def inv_prop(self, state_above):
if not isinstance(state_above, tuple):
expected_space = VectorSpace(self.output_space.get_total_dimension())
state_above = expected_space.format_as(state_above, self.output_space)
self.output_space.validate(state_above)
return tuple(layer.inv_prop(state) for layer,state in safe_zip(self.layers, state_above))
def test_np_format_as_vector2conv2D():
vector_space = VectorSpace(dim=8*8*3, sparse=False)
conv2d_space = Conv2DSpace(shape=(8,8), num_channels=3,
axes=('b','c',0,1))
data = np.arange(5*8*8*3).reshape(5, 8*8*3)
rval = vector_space.np_format_as(data, conv2d_space)
assert np.all(rval == data.reshape((5,3,8,8)))
def __init__(self, nvis, nhid, hidden_transition_model, irange=0.05,
non_linearity='sigmoid', use_ground_truth=True):
allowed_non_linearities = {'sigmoid': T.nnet.sigmoid,
'tanh': T.tanh}
self.nvis = nvis
self.nhid = nhid
self.hidden_transition_model = hidden_transition_model
self.use_ground_truth = use_ground_truth
self.alpha = sharedX(1)
self.alpha_decrease_rate = 0.999
assert non_linearity in allowed_non_linearities
self.non_linearity = allowed_non_linearities[non_linearity]
# Space initialization
self.input_space = VectorSpace(dim=self.nvis)
self.hidden_space = VectorSpace(dim=self.nhid)
self.output_space = VectorSpace(dim=1)
self.input_source = 'features'
self.target_source = 'targets'
# Features-to-hidden matrix
W_value = numpy.random.uniform(low=-irange, high=irange,
size=(self.nvis, self.nhid))
self.W = sharedX(W_value, name='W')
# Hidden biases
b_value = numpy.zeros(self.nhid)
self.b = sharedX(b_value, name='b')
# Hidden-to-out matrix
U_value = numpy.random.uniform(low=-irange, high=irange,
size=(self.nhid, 1))
self.U = sharedX(U_value, name='U')
# Output bias
c_value = numpy.zeros(1)
self.c = sharedX(c_value, name='c')
def test_vector_to_conv_c01b_invertible():
"""
Tests that the format_as methods between Conv2DSpace
and VectorSpace are invertible for the ('c', 0, 1, 'b')
axis format.
"""
rng = np.random.RandomState([2013, 5, 1])
batch_size = 3
rows = 4
cols = 5
channels = 2
conv = Conv2DSpace([rows, cols], channels = channels, axes = ('c', 0, 1, 'b'))
vec = VectorSpace(conv.get_total_dimension())
X = conv.make_batch_theano()
Y = conv.format_as(X, vec)
Z = vec.format_as(Y, conv)
A = vec.make_batch_theano()
B = vec.format_as(A, conv)
C = conv.format_as(B, vec)
f = function([X, A], [Z, C])
X = rng.randn(*(conv.get_origin_batch(batch_size).shape)).astype(X.dtype)
A = rng.randn(*(vec.get_origin_batch(batch_size).shape)).astype(A.dtype)
Z, C = f(X,A)
np.testing.assert_allclose(Z, X)
np.testing.assert_allclose(C, A)
def simulate(inputs, model):
space = VectorSpace(inputs.shape[1])
X = space.get_theano_batch()
Y = model.fprop(space.format_as(X, model.get_input_space()))
f = theano.function([X], Y)
result = []
for x in xrange(0, len(inputs), 100):
result.extend(f(inputs[x:x + 100]))
return result
def test_np_format_as_conv2d_vector_conv2d():
conv2d_space1 = Conv2DSpace(shape=(8, 8), num_channels=3,
axes=('c', 'b', 1, 0))
vector_space = VectorSpace(dim=8*8*3, sparse=False)
conv2d_space0 = Conv2DSpace(shape=(8, 8), num_channels=3,
axes=('b', 'c', 0, 1))
data = np.arange(5*8*8*3).reshape(5, 3, 8, 8)
vecval = conv2d_space0.np_format_as(data, vector_space)
rval1 = vector_space.np_format_as(vecval, conv2d_space1)
rval2 = conv2d_space0.np_format_as(data, conv2d_space1)
assert np.allclose(rval1, rval2)
nval = data.transpose(1, 0, 3, 2)
assert np.allclose(nval, rval1)
def test_np_format_as_vector2conv2D():
vector_space = VectorSpace(dim=8*8*3, sparse=False)
conv2d_space = Conv2DSpace(shape=(8, 8), num_channels=3,
axes=('b', 'c', 0, 1))
data = np.arange(5*8*8*3).reshape(5, 8*8*3)
rval = vector_space.np_format_as(data, conv2d_space)
# Get data in a Conv2DSpace with default axes
new_axes = conv2d_space.default_axes
axis_to_shape = {'b': 5, 'c': 3, 0: 8, 1: 8}
new_shape = tuple([axis_to_shape[ax] for ax in new_axes])
nval = data.reshape(new_shape)
# Then transpose
nval = nval.transpose(*[new_axes.index(ax) for ax in conv2d_space.axes])
assert np.all(rval == nval)
def __init__(self,
nvis,
bias_from_marginals = None):
"""
nvis: the dimension of the space
bias_from_marginals: a dataset, whose marginals are used to
initialize the visible biases
"""
self.__dict__.update(locals())
del self.self
# Don't serialize the dataset
del self.bias_from_marginals
self.space = VectorSpace(nvis)
self.input_space = self.space
origin = self.space.get_origin()
if bias_from_marginals is None:
init_bias = np.zeros((nvis,))
else:
X = bias_from_marginals.get_design_matrix()
assert X.max() == 1.
assert X.min() == 0.
