def functionPlot(*args):
lstm.save('lstm.net')
str = ''
x = to_one_hot_vect(char_to_ix['c'], vocab_size)
for i in range(200):
y = sm.forward(lstm.forward(x))
str += ix_to_char[np.random.choice(range(vocab_size), p=y.ravel())]
x = to_one_hot_vect(np.argmax(y), vocab_size)
print str
display.show(*args)
def gen_data():
# N clusters:
# data, targets = datasets.make_classification(
# n_samples=n, n_features=2, n_informative=2, n_redundant=0, n_classes=num_classes, class_sep=3.0, n_clusters_per_class=1)
data, targets = datasets.make_gaussian_quantiles(mean=(0, 0),
cov=1,
n_samples=n,
n_classes=num_classes)
# Circles:
# data, targets = datasets.make_circles(
# n_samples=n, shuffle=True, noise=0.1, random_state=None, factor=0.1)
# Moons:
# data, targets = datasets.make_moons(n_samples=n, shuffle=True, noise=0.05)
# print data
# print targets
targets = [to_one_hot_vect(target, num_classes) for target in targets]
train = zip(
np.array(data[:n * 9 / 10]).astype(np.float),
np.array(targets[:n * 9 / 10]).astype(np.float))
test = zip(
np.array(data[n / 10:]).astype(np.float),
np.array(targets[n / 10:]).astype(np.float))
return train, test
def reinforcement(self, x, r):
self.states_history.append(x)
self.command_history.append(self.command)
if len(self.command_history) >= self.memory_size:
# print np.argmax(self.command_history,axis=1)
if r != 0:
self.target = r * np.multiply(self.e,
np.array(self.command_history))
# self.target = r*np.array(self.command_history)
self.trainer.learn_minibatch(
self.model,
zip(self.states_history, self.target),
self.loss,
self.optimiser,
)
self.command = utils.to_one_hot_vect(
np.argmax(self.model.forward(x, True)), self.action_size)
return self.command
def load_mnist_dataset(dataset="training", path="."):
if dataset is "training":
fname_img = os.path.join(path, 'train-images.idx3-ubyte')
fname_lbl = os.path.join(path, 'train-labels.idx1-ubyte')
elif dataset is "testing":
fname_img = os.path.join(path, 't10k-images.idx3-ubyte')
fname_lbl = os.path.join(path, 't10k-labels.idx1-ubyte')
else:
raise ValueError, "dataset must be 'testing' or 'training'"
# Load everything in some numpy arrays
with open(fname_lbl, 'rb') as flbl:
magic, num = struct.unpack(">II", flbl.read(8))
lbl = numpy.fromfile(flbl, dtype=numpy.int8)
with open(fname_img, 'rb') as fimg:
magic, num, rows, cols = struct.unpack(">IIII", fimg.read(16))
img = numpy.fromfile(fimg, dtype=numpy.uint8).reshape(
len(lbl), rows * cols).astype(numpy.float)
return zip(img, numpy.array([to_one_hot_vect(v, 10) for v in lbl]))
def lossFun(inputs, targets, hprev):
"""
inputs,targets are both list of integers.
hprev is Hx1 array of initial hidden state
returns the loss, gradients on model parameters, and last hidden state
"""
xs, hs, ys, ps = {}, {}, {}, {}
hs[-1] = np.copy(hprev)
loss = 0
# forward pass
for t in xrange(len(inputs)):
xs[t] = np.zeros((vocab_size, 1)) # encode in 1-of-k representation
xs[t][inputs[t]] = 1
hs[t] = np.tanh(np.dot(Wxh, xs[t]) + np.dot(Whh, hs[t - 1]) +
bh) # hidden state
ys[t] = np.dot(
Why, hs[t]) + by # unnormalized log probabilities for next chars
ps[t] = np.exp(ys[t] - np.max(ys[t])) / np.sum(
np.exp(ys[t] - np.max(ys[t]))) # probabilities for next chars
loss += -np.log(ps[t][targets[t], 0]) # softmax (cross-entropy loss)
assert_array_equal(van.window_step, t)
assert_array_equal(van.state[t - 1], hs[t - 1].T[0])
assert_array_equal(
van.statenet[t].forward([xs[t].T[0], hs[t - 1].T[0]]), hs[t].T[0])
assert_array_equal(
van.statenet[t].net.elements[0].elements[0].elements[1].W.get(),
