot(A, v) b = numpy.random.random(n) c, x, j, i = hf.cg(b) print print 'error on b =', abs(numpy.dot(A, x) - b).mean() print 'error on x =', abs(numpy.linalg.solve(A, b) - x).mean()
def RNN_Evaluation(sample,
lable,
n_hidden=10,
activation_func='tanh',
n_updates=20,
k_fold=5):
X = sample
y = lable
kf = KFold(n_splits=k_fold, shuffle=True)
split_num = kf.get_n_splits(X)
k = 1
G1, G2, S, Total = 0, 0, 0, 0
(AUC, p, r, f1) = (0, 0, 0, 0)
for train_index, test_index in kf.split(X):
# print("TRAIN:", train_index, "TEST:", test_index)
X_train, X_test = X[train_index], X[test_index]
y_train, y_test = y[train_index], y[test_index]
n_in = 20
n_classes = 3
n_out = n_classes # restricted to single softmax per time step
np.random.seed(0)
train_sample = [i for i in X_train]
train_lable = [i for i in y_train]
gradient_dataset = SequenceDataset([train_sample, train_lable],
batch_size=None,
number_batches=500)
cg_dataset = SequenceDataset([train_sample, train_lable],
batch_size=None,
number_batches=100)
model = MetaRNN(n_in=n_in,
n_hidden=n_hidden,
n_out=n_out,
activation=activation_func,
output_type='softmax',
use_symbolic_softmax=True)
opt = hf_optimizer(
p=model.rnn.params,
inputs=[model.x, model.y],
s=model.rnn.y_pred,
costs=[model.rnn.loss(model.y),
model.rnn.errors(model.y)],
h=model.rnn.h)
opt.train(gradient_dataset, cg_dataset, num_updates=n_updates)
y_test_vector = np.zeros((X_test.shape[0], 3), dtype='int64')
for count in range(0, X_test.shape[0]):
if (y_test[count][0] == 0):
y_test_vector[count][0] = 1
elif (y_test[count][0] == 1):
y_test_vector[count][1] = 1
else:
y_test_vector[count][2] = 1
(AUC_k, p_k, r_k, f1_k) = evaluation.evaluate(model, X_test,
y_test_vector, 0.8)
print("%s / %s Iteration:AUC: %s, Prec: %s, Rec: %s, F1: %s" %
(k, k_fold, AUC_k, p_k, r_k, f1_k))
AUC = AUC + AUC_k
p = p + p_k
r = r + r_k
f1 = f1 + f1_k
print("Average: AUC: %s, Prec: %s, Rec: %s, F1: %s" %
(AUC / k, p / k, r / k, f1 / k))
k += 1
AUC = AUC / k_fold
p = p / k_fold
r = r / k_fold
f1 = f1 / k_fold
return AUC, p, r, f1
def test_binary(multiple_out=False, n_updates=250):
""" Test RNN with binary outputs. """
n_hidden = 10
n_in = 5
if multiple_out:
n_out = 2
else:
n_out = 1
n_steps = 10
n_seq = 100
np.random.seed(0)
# simple lag test
seq = np.random.randn(n_seq, n_steps, n_in)
targets = np.zeros((n_seq, n_steps, n_out), dtype='int32')
# whether lag 1 (dim 3) is greater than lag 2 (dim 0)
targets[:, 2:, 0] = np.cast[np.int32](seq[:, 1:-1, 3] > seq[:, :-2, 0])
if multiple_out:
# whether product of lag 1 (dim 4) and lag 1 (dim 2)
# is less than lag 2 (dim 0)
targets[:, 2:, 1] = np.cast[np.int32](
(seq[:, 1:-1, 4] * seq[:, 1:-1, 2]) > seq[:, :-2, 0])
# SequenceDataset wants a list of sequences
# this allows them to be different lengths, but here they're not
seq = [i for i in seq]
targets = [i for i in targets]
gradient_dataset = SequenceDataset([seq, targets], batch_size=None,
number_batches=500)
cg_dataset = SequenceDataset([seq, targets], batch_size=None,
number_batches=100)
model = MetaRNN(n_in=n_in, n_hidden=n_hidden, n_out=n_out,
activation='tanh', output_type='binary')
# optimizes negative log likelihood
# but also reports zero-one error
opt = hf_optimizer(p=model.rnn.params, inputs=[model.x, model.y],