assert not np.any( (X > 0.) * (X < 1.) )
mean = X.mean(axis=0)
mean = np.clip(mean, 1e-7, 1-1e-7)
init_bias = inverse_sigmoid_numpy(mean)
self.bias = sharedX(init_bias, 'visible_bias')
def set_input_space(self, space):
self.input_space = space
if not isinstance(space, Space):
raise TypeError("Expected Space, got "+
str(space)+" of type "+str(type(space)))
self.input_dim = space.get_total_dimension()
self.needs_reformat = not isinstance(space, VectorSpace)
self.desired_space = VectorSpace(self.input_dim)
if not self.needs_reformat:
assert self.desired_space == self.input_space
rng = self.dbm.rng
if self.irange is not None:
assert self.sparse_init is None
W = rng.uniform(-self.irange,self.irange, (self.input_dim,self.n_classes))
else:
assert self.sparse_init is not None
W = np.zeros((self.input_dim, self.n_classes))
for i in xrange(self.n_classes):
for j in xrange(self.sparse_init):
idx = rng.randint(0, self.input_dim)
while W[idx, i] != 0.:
idx = rng.randint(0, self.input_dim)
W[idx, i] = rng.randn()
self.W = sharedX(W, 'softmax_W' )
self._params = [ self.b, self.W ]
def _format_as(self, batch, space):
raise NotImplementedError()
if isinstance(space, CompositeSpace):
pos = 0
pieces = []
for component in space.components:
width = component.get_total_dimension()
subtensor = batch[:, pos:pos + width]
pos += width
formatted = VectorSpace(width).format_as(subtensor, component)
pieces.append(formatted)
return tuple(pieces)
if isinstance(space, Conv2DSpace):
if space.axes[0] != 'b':
raise NotImplementedError(
"Will need to reshape to ('b',*) then do a dimshuffle. Be sure to make this the inverse of space._format_as(x, self)"
)
dims = {
'b': batch.shape[0],
'c': space.num_channels,
0: space.shape[0],
1: space.shape[1]
}
shape = tuple([dims[elem] for elem in space.axes])
rval = batch.reshape(shape)
return rval
raise NotImplementedError(
"VectorSpace doesn't know how to format as " + str(type(space)))
def __init__(self, classes_number, which_set):
self.classes_number = classes_number
self.path = '/home/gortolan/MachineLearning/'
self.which_set = which_set
denseMatrix = pickle.load(
open(self.path + self.which_set + '_cons_small.pkl', "rb"))
self.x = denseMatrix.X
self.y = denseMatrix.y
X_space = VectorSpace(dim=273)
X_source = 'features'
Y_space = VectorSpace(dim=32)
Y_source = 'targets'
space = VectorSpace(X_space, Y_space)
source = VectorSpace(X_source, Y_source)
self.data_specs = (space, source)
super(TIMIT, self).__init__(X=self.x, y=self.y, y_labels=32)
def test_conditional_initialize_parameters():
"""
Conditional.initialize_parameters does the following:
* Set its input_space and ndim attributes
* Calls its MLP's set_mlp method
* Sets its MLP's input_space
* Validates its MLP
* Sets its params and param names
"""
mlp = MLP(layers=[Linear(layer_name='h', dim=5, irange=0.01,
max_col_norm=0.01)])
conditional = DummyConditional(mlp=mlp, name='conditional')
vae = DummyVAE()
conditional.set_vae(vae)
input_space = VectorSpace(dim=5)
conditional.initialize_parameters(input_space=input_space, ndim=5)
testing.assert_same_object(input_space, conditional.input_space)
testing.assert_equal(conditional.ndim, 5)
testing.assert_same_object(mlp.get_mlp(), conditional)
testing.assert_same_object(mlp.input_space, input_space)
mlp_params = mlp.get_params()
conditional_params = conditional.get_params()
assert all([mp in conditional_params for mp in mlp_params])
assert all([cp in mlp_params for cp in conditional_params])
def set_input_space(self, space):
self.input_space = space
if not isinstance(space, Space):
raise TypeError("Expected Space, got "+
str(space)+" of type "+str(type(space)))
self.input_dim = space.get_total_dimension()
self.needs_reformat = not isinstance(space, VectorSpace)
desired_dim = self.input_dim
self.desired_space = VectorSpace(desired_dim)
if not self.needs_reformat:
assert self.desired_space == self.input_space
rng = self.mlp.rng
self._params = []
V = np.zeros((self.n_classes, self.input_dim),dtype=np.float32)
self.V = sharedX(V, self.layer_name + "_V" )
U = np.identity( self.input_dim)
self.U = sharedX(U, self.layer_name + "_U")
Q = np.zeros((self.input_dim, self.input_dim),dtype=np.float32)
self.Q = sharedX(Q, self.layer_name + "_Q")
Ui = np.identity(self.input_dim,dtype=np.float32)
self.Ui = sharedX(Ui, self.layer_name + "_Ui")
self._params = [ self.U, self.Ui, self.V, self.Q]
def __init__(self, nvis, nhid):
super(AEModel, self).__init__()
self.nvis = nvis
self.nhid = nhid
self.W = sharedX(np.random.uniform(-1e-3, 1e-3, (nhid, nvis)),
name="W")
self.W_prime = self.W.T
self.theta = sharedX(np.zeros(nhid))
self.theta_prime = sharedX(np.zeros(nvis))
self._params = [self.W, self.theta, self.theta_prime]
self.input_space = VectorSpace(dim=nvis)
self.output_space = VectorSpace(dim=nhid)
def _build_output_space(self, space):
if isinstance(space, IndexSpace):
return VectorSpace(self.dim * space.dim)
if isinstance(space, CompositeSpace):
return CompositeSpace(
[self._build_output_space(c) for c in space.components])
assert False
def create_input_space(self):
ws = (self.ws * 2 + 1)
return CompositeSpace([
IndexSpace(max_labels=self.vocab_size, dim=ws),
IndexSpace(max_labels=self.total_feats, dim=self.feat_num),
VectorSpace(dim=self.extender_dim * ws)
])
def __init__(self,
nvis,
bias_from_marginals = None):
"""
nvis: the dimension of the space
bias_from_marginals: a dataset, whose marginals are used to
initialize the visible biases
"""
self.__dict__.update(locals())
del self.self
# Don't serialize the dataset
del self.bias_from_marginals
self.space = VectorSpace(nvis)
self.input_space = self.space
origin = self.space.get_origin()
if bias_from_marginals is None:
init_bias = np.zeros((nvis,))
else:
init_bias = init_tanh_bias_from_marginals(bias_from_marginals)
self.bias = sharedX(init_bias, 'visible_bias')
def test_fprop(self):
"""
Use an RNN without non-linearity to create the Mersenne numbers
(2 ** n - 1) to check whether fprop works correctly.