Wxh)
assert_array_equal(
van.statenet[t].net.elements[0].elements[1].elements[1].W.get(),
Whh)
assert_array_equal(van.statenet[t].net.elements[0].elements[2].W.get(),
bh.T[0])
assert_array_equal(
vantr.statenet[t].net.elements[0].elements[0].elements[1].W.get(),
Wxh)
assert_array_equal(
vantr.statenet[t].net.elements[0].elements[1].elements[1].W.get(),
Whh)
assert_array_equal(
vantr.statenet[t].net.elements[0].elements[2].W.get(), bh.T[0])
assert_array_equal(
vantr.outputnet[t].net.elements[0].elements[1].W.get(), Why)
assert_array_equal(vantr.outputnet[t].net.elements[1].W.get(), by.T[0])
#
# #Neg
# assert_array_almost_equal(van.outputnet[t].net.elements[0].elements[0].elements[1].W,Why)
# assert_array_almost_equal(van.outputnet[t].net.elements[0].elements[1].W,by.T[0])
# assert_array_almost_equal(van.outputnet[t].forward(hs[t].T[0]),ps[t].T[0])
# assert_array_almost_equal(van.outputnet[t].forward(van.statenet[t].forward([xs[t].T[0],hs[t-1].T[0]])),ps[t].T[0])
# assert_array_almost_equal(van.forward(xs[t].T[0]),ps[t].T[0])
#
# Cross
assert_array_equal(
van.outputnet[t].net.elements[0].elements[1].W.get(), Why)
assert_array_equal(van.outputnet[t].net.elements[1].W.get(), by.T[0])
assert_array_equal(van.outputnet[t].forward(hs[t].T[0]), ys[t].T[0])
assert_array_equal(
van.outputnet[t].forward(van.statenet[t].forward(
[xs[t].T[0], hs[t - 1].T[0]])), ys[t].T[0])
assert_array_equal(
van.outputnet[t].forward(van.statenet[t].forward(
[xs[t].T[0], van.state[t - 1]])), ys[t].T[0])
assert_array_equal(van.forward(xs[t].T[0]), ys[t].T[0])
assert_array_equal(soft.forward(ys[t].T[0]), ps[t].T[0])
# backward pass: compute gradients going backwards
dWxh, dWhh, dWhy = np.zeros_like(Wxh), np.zeros_like(Whh), np.zeros_like(
Why)
dbh, dby = np.zeros_like(bh), np.zeros_like(by)
dhnext = np.zeros_like(hs[0])
for t in reversed(xrange(len(inputs))):
dy = np.copy(ps[t])
dy[targets[
t]] -= 1 # backprop into y. see http://cs231n.github.io/neural-networks-case-study/#grad if confused here
dWhy += np.dot(dy, hs[t].T)
dby += dy
dh = np.dot(Why.T, dy) + dhnext # backprop into h
dhraw = (1 - hs[t] * hs[t]) * dh # backprop through tanh nonlinearity
dbh += dhraw
dWxh += np.dot(dhraw, xs[t].T)
dWhh += np.dot(dhraw, hs[t - 1].T)
#
# #Neg
# van.backward(negLog.dJdy_gradient(ps[t].T[0],to_one_hot_vect(targets[t],vocab_size)),opt)
# assert_array_almost_equal(van.outputnet[t].net.elements[0].elements[1].x,hs[t].T[0])
# assert_array_almost_equal(van.outputnet[t].net.elements[0].elements[0].elements[1].dW,dWhy)
# assert_array_almost_equal(van.outputnet[t].net.elements[0].elements[1].dW,dby.T[0])
# assert_array_almost_equal(van.outputnet[t].net.elements[0].elements[0].elements[1].W,Why)
# assert_array_almost_equal(van.outputnet[t].net.elements[0].elements[1].W,by.T[0])
#
#Cross
assert_array_equal(van.outputnet[t].net.elements[0].elements[1].x,
hs[t].T[0])
assert_array_equal(van.outputnet[t].net.forward(hs[t].T[0]),
ys[t].T[0])
assert_array_equal(
soft.forward(van.outputnet[t].net.forward(hs[t].T[0])), ps[t].T[0])
assert_array_equal(
soft.forward(van.outputnet[t].net.forward(hs[t].T[0])) -
to_one_hot_vect(targets[t], vocab_size), dy.T[0])
err = cross.dJdy_gradient(ys[t].T[0],
to_one_hot_vect(targets[t], vocab_size))
assert_array_equal(
soft.forward(van.outputnet[t].net.forward(hs[t].T[0])) -
to_one_hot_vect(targets[t], vocab_size), dy.T[0])
assert_array_equal(
ps[t].T[0] - to_one_hot_vect(targets[t], vocab_size), dy.T[0])
assert_array_equal(err, dy.T[0])
van.backward(err, opt)