s=model.rnn.y_pred,
costs=[model.rnn.loss(model.y),
model.rnn.errors(model.y)], h=model.rnn.h)
# using settings of initial_lambda and mu given in Nicolas' RNN example
# seem to do a little worse than the default
opt.train(gradient_dataset, cg_dataset, num_updates=n_updates)
seqs = xrange(10)
plt.close('all')
for seq_num in seqs:
fig = plt.figure()
ax1 = plt.subplot(211)
plt.plot(seq[seq_num])
ax1.set_title('input')
ax2 = plt.subplot(212)
true_targets = plt.step(xrange(n_steps), targets[seq_num], marker='o')
guess = model.predict_proba(seq[seq_num])
guessed_targets = plt.step(xrange(n_steps), guess)
plt.setp(guessed_targets, linestyle='--', marker='d')
for i, x in enumerate(guessed_targets):
x.set_color(true_targets[i].get_color())
ax2.set_ylim((-0.1, 1.1))
ax2.set_title('solid: true output, dashed: model output (prob)')
def test_softmax(n_updates=250):
""" Test RNN with softmax outputs. """
n_hidden = 10
n_in = 5
n_steps = 10
n_seq = 100
n_classes = 3
n_out = n_classes # restricted to single softmax per time step
np.random.seed(0)
# simple lag test
seq = np.random.randn(n_seq, n_steps, n_in)
targets = np.zeros((n_seq, n_steps), dtype='int32')
thresh = 0.5
# if lag 1 (dim 3) is greater than lag 2 (dim 0) + thresh
# class 1
# if lag 1 (dim 3) is less than lag 2 (dim 0) - thresh
# class 2
# if lag 2(dim0) - thresh <= lag 1 (dim 3) <= lag2(dim0) + thresh
# class 0
targets[:, 2:][seq[:, 1:-1, 3] > seq[:, :-2, 0] + thresh] = 1
targets[:, 2:][seq[:, 1:-1, 3] < seq[:, :-2, 0] - thresh] = 2
#targets[:, 2:, 0] = np.cast[np.int](seq[:, 1:-1, 3] > seq[:, :-2, 0])
# SequenceDataset wants a list of sequences
# this allows them to be different lengths, but here they're not
seq = [i for i in seq]
targets = [i for i in targets]
gradient_dataset = SequenceDataset([seq, targets], batch_size=None,
number_batches=500)
cg_dataset = SequenceDataset([seq, targets], batch_size=None,
number_batches=100)
model = MetaRNN(n_in=n_in, n_hidden=n_hidden, n_out=n_out,
activation='tanh', output_type='softmax',
use_symbolic_softmax=True)
# optimizes negative log likelihood
# but also reports zero-one error
opt = hf_optimizer(p=model.rnn.params, inputs=[model.x, model.y],
s=model.rnn.y_pred,
costs=[model.rnn.loss(model.y),
model.rnn.errors(model.y)], h=model.rnn.h)
# using settings of initial_lambda and mu given in Nicolas' RNN example
# seem to do a little worse than the default
opt.train(gradient_dataset, cg_dataset, num_updates=n_updates)
seqs = xrange(10)
plt.close('all')
for seq_num in seqs:
fig = plt.figure()
ax1 = plt.subplot(211)
plt.plot(seq[seq_num])
ax1.set_title('input')
ax2 = plt.subplot(212)
# blue line will represent true classes
true_targets = plt.step(xrange(n_steps), targets[seq_num], marker='o')
# show probabilities (in b/w) output by model
guess = model.predict_proba(seq[seq_num])
guessed_probs = plt.imshow(guess.T, interpolation='nearest',
cmap='gray')
ax2.set_title('blue: true class, grayscale: probs assigned by model')
#gradient_dataset = SequenceDataset([x_seq_list, y_seq_list], batch_size=None, number_batches=100)
#cg_dataset = SequenceDataset([x_seq_list, y_seq_list], batch_size=None, number_batches=50)
print("Convert to SequenceDataset ...")