"""
rnn = RNN(input_space=SequenceSpace(VectorSpace(dim=1)),
layers=[
Recurrent(dim=1,
layer_name='recurrent',
irange=0.1,
indices=[-1],
nonlinearity=lambda x: x)
])
W, U, b = rnn.layers[0].get_params()
W.set_value([[1]])
U.set_value([[2]])
X_data, X_mask = rnn.get_input_space().make_theano_batch()
y_hat = rnn.fprop((X_data, X_mask))
seq_len = 20
X_data_vals = np.ones((seq_len, seq_len, 1))
X_mask_vals = np.triu(np.ones((seq_len, seq_len)))
f = function([X_data, X_mask], y_hat, allow_input_downcast=True)
np.testing.assert_allclose(2**np.arange(1, seq_len + 1) - 1,
f(X_data_vals, X_mask_vals).flatten())
def __init__(self,
nvis,
nhid,
init_bias_hid,
init_beta,
init_scale,
min_beta,
fixed_point_orthogonalize=False,
censor_beta_norms=True):
self.__dict__.update(locals())
del self.self
self.rng = np.random.RandomState([2012, 11, 13])
self.scale = sharedX(np.zeros((nhid, )) + init_scale)
self.scale.name = 'scale'
self.b = sharedX(np.zeros((nhid, )) + init_bias_hid)
self.b.name = 'b'
self.beta = sharedX(np.zeros(nvis, ) + init_beta)
self.beta.name = 'beta'
self.W = sharedX(
random_ortho_columns(nvis, nhid, self.beta.get_value(), self.rng))
self.W.name = 'W'
self._params = [self.scale, self.W, self.b, self.beta]
self.input_space = VectorSpace(nvis)
def test_set_get_weights_Softmax():
"""
Tests setting and getting weights for Softmax layer.
"""
num_classes = 2
dim = 3
conv_dim = [3, 4, 5]
# VectorSpace input space
layer = Softmax(num_classes, 's', irange=.1)
softmax_mlp = MLP(layers=[layer], input_space=VectorSpace(dim=dim))
vec_weights = np.random.randn(dim, num_classes).astype(config.floatX)
layer.set_weights(vec_weights)
assert np.allclose(layer.W.get_value(), vec_weights)
layer.W.set_value(vec_weights)
assert np.allclose(layer.get_weights(), vec_weights)
# Conv2DSpace input space
layer = Softmax(num_classes, 's', irange=.1)
softmax_mlp = MLP(layers=[layer],
input_space=Conv2DSpace(shape=(conv_dim[0], conv_dim[1]),
num_channels=conv_dim[2]))
conv_weights = np.random.randn(conv_dim[0], conv_dim[1], conv_dim[2],
num_classes).astype(config.floatX)
layer.set_weights(conv_weights.reshape(np.prod(conv_dim), num_classes))
assert np.allclose(layer.W.get_value(),
conv_weights.reshape(np.prod(conv_dim), num_classes))
layer.W.set_value(conv_weights.reshape(np.prod(conv_dim), num_classes))
assert np.allclose(layer.get_weights_topo(),
np.transpose(conv_weights, axes=(3, 0, 1, 2)))
def __init__(self, mlp, n_classes = None, input_source='features', input_space=None, scale = False):
"""
Parameters
----------
mlp: Pylearn2 MLP class
The frame based classifier
"""
if n_classes is None:
if hasattr(mlp .layers[-1], 'dim'):
self.n_classes = mlp.layers[-1].dim
elif hasattr(mlp.layers[-1], 'n_classes'):
self.n_classes = mlp.layers[-1].n_classes
else:
raise ValueError("n_classes was not provided and couldn't be infered from the mlp's last layer")
else:
self.n_classes = n_classes
self.mlp = mlp
self.scale = scale
self.input_source = input_source
assert isinstance(input_space, FaceTubeSpace)
self.input_space = input_space
self.input_size = (input_space.shape[0]
* input_space.shape[1]
* input_space.num_channels)
self.output_space = VectorSpace(dim=7)
#rng = self.mlp.rng
#self.W = theano.shared(rng.uniform(size=(n_classes, n_classes, n_classes)).astype(config.floatX))
#self.W.name = 'crf_w'
self.init_transition_matrix()
self.name = 'crf'
def __init__(self,
nvis,
bias_from_marginals = None):
"""
nvis: the dimension of the space
bias_from_marginals: a dataset, whose marginals are used to
initialize the visible biases
"""
self.__dict__.update(locals())
del self.self
# Don't serialize the dataset
del self.bias_from_marginals
self.space = VectorSpace(nvis)
self.input_space = self.space
origin = self.space.get_origin()
if bias_from_marginals is None:
init_bias = np.zeros((nvis,))
else:
# data is in [-1, 1], but want biases for a sigmoid
init_bias = init_sigmoid_bias_from_array(bias_from_marginals.X / 2. + 0.5)
# init_bias =
self.boltzmann_bias = sharedX(init_bias, 'visible_bias')
def set_input_space(self, space):
""" Note: this resets parameters! """
self.input_space = space
if isinstance(space, VectorSpace):
self.requires_reformat = False
self.input_dim = space.dim
else:
self.requires_reformat = True
self.input_dim = space.get_total_dimension()
self.desired_space = VectorSpace(self.input_dim)
self.output_space = VectorSpace(self.dim)
rng = self.dbm.rng
if self.irange is not None:
assert self.sparse_init is None
W = rng.uniform(-self.irange,
self.irange,
(self.input_dim, self.dim)) * \
(rng.uniform(0.,1., (self.input_dim, self.dim))
< self.include_prob)
else:
assert self.sparse_init is not None
W = np.zeros((self.input_dim, self.dim))
W *= self.sparse_stdev
W = sharedX(W)
W.name = self.layer_name + '_W'
self.transformer = MatrixMul(W)
W ,= self.transformer.get_params()
assert W.name is not None
def get_output_space(self):
"""
.. todo::
WRITEME
"""
return VectorSpace(self.num_classes)
def __init__(self,
corruptor,
nvis,
nhid,
act_enc,
act_dec,
tied_weights=False,
irange=1e-3,
rng=9001):
"""
.. todo::
WRITEME
"""
# sampling dot only supports tied weights
assert tied_weights == True
self.names_to_del = set()
super(SparseDenoisingAutoencoder,
self).__init__(corruptor,
nvis,
nhid,
act_enc,
act_dec,
tied_weights=tied_weights,
irange=irange,
rng=rng)
# this step is crucial to save loads of space because w_prime is never used in
# training the sparse da.
del self.w_prime
self.input_space = VectorSpace(nvis, sparse=True)
def __init__(self,
load_path=None,
from_scipy_sparse_dataset=None,
zipped_npy=True):
self.load_path = load_path
self.y = None
if self.load_path is not None:
if zipped_npy is True:
logger.info('... loading sparse data set from a zip npy file')
self.X = scipy.sparse.csr_matrix(numpy.load(
gzip.open(load_path)),
dtype=floatX)
else:
logger.info('... loading sparse data set from a npy file')
self.X = scipy.sparse.csr_matrix(numpy.load(load_path).item(),
dtype=floatX)
else:
logger.info('... building from given sparse dataset')
self.X = from_scipy_sparse_dataset
if not scipy.sparse.issparse(from_scipy_sparse_dataset):
msg = "from_scipy_sparse_dataset is not sparse : %s" \
% type(self.X)
raise TypeError(msg)
X_space = VectorSpace(dim=self.X.shape[1], sparse=True)
self.X_space = X_space
space = self.X_space
source = 'features'
self._iter_data_specs = (space, source)
self.data_specs = (space, source)
def __init__(self,
k,
nvis,
convergence_th=1e-6,
max_iter=None,
verbose=False):
"""
Parameters in conf:
:type k: int
:param k: number of clusters.