assert_array_equal(
van.outputnet[t].net.elements[0].elements[1].W.get_dW(), dWhy)
assert_array_equal(
van.outputnet[t].net.elements[0].elements[1].W.get(), Why)
assert_array_equal(van.outputnet[t].net.elements[1].W.get_dW(),
dby.T[0])
assert_array_almost_equal(van.outputnet[t].net.elements[1].W.get(),
by.T[0])
#
assert_array_equal(
van.statenet[t].net.elements[0].elements[0].elements[1].W.get_dW(),
dWxh)
assert_array_equal(
van.statenet[t].net.elements[0].elements[1].elements[1].W.get_dW(),
dWhh)
assert_array_equal(
van.statenet[t].net.elements[0].elements[2].W.get_dW(), dbh.T[0])
assert_array_equal(
van.statenet[t].net.elements[0].elements[0].elements[1].W.get(),
Wxh)
assert_array_equal(
van.statenet[t].net.elements[0].elements[1].elements[1].W.get(),
Whh)
assert_array_equal(van.statenet[t].net.elements[0].elements[2].W.get(),
bh.T[0])
assert_array_equal(van.dJdh[t], dhnext.T[0])
dhnext = np.dot(Whh.T, dhraw)
opt.update_model()
trainer.learn_window(
vantr,
zip(to_hot_vect(inputs, vocab_size), to_hot_vect(targets, vocab_size)),
crosstr, opttr)
for dparam in [dWxh, dWhh, dWhy, dbh, dby]:
np.clip(dparam, -5, 5,
out=dparam) # clip to mitigate exploding gradients
return loss, dWxh, dWhh, dWhy, dbh, dby, hs[len(inputs) - 1]
def test_all(self):
n, p, epoch = 0, 0, -1
mWxh, mWhh, mWhy = np.zeros_like(Wxh), np.zeros_like(
Whh), np.zeros_like(Why)
mbh, mby = np.zeros_like(bh), np.zeros_like(
by) # memory variables for Adagrad
smooth_loss = -np.log(
1.0 / vocab_size) * seq_length # loss at iteration 0
while n <= 400:
print(n, p, epoch)
# prepare inputs (we're sweeping from left to right in steps seq_length long)
if p + seq_length + 1 > len(data) or n == 0:
van.clear_memory()
vantr.clear_memory()
hprev = np.zeros((hidden_size, 1)) # reset RNN memory
p = 0 # go from start of data
epoch += 1
# print (n,p,epoch)
inputs = [char_to_ix[ch] for ch in data[p:p + seq_length]]
targets = [char_to_ix[ch] for ch in data[p + 1:p + seq_length + 1]]
if epoch == epochs:
trainer2.learn_throughtime(
vantr2,
zip(to_hot_vect(inputs_all, vocab_size),
to_hot_vect(targets_all, vocab_size)),
CrossEntropyLoss(),
AdaGrad(learning_rate=learning_rate, clip=5), epochs)
assert_array_equal(
vantr2.statenet[0].net.elements[0].elements[0].elements[1].
W.get(), Wxh)
assert_array_equal(
vantr2.statenet[0].net.elements[0].elements[1].elements[1].
W.get(), Whh)
assert_array_equal(
vantr2.statenet[0].net.elements[0].elements[2].W.get(),
bh.T[0])
assert_array_equal(
vantr2.outputnet[0].net.elements[0].elements[1].W.get(),
Why)
assert_array_equal(vantr2.outputnet[0].net.elements[1].W.get(),
by.T[0])
txtvan = ''
x = to_one_hot_vect(inputs[0], vocab_size)
for i in range(200):
y = soft.forward(vantr2.forward(x))
txtvan += ix_to_char[np.argmax(
y)] #np.random.choice(range(vocab_size), p=y.ravel())]
x = to_one_hot_vect(np.argmax(y), vocab_size)
vantr2.clear_memory()
sample_ix = sample(hprev, inputs[0], 200)
txt = ''.join(ix_to_char[ix] for ix in sample_ix)
print '----\n %s \n %s \n----' % (txt, txtvan)
epoch = 0
# sample from the model now and then
# if n % epochs == 0:
# sample_ix = sample(hprev, inputs[0], 200)
# txt = ''.join(ix_to_char[ix] for ix in sample_ix)
# print '----\n %s \n %s ----' % (txt,txtvan )
# forward seq_length characters through the net and fetch gradient
loss, dWxh, dWhh, dWhy, dbh, dby, hprev = lossFun(
inputs, targets, hprev)
smooth_loss = smooth_loss * 0.999 + loss * 0.001
if n % epochs == 0:
print 'iter %d, loss: %f' % (n, smooth_loss) # print progress