gradient_dataset = SequenceDataset([x_seq_list, x_seq_list], batch_size=None, number_batches=500)
cg_dataset = SequenceDataset([x_seq_list, x_seq_list], batch_size=None, number_batches=1000)
model = MetaRNN(n_in=n_in, n_hidden=n_hidden, n_out=n_out, n_epochs=n_epochs,
activation='tanh', output_type='softmax',
use_symbolic_softmax=True, L2_reg=0.0001)
# optimizes negative log likelihood
# but also reports zero-one error
opt = hf_optimizer(p=model.rnn.params, inputs=[model.x, model.y],
s=model.rnn.y_pred,
n_in=n_in, word2vec=word2vec, word2label=word2label,
costs=[model.rnn.loss(model.y),
model.rnn.errors(model.y)], h=model.rnn.h)
print("Training ...")
opt.train(gradient_dataset, cg_dataset, num_updates=n_updates, save_progress='save_param')
seqs = range(10)
plt.close('all')
for seq_num in seqs:
fig = plt.figure()
ax1 = plt.subplot(211)
plt.plot(x_seq_list[seq_num])
ax1.set_title('input')
ax2 = plt.subplot(212)
def RNN_leave_one_cross_validation(seq, targets, n_updates=250, n_seq=182):
""" Test RNN with softmax outputs. """
length = len(seq)
right, first, second, third = 0, 0, 0, 0
false_list = []
for k in range(0, length):
n_hidden = 10
n_in = 40
n_classes = 3
n_out = n_classes # restricted to single softmax per time step
np.random.seed(0)
train_sample = copy.deepcopy(seq)
train_lable = copy.deepcopy(targets)
test_sample = seq[k]
train_sample = np.delete(train_sample, k, 0)
train_lable = np.delete(train_lable, k, 0)
train_sample = [i for i in train_sample]
train_lable = [i for i in train_lable]
gradient_dataset = SequenceDataset([train_sample, train_lable],
batch_size=None,
number_batches=500)
cg_dataset = SequenceDataset([train_sample, train_lable],
batch_size=None,
number_batches=100)
model = MetaRNN(n_in=n_in,
n_hidden=n_hidden,
n_out=n_out,
activation='tanh',
output_type='softmax',
use_symbolic_softmax=True)
opt = hf_optimizer(
p=model.rnn.params,
inputs=[model.x, model.y],
s=model.rnn.y_pred,
costs=[model.rnn.loss(model.y),
model.rnn.errors(model.y)],
h=model.rnn.h)
opt.train(gradient_dataset, cg_dataset, num_updates=n_updates)
guess = model.predict_proba(test_sample)
tmp_list = np.ndarray.tolist(guess.T)
# print tmp_list
if (targets[k][0] == 0):
if (tmp_list.index(max(tmp_list)) == targets[k][0]):
print k, True
first += 1
right += 1
else:
print k, False
false_list.append(k)
elif (targets[k][0] == 1):
if (tmp_list.index(max(tmp_list)) == targets[k][0]):
print k, True
second += 1
right += 1
else:
print k, False
false_list.append(k)
else:
if (tmp_list.index(max(tmp_list)) == targets[k][0]):
print k, True
third += 1
right += 1
else:
print k, False
false_list.append(k)
print "class G1:", 1.0 * first / 59
print "class S:", 1.0 * second / 58
print "class G2:", 1.0 * third / 65
print "class total:", 1.0 * right / length
print false_list
targets[:, 2:][seq[:, 1:-1, 3] < seq[:, :-2, 0] - thresh] = 2
#targets[:, 2:, 0] = np.cast[np.int](seq[:, 1:-1, 3] > seq[:, :-2, 0])
'''
gradient_dataset = SequenceDataset([x_seq_list, y_seq_list], batch_size=None, number_batches=500)
cg_dataset = SequenceDataset([x_seq_list, y_seq_list], batch_size=None, number_batches=100)
model = MetaRNN(n_in=n_in, n_hidden=n_hidden, n_out=n_out, n_epochs=n_epochs,
activation='tanh', output_type='softmax',
use_symbolic_softmax=True, L2_reg=0.01)
# optimizes negative log likelihood
# but also reports zero-one error
opt = hf_optimizer(p=model.rnn.params, inputs=[model.x, model.y],
s=model.rnn.y_pred,
costs=[model.rnn.loss(model.y),
model.rnn.errors(model.y)], h=model.rnn.h)
opt.train(gradient_dataset, cg_dataset, num_updates=n_updates, save_progress='save_param')
seqs = range(10)