:type convergence_th: float
:param convergence_th: threshold of distance to clusters under which
kmeans stops iterating.
:type max_iter: int
:param max_iter: maximum number of iterations. Defaults to infinity.
"""
Block.__init__(self)
Model.__init__(self)
self.input_space = VectorSpace(nvis)
self.k = k
self.convergence_th = convergence_th
if max_iter:
if max_iter < 0:
raise Exception('KMeans init: max_iter should be positive.')
self.max_iter = max_iter
else:
self.max_iter = float('inf')
self.verbose = verbose
def set_input_space(self, space):
self.input_space = space
if not isinstance(space, Space):
raise TypeError("Expected Space, got "+
str(space)+" of type "+str(type(space)))
self.input_dim = space.get_total_dimension()
self.needs_reformat = not isinstance(space, VectorSpace)
if self.no_affine:
desired_dim = self.n_classes
assert self.input_dim == desired_dim
else:
desired_dim = self.input_dim
self.desired_space = VectorSpace(desired_dim)
if not self.needs_reformat:
assert self.desired_space == self.input_space
rng = self.mlp.rng
if self.irange is not None:
assert self.istdev is None
assert self.sparse_init is None
W = rng.uniform(-self.irange,self.irange, (self.input_dim,self.n_groups,self.n_classes))
elif self.istdev is not None:
assert self.sparse_init is None
W = rng.randn(self.input_dim,self.n_groups,self.n_classes) * self.istdev
else:
raise NotImplementedError()
self.W = sharedX(W, 'softmax_W' )
self._params = [ self.b, self.W ]
def test_np_format_as_vector2conv2d():
vector_space = VectorSpace(dim=8 * 8 * 3, sparse=False)
conv2d_space = Conv2DSpace(shape=(8, 8),
num_channels=3,
axes=('b', 'c', 0, 1))
data = np.arange(5 * 8 * 8 * 3).reshape(5, 8 * 8 * 3)
rval = vector_space.np_format_as(data, conv2d_space)
# Get data in a Conv2DSpace with default axes
new_axes = conv2d_space.default_axes
axis_to_shape = {'b': 5, 'c': 3, 0: 8, 1: 8}
new_shape = tuple([axis_to_shape[ax] for ax in new_axes])
nval = data.reshape(new_shape)
# Then transpose
nval = nval.transpose(*[new_axes.index(ax) for ax in conv2d_space.axes])
assert np.all(rval == nval)
def set_input_space(self, space):
""" Note: this resets parameters! """
self.input_space = space
if isinstance(space, VectorSpace):
self.requires_reformat = False
self.input_dim = space.dim
else:
self.requires_reformat = True
self.input_dim = space.get_total_dimension()
self.desired_space = VectorSpace(self.input_dim)
if not (self.detector_layer_dim % self.pool_size == 0):
raise ValueError("detector_layer_dim = %d, pool_size = %d. Should be divisible but remainder is %d" %
(self.detector_layer_dim, self.pool_size, self.detector_layer_dim % self.pool_size))
self.h_space = VectorSpace(self.detector_layer_dim)
self.pool_layer_dim = self.detector_layer_dim / self.pool_size
self.output_space = VectorSpace(self.pool_layer_dim)
rng = self.dbm.rng
if self.irange is not None:
assert self.sparse_init is None
W = rng.uniform(-self.irange,
self.irange,
(self.input_dim, self.detector_layer_dim)) * \
(rng.uniform(0.,1., (self.input_dim, self.detector_layer_dim))
< self.include_prob)
else:
assert self.sparse_init is not None
W = np.zeros((self.input_dim, self.detector_layer_dim))
for i in xrange(self.detector_layer_dim):
for j in xrange(self.sparse_init):
idx = rng.randint(0, self.input_dim)
while W[idx, i] != 0:
idx = rng.randint(0, self.input_dim)
W[idx, i] = rng.randn()
W = sharedX(W)
W.name = self.layer_name + '_W'
self.transformer = MatrixMul(W)
W ,= self.transformer.get_params()
assert W.name is not None
def test_np_format_as_conv2D_vector_conv2D():
conv2d_space1 = Conv2DSpace(shape=(8, 8),
num_channels=3,
axes=('c', 'b', 1, 0))
vector_space = VectorSpace(dim=8 * 8 * 3, sparse=False)
conv2d_space0 = Conv2DSpace(shape=(8, 8),
num_channels=3,
axes=('b', 'c', 0, 1))
data = np.arange(5 * 8 * 8 * 3).reshape(5, 3, 8, 8)
vecval = conv2d_space0.np_format_as(data, vector_space)
rval1 = vector_space.np_format_as(vecval, conv2d_space1)
rval2 = conv2d_space0.np_format_as(data, conv2d_space1)
assert np.allclose(rval1, rval2)
nval = data.transpose(1, 0, 3, 2)
assert np.allclose(nval, rval1)
def set_input_space(self, space):
""" Note: this resets parameters! """
self.input_space = space
if isinstance(space, VectorSpace):
self.requires_reformat = False
self.input_dim = space.dim
else:
self.requires_reformat = True
self.input_dim = space.get_total_dimension()
self.desired_space = VectorSpace(self.input_dim)
self.output_space = VectorSpace(self.dim)
rng = self.dbm.rng
if self.irange is not None:
assert self.sparse_init is None
W = rng.uniform(-self.irange,
self.irange,
(self.input_dim, self.dim)) * \
(rng.uniform(0.,1., (self.input_dim, self.dim))
< self.include_prob)
else:
assert self.sparse_init is not None
W = np.zeros((self.input_dim, self.dim))
W *= self.sparse_stdev
W = sharedX(W)
W.name = self.layer_name + '_W'
self.W = W
if self.sampling_b_stdev is not None:
self.noisy_sampling_b = sharedX(np.zeros((self.dbm.batch_size, self.dim)))
self.layer_below.noisy_sampling_b = sharedX(np.zeros((self.dbm.batch_size, self.layer_below.nvis)))
if self.sampling_W_stdev is not None:
self.noisy_sampling_W = sharedX(np.zeros((self.input_dim, self.dim)), 'noisy_sampling_W')
updates = OrderedDict()
updates[self.boltzmann_b] = self.boltzmann_b
updates[self.W] = self.W
updates[self.layer_below.boltzmann_bias] = self.layer_below.boltzmann_bias
self.censor_updates(updates)
f = function([], updates=updates)
f()
def __init__(self, nvis, nhid, nclasses):
super(MLP, self).__init__()
self.nvis, self.nhid, self.nclasses = nvis, nhid, nclasses
self.W = sharedX(numpy.random.normal(scale=0.01,
size=(self.nvis, self.nhid)),
name='W')
self.b = sharedX(numpy.zeros(self.nhid), name='b')
self.V = sharedX(numpy.random.normal(scale=0.01,
size=(self.nhid, self.nclasses)),
name='V')
