def __init__(self, n_in, n_out, block_size, activation, do_dropout=False,
reinforce_params="reinforce",
default_params="default"):
self.block_size = block_size
self.nblocks = n_out / block_size
self.do_dropout = do_dropout
assert n_out % block_size == 0
self.h = HiddenLayer(n_in, n_out, activation)
shared.bind(reinforce_params)
self.d = HiddenLayer(n_in, self.nblocks, T.nnet.sigmoid)
shared.bind(default_params)
def __init__(self,
n_in,
n_out,
block_size,
activation,
rate,
do_dropout=False):
self.rate = rate
self.block_size = block_size
self.nblocks = n_out / block_size
self.do_dropout = do_dropout
assert n_out % block_size == 0
self.h = HiddenLayer(n_in, n_out, activation)
class PolicyDropoutLayer:
def __init__(self, n_in, n_out, block_size, activation, do_dropout=False):
self.block_size = block_size
self.nblocks = n_out / block_size
self.do_dropout = do_dropout
assert n_out % block_size == 0
self.h = HiddenLayer(n_in, n_out, activation)
shared.bind("reinforce")
self.d = HiddenLayer(n_in, self.nblocks, T.nnet.sigmoid)
shared.bind("default")
def __call__(self, x, xmask=None):
probs = self.d(x) * 0.98 + 0.01
mask = srng.uniform(probs.shape) < probs
print xmask
mask.name = "mask!"
masked = self.h.activation(sparse_dot(x, xmask, self.h.W, mask, self.h.b, self.block_size))
if not "this is the equivalent computation in theano":
h = self.h(x)
if self.do_dropout:
h = h * (srng.uniform(h.shape) < 0.5)
h_r = h.reshape([h.shape[0], self.nblocks, self.block_size])
masked = h_r * mask.dimshuffle(0,1,'x')
masked = masked.reshape(h.shape)
self.sample_probs = T.prod(mask*probs+(1-probs)*(1-mask), axis=1)
self.probs = probs
return masked, mask
class PolicyDropoutLayer:
def __init__(self, n_in, n_out, block_size, activation, do_dropout=False,
reinforce_params="reinforce",
default_params="default"):
self.block_size = block_size
self.nblocks = n_out / block_size
self.do_dropout = do_dropout
assert n_out % block_size == 0
self.h = HiddenLayer(n_in, n_out, activation)
shared.bind(reinforce_params)
self.d = HiddenLayer(n_in, self.nblocks, T.nnet.sigmoid)
shared.bind(default_params)
def __call__(self, x, xmask=None):
probs = self.d(x) * 0.98 + 0.01
mask = srng.uniform(probs.shape) < probs
print xmask
mask.name = "mask!"
masked = self.h.activation(sparse_dot(x, xmask, self.h.W, mask, self.h.b, self.block_size))
if not "this is the equivalent computation in theano":
h = self.h(x)
if self.do_dropout:
h = h * (srng.uniform(h.shape) < 0.5)
h_r = h.reshape([h.shape[0], self.nblocks, self.block_size])
masked = h_r * mask.dimshuffle(0,1,'x')
masked = masked.reshape(h.shape)
self.sample_probs = T.prod(mask*probs+(1-probs)*(1-mask), axis=1)
self.probs = probs
return masked, mask
def __init__(self, n_in, n_out, block_size, activation, rate, do_dropout=False):
self.rate = rate
self.block_size = block_size
self.nblocks = n_out / block_size
self.do_dropout = do_dropout
assert n_out % block_size == 0
self.h = HiddenLayer(n_in, n_out, activation)
def __init__(self, n_in, n_out, block_size, activation, do_dropout=False):
self.block_size = block_size
self.nblocks = n_out / block_size
self.do_dropout = do_dropout
assert n_out % block_size == 0
self.h = HiddenLayer(n_in, n_out, activation)
shared.bind("reinforce")
self.d = HiddenLayer(n_in, self.nblocks, T.nnet.sigmoid)
shared.bind("default")
class PolicyDropoutLayer:
def __init__(self, n_in, n_out, block_size, activation, rate, do_dropout=False):
self.rate = rate
self.block_size = block_size
self.nblocks = n_out / block_size
self.do_dropout = do_dropout
assert n_out % block_size == 0
self.h = HiddenLayer(n_in, n_out, activation)
def __call__(self, x, xmask=None):
mask = srng.uniform((x.shape[0],self.nblocks)) < self.rate
masked = self.h.activation(sparse_dot(x, xmask, self.h.W, mask, self.h.b, self.block_size))
return masked, mask
class PolicyDropoutLayer:
def __init__(self,
n_in,
n_out,
block_size,
activation,
rate,
do_dropout=False):
self.rate = rate
self.block_size = block_size
self.nblocks = n_out / block_size
self.do_dropout = do_dropout
assert n_out % block_size == 0
self.h = HiddenLayer(n_in, n_out, activation)
def __call__(self, x, xmask=None):
mask = srng.uniform((x.shape[0], self.nblocks)) < self.rate
masked = self.h.activation(
sparse_dot(x, xmask, self.h.W, mask, self.h.b, self.block_size))
return masked, mask
def build_model(new_model=True):
momentum_epsilon = 0.9
nhidden = [64, 64]
L2reg = 0.001
vanilla = True
hyperparams = locals()
if new_model:
expid = str(uuid.uuid4())
import os
import os.path
code = file(os.path.abspath(__file__), 'r').read()
os.mkdir(expid)
os.chdir(expid)
file('code.py', 'w').write(code)
print expid
f = file("params.txt", 'w')
for i in hyperparams:
f.write("%s:%s\n" % (i, str(hyperparams[i])))
f.close()
params = []
shared.bind(params)
rect = lambda x: T.maximum(0, x)
act = T.tanh
model = StackModel([
HiddenLayer(32 * 32 * 3, nhidden[0], act),
Dropout(),
HiddenLayer(nhidden[0], nhidden[1], act),
Dropout(),
HiddenLayer(nhidden[-1], 10, T.nnet.softmax)
])
x = T.matrix()
y = T.ivector()
lr = T.scalar()
y_hat, = model(x)
loss = T.nnet.categorical_crossentropy(y_hat, y)
cost = T.sum(loss)
l2 = lambda x: sum([T.sum(i**2) for i in x])
updates = []
error = T.sum(T.neq(y_hat.argmax(axis=1), y))
nn_regularization = L2reg * l2(params)
grads = T.grad(cost + nn_regularization, params)
updates += gradient_descent(params, grads, lr)
learn = theano.function([x, y, lr], [cost, error],
updates=updates,
allow_input_downcast=True)
test = theano.function([x, y], [cost, error], allow_input_downcast=True)
return model, learn, test