plt.close('all')
for seq_num in seqs:
fig = plt.figure()
ax1 = plt.subplot(211)
plt.plot(x_seq_list[seq_num])
ax1.set_title('input')
ax2 = plt.subplot(212)
# blue line will represent true classes
def RNN_k_fold_croos_validation(sample, lable, n_updates=20):
X = sample
y = lable
kf = KFold(n_splits=5, shuffle=True)
split_num = kf.get_n_splits(X)
k = 1
G1, G2, S, Total = 0, 0, 0, 0
for train_index, test_index in kf.split(X):
# print("TRAIN:", train_index, "TEST:", test_index)
X_train, X_test = X[train_index], X[test_index]
y_train, y_test = y[train_index], y[test_index]
n_hidden = 10
n_in = 40
n_classes = 3
n_out = n_classes # restricted to single softmax per time step
np.random.seed(0)
train_sample = [i for i in X_train]
train_lable = [i for i in y_train]
gradient_dataset = SequenceDataset([train_sample, train_lable],
batch_size=None,
number_batches=500)
cg_dataset = SequenceDataset([train_sample, train_lable],
batch_size=None,
number_batches=100)
model = MetaRNN(n_in=n_in,
n_hidden=n_hidden,
n_out=n_out,
activation='tanh',
output_type='softmax',
use_symbolic_softmax=True)
opt = hf_optimizer(
p=model.rnn.params,
inputs=[model.x, model.y],
s=model.rnn.y_pred,
costs=[model.rnn.loss(model.y),
model.rnn.errors(model.y)],
h=model.rnn.h)
opt.train(gradient_dataset, cg_dataset, num_updates=n_updates)
right, first, second, third = 0, 0, 0, 0
first_sum, second_sum, third_sum = 0, 0, 0
for count in range(0, X_test.shape[0]):
guess = model.predict_proba(X_test[count])
tmp_list = np.ndarray.tolist(guess.T)
# print test_sample.shape
if (y_test[count][0] == 0):
first_sum += 1
if (tmp_list.index(max(tmp_list)) == y_test[count][0]):
print True
first += 1
right += 1
else:
print k, False
elif (y_test[count][0] == 1):
second_sum += 1
if (tmp_list.index(max(tmp_list)) == y_test[count][0]):
print True
second += 1
right += 1
else:
print False
else:
third_sum += 1
if (tmp_list.index(max(tmp_list)) == y_test[count][0]):
print True
third += 1
right += 1
else:
print False
print "...................................................................................................."
print k, "interation"
print "...................................................................................................."
G1 = G1 + 1.0 * first / first_sum
S = S + 1.0 * second / second_sum
G2 = G2 + 1.0 * third / third_sum
Total = Total + 1.0 * right / X_test.shape[0]
print "class G1:", G1 / k
print "class S:", S / k
print "class G2:", G2 / k
print "class total:", Total / k
k += 1
print "...................................................................................................."
print "Final Result:"
print "...................................................................................................."
print "class G1:", G1 / split_num
print "class S:", S / split_num
print "class G2:", G2 / split_num
print "class total:", Total / split_num
def test_dCue(n_updates=1000):
""" Test RNN with softmax outputs. """
n_hidden = 50
n_in = 2
n_steps = 60
n_seq = 40
n_out = 2
cueOn = 0.2
cueOff = 0.25
decideOn = 0.25
np.random.seed(np.random.randint(1e3))
cueOn = round(cueOn*n_steps)
cueOff = round(cueOff*n_steps)
decideOn = round(decideOn*n_steps)
trialConditions = np.random.randint(0,2,n_seq)
seq = np.zeros((n_seq, n_steps, n_in))
targets = np.zeros((n_seq, n_steps, n_in), dtype='int32')
for trial in xrange(n_seq):
seq[trial,cueOn:cueOff,trialConditions[trial]] = 1
targets[trial,decideOn:,trialConditions[trial]] = 1
# SequenceDataset wants a list of sequences
# this allows them to be different lengths, but here they're not
seq = [i for i in seq]
targets = [i for i in targets]