self.c = sharedX(numpy.zeros(self.nclasses), name='c')
self._params = [self.W, self.b, self.V, self.c]
self.input_space = VectorSpace(dim=self.nvis)
self.output_space = VectorSpace(dim=self.nclasses)
def __init__(self, dim, layer_name, operation, nonlinearity=None, **kwargs):
super(GlueLayer, self).__init__(**kwargs)
self.dim = dim
self.layer_name = layer_name
self.nonlinearity = nonlinearity
self.operation = operation
self.output_space = VectorSpace(self.dim)
self._params = []
def __init__(self,
n_classes,
layer_name,
irange=None,
sparse_init=None,
W_lr_scale=None):
if isinstance(W_lr_scale, str):
W_lr_scale = float(W_lr_scale)
self.__dict__.update(locals())
del self.self
assert isinstance(n_classes, int)
self.output_space = VectorSpace(n_classes)
self.b = sharedX(np.zeros((n_classes, )), name='softmax_b')
def model1():
#pdb.set_trace()
# train set X has dim (60,000, 784), y has dim (60,000, 10)
train_set = MNIST(which_set='train', one_hot=True)
# test set X has dim (10,000, 784), y has dim (10,000, 10)
valid_set = MNIST(which_set='test', one_hot=True)
test_set = MNIST(which_set='test', one_hot=True)
#import pdb
#pdb.set_trace()
#print train_set.X.shape[1]
# =====<Create the MLP Model>=====
h2_layer = NoisyRELU(layer_name='h1',
sparse_init=15,
noise_factor=5,
dim=1000,
desired_active_rate=0.2,
bias_factor=20,
max_col_norm=1)
#h2_layer = RectifiedLinear(layer_name='h2', dim=100, sparse_init=15, max_col_norm=1)
#print h1_layer.get_params()
#h2 = RectifiedLinear(layer_name='h2', dim=500, sparse_init=15, max_col_norm=1)
y_layer = Softmax(layer_name='y', n_classes=10, irange=0., max_col_norm=1)
mlp = MLP(batch_size=200,
input_space=VectorSpace(dim=train_set.X.shape[1]),
layers=[h2_layer, y_layer])
# =====<Create the SGD algorithm>=====
sgd = SGD(init_momentum=0.1,
learning_rate=0.01,
monitoring_dataset={'valid': valid_set},
cost=MethodCost('cost_from_X'),
termination_criterion=MonitorBased(
channel_name='valid_y_misclass', prop_decrease=0.001, N=50))
#sgd.setup(model=mlp, dataset=train_set)
# =====<Extensions>=====
ext = [MomentumAdjustor(start=1, saturate=10, final_momentum=0.9)]
# =====<Create Training Object>=====
save_path = './mlp_model1.pkl'
train_obj = Train(dataset=train_set,
model=mlp,
algorithm=sgd,
extensions=ext,
save_path=save_path,
save_freq=0)
#train_obj.setup_extensions()
#import pdb
#pdb.set_trace()
train_obj.main_loop()
# =====<Run the training>=====
'''
def test_broadcastable():
v = VectorSpace(5).make_theano_batch(batch_size=1)
np.testing.assert_(v.broadcastable[0])
c = Conv2DSpace((5, 5), channels=3,
axes=['c', 0, 1, 'b']).make_theano_batch(batch_size=1)
np.testing.assert_(c.broadcastable[-1])
d = Conv2DSpace((5, 5), channels=3,
axes=['b', 0, 1, 'c']).make_theano_batch(batch_size=1)
np.testing.assert_(d.broadcastable[0])
def __init__(self, nvis, nclasses):
super(LogisticRegression, self).__init__()
# Number of input nodes
self.nvis = nvis
# Number of output nodes
self.nclasses = nclasses
W_value = np.random.uniform(size=(self.nvis, self.nclasses))
self.W = sharedX(W_value, 'W') # sharedX formats for GPUs
b_value = np.zeros(self.nclasses)
self.b = sharedX(b_value, 'b')
self._params = [self.W, self.b]
self.input_space = VectorSpace(dim=self.nvis)
self.output_space = VectorSpace(dim=self.nclasses)
def __init__(self,
mlp,
input_condition_space,
condition_distribution,
noise_dim=100,
*args,
**kwargs):
super(ConditionalGenerator, self).__init__(mlp, *args, **kwargs)
self.noise_dim = noise_dim
self.noise_space = VectorSpace(dim=self.noise_dim)
self.condition_space = input_condition_space
self.condition_distribution = condition_distribution
self.input_space = CompositeSpace(
[self.noise_space, self.condition_space])
self.mlp.set_input_space(self.input_space)
def __init__(self, n_classes, layer_name, irange = None,
istdev = None,
sparse_init = None):
self.__dict__.update(locals())
del self.self
self.output_space = VectorSpace(n_classes)
self.b = sharedX(np.zeros((n_classes,)), name = 'hingeloss_b')
def __init__(self, nvis, nclasses):
super(LogisticRegressionLayer, self).__init__()
assert nvis >= 0, "Number of visible units must be non-negative"
self.input_space = VectorSpace(nvis)
self.output_space = VectorSpace(nclasses)
assert nclasses >= 0, "Number of classes must be non-negative"
self.nvis = nvis
self.nclasses = nclasses
# initialize with 0 the weights W as a matrix of shape (nvis, nclasses)
self.W = sharedX(numpy.zeros((nvis, nclasses)), name='W', borrow=True)
# initialize the biases b as a vector of nclasses 0s
self.b = sharedX(numpy.zeros((nclasses,)), name='b', borrow=True)
# parameters of the model
self._params = [self.W, self.b]
def __init__(self,
n_classes,
layer_name,
C=0.1,
irange=None,
istdev=None,
sparse_init=None,
W_lr_scale=None,
b_lr_scale=None,
max_row_norm=None,
no_affine=False,
max_col_norm=None,
init_bias_target_marginals=None,
binary_target_dim=None):
super(L2SquareHinge, self).__init__()
if isinstance(W_lr_scale, str):
W_lr_scale = float(W_lr_scale)
self.__dict__.update(locals())
del self.self
del self.init_bias_target_marginals
assert isinstance(n_classes, py_integer_types)
if binary_target_dim is not None:
assert isinstance(binary_target_dim, py_integer_types)
self._has_binary_target = True
self._target_space = IndexSpace(dim=binary_target_dim,
max_labels=n_classes)
else:
self._has_binary_target = False
self.output_space = VectorSpace(n_classes)
self.b = sharedX(np.zeros((n_classes, )), name='hinge_b')
if init_bias_target_marginals:
y = init_bias_target_marginals.y
if init_bias_target_marginals.y_labels is None:
marginals = y.mean(axis=0)
else:
# compute class frequencies
if np.max(y.shape) != np.prod(y.shape):
raise AssertionError("Use of "
"`init_bias_target_marginals` "
"requires that each example has "
"a single label.")