# print 'iter %d, loss: %f' % (n, smooth_loss) # print progress
# perform parameter update with Adagrad
for param, dparam, mem in zip([Wxh, Whh, Why, bh, by],
[dWxh, dWhh, dWhy, dbh, dby],
[mWxh, mWhh, mWhy, mbh, mby]):
mem += dparam * dparam
param += -learning_rate * dparam / np.sqrt(
mem + 1e-8) # adagrad update
p += seq_length # move data pointer
n += 1 # iteration counter
def forward(self, x, update=False):
self.x = x
self.y = np.sign(self.W.dot(x) + np.random.normal(0, self.sigma))
self.e = self.delta * self.e + (1 - self.delta) * self.y * self.x
return utils.to_one_hot_vect(self.y / 2.0 + 1.0, 2)
def forward(self, x, update=False):
self.states_history.append(x)
self.command_history.append(self.command)
self.command = utils.to_one_hot_vect(np.argmax(self.model.forward(x)),
self.action_size)
return self.command
def forward(self, x, update=False):
self.x = np.argmax(x)
return utils.to_one_hot_vect(self.policy(self.x), self.action_size)
def functionPlot(*args):
lstm.save('lstm.net')
str = ''
x = to_one_hot_vect(char_to_ix['c'], vocab_size)
for i in range(200):
y = sm.forward(lstm.forward(x))
str += ix_to_char[np.random.choice(range(vocab_size), p=y.ravel())]
x = to_one_hot_vect(np.argmax(y), vocab_size)
print str
display.show(*args)
trainer = Trainer(show_training=True, show_function=functionPlot)
train = [to_one_hot_vect(char_to_ix[ch], vocab_size) for ch in data[0:-1]]
target = [to_one_hot_vect(char_to_ix[ch], vocab_size) for ch in data[1:]]
J, dJdy = trainer.learn_throughtime(lstm, zip(train, target),
CrossEntropyLoss(), opt, 1, window_size)
# J, dJdy = trainer.learn_window(
# v,
# zip(train[:5],target[:5]),
# NegativeLogLikelihoodLoss(),
# #CrossEntropyLoss(),
# AdaGrad(learning_rate=1e-1),
# )
# print J
# J, dJdy = trainer.learn_window(
# print v.forward(x)
# print v.backward(x)
epochs = 50
# opt = GradientDescent(learning_rate=0.01),
# opt = GradientDescentMomentum(learning_rate=0.01,momentum=0.5),
opt = AdaGrad(learning_rate=0.1) #,clip=100.0),
display = ShowTraining(
epochs_num=epochs
) #, weights_list = {'Wx':v.Wxh,'Whh':v.Whh,'Why':v.Why,'by':v.by,'bh':v.bh})
trainer = Trainer(show_training=True, show_function=display.show)
train = [to_one_hot_vect(char_to_ix[ch], vocab_size) for ch in data[0:-1]]
target = [to_one_hot_vect(char_to_ix[ch], vocab_size) for ch in data[1:]]
# J, dJdy = trainer.learn_window(
# v,
# zip(train[:5],target[:5]),
# NegativeLogLikelihoodLoss(),
# #CrossEntropyLoss(),
# AdaGrad(learning_rate=1e-1),
# )
# print J
# J, dJdy = trainer.learn_window(
# v,
# zip(train[:5],target[:5]),
# NegativeLogLikelihoodLoss(),
def forward(self, x, update = False):
self.x = x
return utils.to_one_hot_vect(np.random.choice(range(x.size), p=x.ravel()),x.size)
# from computationalgraph import Input, HotVect
# x = Input(['x','a'],'x')
# a = Input(['x','a'],'a')
# c=ComputationalGraphLayer(x*HotVect(a))
# c.forward([[10,10],1])
printer = ShowTraining(epochs_num=epochs, weights_list={'Q': W1})
trainer = Trainer(show_training=True, show_function=printer.show)
data_train = np.random.rand(1000, 2) * 5
train = []
for x in data_train:
out = Q_hat.forward(x)
train.append(
Q_hat.forward(x) * utils.to_one_hot_vect(np.argmax(out), out.size))
data_test = np.random.rand(1000, 2) * 5
test = []
for x in data_test:
out = Q_hat.forward(x)
test.append(
Q_hat.forward(x) * utils.to_one_hot_vect(np.argmax(out), out.size))
J_list, dJdy_list, J_test = trainer.learn(
model=Q,
train=zip(data_train, train),
test=zip(data_test, test),
# loss = NegativeLogLikelihoodLoss(),
# loss = CrossEntropyLoss(),
loss=SquaredLoss(),