gradient_dataset = SequenceDataset([seq, targets], batch_size=None,
number_batches=2)
cg_dataset = SequenceDataset([seq, targets], batch_size=None,
number_batches=2)
model = MetaRNN(n_in=n_in, n_hidden=n_hidden, n_out=n_out,
activation='tanh', output_type='binary')
# optimizes negative log likelihood
# but also reports zero-one error
opt = hf_optimizer(p=model.rnn.params, inputs=[model.x, model.y],
s=model.rnn.y_pred,
costs=[model.rnn.loss(model.y),
model.rnn.errors(model.y)], h=model.rnn.h)
# using settings of initial_lambda and mu given in Nicolas' RNN example
# seem to do a little worse than the default
opt.train(gradient_dataset, cg_dataset, num_updates=n_updates)
# seqs = xrange(10)
#
# plt.close('all')
# for seq_num in seqs:
# fig = plt.figure()
# ax1 = plt.subplot(211)
# plt.plot(seq[seq_num])
# ax1.set_title('input')
#
# ax2 = plt.subplot(212)
# # blue line will represent true classes
# true_targets = plt.step(xrange(n_steps), targets[seq_num], marker='o')
#
# # show probabilities (in b/w) output by model
# guess = model.predict_proba(seq[seq_num])
# guessed_probs = plt.imshow(guess.T, interpolation='nearest',
# cmap='gray')
# ax2.set_title('blue: true class, grayscale: probs assigned by model')
return model
valid_dataset = SequenceDataset(valid_dataset, batch_size=500, number_batches=20) #p, inputs, s, costs = My_simple_NN([784, 512, 512, 512, 10]) p, inputs, s, costs = DNN([784, 1000, 500, 250, 30, 10]) #p, inputs, s, costs = CNN() #numpy_rng = numpy.random.RandomState(89677) #NN=SdA(numpy_rng=numpy_rng, # n_ins=28 * 28, # hidden_layers_sizes=[1000, 1000, 1000], # n_outs=10) #p, inputs, s, costs = NN.model() #sgd_optimizer(p, inputs, costs, gradient_dataset, valid_dataset, acceleration=False, lr_init=0.08, momentum=0.5, logfile='./Log/log_gd_test2') dnn = hf_optimizer(p, inputs, s, costs) print 'Done building model' dnn.train(gradient_dataset, cg_dataset, num_updates=500, initial_lambda=30, preconditioner=False, validation=valid_dataset,logfile='./Log/log_HF_DNN') #dnn.train_gd(gradient_dataset, valid_dataset, acceleration=False, lr_init=0.08, momentum=0.5, logfile='./Log/log_gd_noacceleration_0.08') #gradient=dnn.get_gradient(gradient_dataset) #dnn.mu=0 #v= numpy.ones(sum(dnn.sizes),dtype=theano.config.floatX) #col = dnn.batch_Gv_test(v, cg_dataset, lambda_=0) #dnn.cg_dataset = cg_dataset #dnn.gradient_dataset = gradient_dataset #dnn.lambda_ = 30 #dnn.preconditioner = False #dnn.max_cg_iterations = 250 #dnn.global_backtracking = False #res=dnn.batch_Jv(v) #after_cost, flat_delta, backtracking, num_cg_iterations = dnn.cg(-gradient)
#targets[:, 2:, 0] = np.cast[np.int](seq[:, 1:-1, 3] > seq[:, :-2, 0])
'''
print("Convert to SequenceDataset ...")
gradient_dataset = SequenceDataset([x_seq_list, y_seq_list], batch_size=None, number_batches=100)
cg_dataset = SequenceDataset([x_seq_list, y_seq_list], batch_size=None, number_batches=50)
model = MetaRNN(n_in=n_in, n_hidden=n_hidden, n_out=n_out, n_epochs=n_epochs,
activation='tanh', output_type='softmax',
use_symbolic_softmax=True, L2_reg=0.0001)
# optimizes negative log likelihood
# but also reports zero-one error
opt = hf_optimizer(p=model.rnn.params, inputs=[model.x, model.y],
s=model.rnn.y_pred, n_in=n_in,
costs=[model.rnn.loss(model.y),
model.rnn.errors(model.y)], h=model.rnn.h)
print("Training ...")
opt.train(gradient_dataset, cg_dataset, initial_lambda=1.0, num_updates=n_updates, save_progress='save_param')
seqs = range(10)
plt.close('all')
for seq_num in seqs:
fig = plt.figure()
ax1 = plt.subplot(211)
plt.plot(x_seq_list[seq_num])
ax1.set_title('input')
ax2 = plt.subplot(212)