marginals = np.bincount(y.flat) / float(y.shape[0])
assert marginals.ndim == 1
b = pseudoinverse_softmax_numpy(marginals).astype(self.b.dtype)
assert b.ndim == 1
assert b.dtype == self.b.dtype
self.b.set_value(b)
else:
assert init_bias_target_marginals is None
def test_finitedataset_source_check():
"""
Check that the FiniteDatasetIterator returns sensible
errors when there is a missing source in the dataset.
"""
dataset = DenseDesignMatrix(X=np.random.rand(20,15).astype(theano.config.floatX),
y=np.random.rand(20,5).astype(theano.config.floatX))
assert_raises(ValueError,
dataset.iterator,
mode='sequential',
batch_size=5,
data_specs=(VectorSpace(15),'featuresX'))
try:
dataset.iterator(mode='sequential',
batch_size=5,
data_specs=(VectorSpace(15),'featuresX'))
except ValueError as e:
assert 'featuresX' in str(e)
def set_topological_view(self, V, axes=('b', 0, 1, 'c')):
"""
Sets the dataset to represent V, where V is a batch
of topological views of examples.
Parameters
----------
V : ndarray
An array containing a design matrix representation of training \
examples.
axes : WRITEME
.. todo::
Why is this parameter named 'V'?
"""
assert not np.any(np.isnan(V))
rows = V.shape[axes.index(0)]
cols = V.shape[axes.index(1)]
channels = V.shape[axes.index('c')]
self.view_converter = DefaultViewConverter([rows, cols, channels],
axes=axes)
self.X = self.view_converter.topo_view_to_design_mat(V)
# self.X_topo_space stores a "default" topological space that
# will be used only when self.iterator is called without a
# data_specs, and with "topo=True", which is deprecated.
self.X_topo_space = self.view_converter.topo_space
assert not np.any(np.isnan(self.X))
# Update data specs
X_space = VectorSpace(dim=self.X.shape[1])
X_source = 'features'
if self.y is None:
space = X_space
source = X_source
else:
y_space = VectorSpace(dim=self.y.shape[-1])
y_source = 'targets'
space = CompositeSpace((X_space, y_space))
source = (X_source, y_source)
self.data_specs = (space, source)
self.X_space = X_space
self._iter_data_specs = (X_space, X_source)
def __init__(self,
alpha_list=[1.4],
beta_list=[0.3],
init_state_list=[numpy.array([0, 0])],
num_samples=1000,
frame_length=1,
rng=None):
# Validate parameters and set member variables
self.alpha_list = alpha_list
self.beta_list = beta_list
if num_samples <= 0:
raise ValueError("num_samples must be positive.")
self.num_samples = num_samples
self.num_examples = len(alpha_list)
self.frame_length = frame_length
self.init_state_list = init_state_list
# Initialize RNG
if rng is None:
self.rng = numpy.random.RandomState(self._default_seed)
else:
self.rng = numpy.random.RandomState(rng)
X, y = self._generate_data()
self.data = (X, y)
# DataSpecs
features_space = VectorSpace(dim=2 * self.frame_length)
features_source = 'features'
targets_space = VectorSpace(dim=2)
targets_source = 'targets'
space = CompositeSpace([features_space, targets_space])
source = tuple([features_source, targets_source])
self.data_specs = (space, source)
# Defaults for iterators
self._iter_mode = resolve_iterator_class('shuffled_sequential')
self._iter_data_specs = (CompositeSpace(
(features_space, targets_space)), (features_source,
targets_source))
def __init__(self, shape, axes=None):
"""
The arguments describe how the data is laid out in the design matrix.
Parameters
----------
shape : tuple
A tuple of 4 ints, describing the shape of each datum.
This is the size of each axis in <axes>, excluding the 'b' axis.
axes : tuple
A tuple of the following elements in any order:
'b' batch axis
's' stereo axis
0 image axis 0 (row)
1 image axis 1 (column)
'c' channel axis
"""
shape = tuple(shape)
if not all(isinstance(s, int) for s in shape):
raise TypeError("Shape must be a tuple/list of ints")
if len(shape) != 4:
raise ValueError("Shape array needs to be of length 4, got %s." %
shape)
datum_axes = list(axes)
datum_axes.remove('b')
if shape[datum_axes.index('s')] != 2:
raise ValueError("Expected 's' axis to have size 2, got %d.\n"
" axes: %s\n"
" shape: %s" %
(shape[datum_axes.index('s')],
axes,
shape))
self.shape = shape
self.set_axes(axes)
def make_conv2d_space(shape, axes):
shape_axes = list(axes)
shape_axes.remove('b')
image_shape = tuple(shape[shape_axes.index(axis)]
for axis in (0, 1))
conv2d_axes = list(axes)
conv2d_axes.remove('s')
return Conv2DSpace(shape=image_shape,
num_channels=shape[shape_axes.index('c')],
axes=conv2d_axes,
dtype=None)
conv2d_space = make_conv2d_space(shape, axes)
self.topo_space = CompositeSpace((conv2d_space, conv2d_space))
self.storage_space = VectorSpace(dim=numpy.prod(shape))
def __init__(self,
layer_name,
num_gates,
irange = 0.05,
routing_protocol = 'nearest'
):
self.__dict__.update(locals())
del self.self
self.output_space = VectorSpace(self.num_gates)
def __init__(self, mlp, input_condition_space, condition_distribution, noise_dim=100, *args, **kwargs):
super(ConditionalGenerator, self).__init__(mlp, *args, **kwargs)
self.noise_dim = noise_dim
self.noise_space = VectorSpace(dim=self.noise_dim)
self.condition_space = input_condition_space
self.condition_distribution = condition_distribution
self.input_space = CompositeSpace([self.noise_space, self.condition_space])
self.mlp.set_input_space(self.input_space)
def get_weights_topo(self):
"""
Returns a topological view of the weights, the first half
corresponds to wxf and the second half to wyf.
Returns
-------
weights : ndarray
Same as the return value of `get_weights` but formatted as a 4D
tensor with the axes being (hidden/factor units, rows, columns,
channels).The the number of channels is either 1 or 3
(because they will be visualized as grayscale or RGB color).
At the moment the function only supports factors whose sqrt
is exact.
"""
if not isinstance(self.input_space.components[0], Conv2DSpace) or not isinstance(
self.input_space.components[1], Conv2DSpace
):
raise NotImplementedError()
wxf = self.wxf.get_value(borrow=False).T
wyf = self.wyf.get_value(borrow=False).T
convx = self.input_space.components[0]
convy = self.input_space.components[1]
vecx = VectorSpace(self.nvisx)
vecy = VectorSpace(self.nvisy)
wxf_view = vecx.np_format_as(
wxf, Conv2DSpace(convx.shape, num_channels=convx.num_channels, axes=("b", 0, 1, "c"))
)
wyf_view = vecy.np_format_as(
wyf, Conv2DSpace(convy.shape, num_channels=convy.num_channels, axes=("b", 0, 1, "c"))
)
h = int(numpy.ceil(numpy.sqrt(self.nfac)))
new_weights = numpy.zeros(
(wxf_view.shape[0] * 2, wxf_view.shape[1], wxf_view.shape[2], wxf_view.shape[3]), dtype=wxf_view.dtype
)
t = 0
while t < (self.nfac // h):
filter_pair = numpy.concatenate((wxf_view[h * t : h * (t + 1), ...], wyf_view[h * t : h * (t + 1), ...]), 0)
new_weights[h * 2 * t : h * 2 * (t + 1), ...] = filter_pair
t += 1
return new_weights
def __init__(self, n_classes, layer_name, irange = None,
sparse_init = None, W_lr_scale = None):
if isinstance(W_lr_scale, str):
W_lr_scale = float(W_lr_scale)
self.__dict__.update(locals())
del self.self
assert isinstance(n_classes, int)
self.output_space = VectorSpace(n_classes)
self.b = sharedX( np.zeros((n_classes,)), name = 'softmax_b')
def set_input_space(self, space):
self.input_space = space
if not isinstance(space, Space):
raise TypeError("Expected Space, got "+
str(space)+" of type "+str(type(space)))
self.input_dim = space.get_total_dimension()
self.needs_reformat = not isinstance(space, VectorSpace)
if self.no_affine:
desired_dim = self.n_classes
assert self.input_dim == desired_dim
else:
desired_dim = self.input_dim
self.desired_space = VectorSpace(desired_dim)
if not self.needs_reformat:
assert self.desired_space == self.input_space
rng = self.mlp.rng
if self.no_affine:
self._params = []
else:
if self.irange is not None:
assert self.istdev is None
assert self.sparse_init is None
W_cluster = rng.uniform(-self.irange,self.irange, (self.input_dim, self.n_clusters))
W_class = rng.uniform(-self.irange,self.irange, (self.n_clusters, self.input_dim, self.n_classes))
elif self.istdev is not None:
assert self.sparse_init is None
W_cluster = rng.randn(self.input_dim, self.n_clusters) * self.istdev
W_class = rng.randn(self.n_clusters, self.input_dim, self.n_classes) * self.istdev
else:
raise NotImplementedError()
# set the extra dummy weights to 0
for key in self.clusters_scope.keys():
#print key
#should probably be reverse
W_class[int(key), :, :self.clusters_scope[key]] = 0.
self.W_class = sharedX(W_class, 'softmax_W_class' )
self.W_cluster = sharedX(W_cluster, 'softmax_W_cluster' )
self._params = [self.b_class, self.W_class, self.b_cluster, self.W_cluster]
class VectorSpaceConverter(mlp.Layer):
def __init__(self, layer_name):
self.layer_name = layer_name
self._params = []
def set_input_space(self, space):
self.input_space = space
self.output_space = VectorSpace(space.get_total_dimension())
def fprop(self, state_below):
return self.input_space.format_as(state_below, self.output_space)
def inv_prop(self, state_above):
return self.output_space.format_as(state_above, self.input_space)
def get_weight_decay(self, coeff):
return 0.0
def get_l1_weight_decay(self, coeff):
return 0.0
def __init__(self, shape, axes=None):
shape = tuple(shape)
if not all(isinstance(s, int) for s in shape):
raise TypeError("Shape must be a tuple/list of ints")
if len(shape) != 4:
raise ValueError("Shape array needs to be of length 4, got %s." %
shape)
datum_axes = list(axes)
datum_axes.remove('b')
if shape[datum_axes.index('s')] != 2:
raise ValueError("Expected 's' axis to have size 2, got %d.\n"
" axes: %s\n"
" shape: %s" %
(shape[datum_axes.index('s')],
axes,
shape))
self.shape = shape
self.set_axes(axes)
def make_conv2d_space(shape, axes):
shape_axes = list(axes)
shape_axes.remove('b')
image_shape = tuple(shape[shape_axes.index(axis)]
for axis in (0, 1))
conv2d_axes = list(axes)
conv2d_axes.remove('s')
return Conv2DSpace(shape=image_shape,
num_channels=shape[shape_axes.index('c')],
axes=conv2d_axes)
conv2d_space = make_conv2d_space(shape, axes)
self.topo_space = CompositeSpace((conv2d_space, conv2d_space))
self.storage_space = VectorSpace(dim=numpy.prod(shape))
def __init__(self, nvis = None, nhid = None,
vis_space = None,
hid_space = None,
transformer = None,
irange=0.5, rng=None, init_bias_vis = None,
init_bias_vis_marginals = None, init_bias_hid=0.0,
base_lr = 1e-3, anneal_start = None, nchains = 100, sml_gibbs_steps = 1,
random_patches_src = None,
monitor_reconstruction = False):
"""
Construct an RBM object.
Parameters
----------
nvis : int
Number of visible units in the model.
(Specifying this implies that the model acts on a vector,
i.e. it sets vis_space = pylearn2.space.VectorSpace(nvis) )
nhid : int
Number of hidden units in the model.
(Specifying this implies that the model acts on a vector)
vis_space:
A pylearn2.space.Space object describing what kind of vector
space the RBM acts on. Don't specify if you used nvis / hid
hid_space:
A pylearn2.space.Space object describing what kind of vector
space the RBM's hidden units live in. Don't specify if you used
nvis / nhid
init_bias_vis_marginals: either None, or a Dataset to use to initialize
the visible biases to the inverse sigmoid of the data marginals
irange : float, optional
The size of the initial interval around 0 for weights.
rng : RandomState object or seed
NumPy RandomState object to use when initializing parameters
of the model, or (integer) seed to use to create one.
init_bias_vis : array_like, optional
Initial value of the visible biases, broadcasted as necessary.
init_bias_hid : array_like, optional
initial value of the hidden biases, broadcasted as necessary.
monitor_reconstruction : if True, will request a monitoring channel to monitor
reconstruction error
random_patches_src: Either None, or a Dataset from which to draw random patches
in order to initialize the weights. Patches will be multiplied by irange
Parameters for default SML learning rule:
base_lr : the base learning rate
anneal_start : number of steps after which to start annealing on a 1/t schedule
nchains: number of negative chains
sml_gibbs_steps: number of gibbs steps to take per update
"""
Model.__init__(self)
Block.__init__(self)
if init_bias_vis_marginals is not None:
assert init_bias_vis is None
X = init_bias_vis_marginals.X
assert X.min() >= 0.0
assert X.max() <= 1.0
marginals = X.mean(axis=0)
#rescale the marginals a bit to avoid NaNs
init_bias_vis = inverse_sigmoid_numpy(.01 + .98 * marginals)
if init_bias_vis is None:
init_bias_vis = 0.0
if rng is None:
# TODO: global rng configuration stuff.
rng = numpy.random.RandomState(1001)
self.rng = rng
if vis_space is None:
#if we don't specify things in terms of spaces and a transformer,
#assume dense matrix multiplication and work off of nvis, nhid
assert hid_space is None
assert transformer is None or isinstance(transformer,MatrixMul)
assert nvis is not None
assert nhid is not None
if transformer is None:
if random_patches_src is None:
W = rng.uniform(-irange, irange, (nvis, nhid))
else:
if hasattr(random_patches_src, '__array__'):
W = irange * random_patches_src.T
assert W.shape == (nvis, nhid)
else:
#assert type(irange) == type(0.01)
#assert irange == 0.01
W = irange * random_patches_src.get_batch_design(nhid).T
self.transformer = MatrixMul( sharedX(
W,
name='W',
borrow=True
)
)
else:
self.transformer = transformer
self.vis_space = VectorSpace(nvis)
self.hid_space = VectorSpace(nhid)
else:
assert hid_space is not None
assert transformer is not None
assert nvis is None
assert nhid is None
self.vis_space = vis_space
self.hid_space = hid_space
self.transformer = transformer
try:
b_vis = self.vis_space.get_origin()
b_vis += init_bias_vis
except ValueError:
raise ValueError("bad shape or value for init_bias_vis")
self.bias_vis = sharedX(b_vis, name='bias_vis', borrow=True)
try:
b_hid = self.hid_space.get_origin()
b_hid += init_bias_hid
except ValueError:
raise ValueError('bad shape or value for init_bias_hid')
self.bias_hid = sharedX(b_hid, name='bias_hid', borrow=True)
self.random_patches_src = random_patches_src
self.register_names_to_del(['random_patches_src'])
self.__dict__.update(nhid=nhid, nvis=nvis)
self._params = safe_union(self.transformer.get_params(), [self.bias_vis, self.bias_hid])
self.base_lr = base_lr
self.anneal_start = anneal_start
self.nchains = nchains
self.sml_gibbs_steps = sml_gibbs_steps
class RBM(Block, Model):
"""
A base interface for RBMs, implementing the binary-binary case.
"""
def __init__(self, nvis = None, nhid = None,
vis_space = None,
hid_space = None,
transformer = None,
irange=0.5, rng=None, init_bias_vis = None,
init_bias_vis_marginals = None, init_bias_hid=0.0,
base_lr = 1e-3, anneal_start = None, nchains = 100, sml_gibbs_steps = 1,
random_patches_src = None,
monitor_reconstruction = False):
"""
Construct an RBM object.
Parameters
----------
nvis : int
Number of visible units in the model.
(Specifying this implies that the model acts on a vector,
i.e. it sets vis_space = pylearn2.space.VectorSpace(nvis) )
nhid : int
Number of hidden units in the model.
(Specifying this implies that the model acts on a vector)
vis_space:
A pylearn2.space.Space object describing what kind of vector
space the RBM acts on. Don't specify if you used nvis / hid
hid_space:
A pylearn2.space.Space object describing what kind of vector
space the RBM's hidden units live in. Don't specify if you used
nvis / nhid
init_bias_vis_marginals: either None, or a Dataset to use to initialize
the visible biases to the inverse sigmoid of the data marginals
irange : float, optional
The size of the initial interval around 0 for weights.
rng : RandomState object or seed
NumPy RandomState object to use when initializing parameters
of the model, or (integer) seed to use to create one.
init_bias_vis : array_like, optional
Initial value of the visible biases, broadcasted as necessary.
init_bias_hid : array_like, optional
initial value of the hidden biases, broadcasted as necessary.
monitor_reconstruction : if True, will request a monitoring channel to monitor
reconstruction error
random_patches_src: Either None, or a Dataset from which to draw random patches
in order to initialize the weights. Patches will be multiplied by irange
Parameters for default SML learning rule:
base_lr : the base learning rate
anneal_start : number of steps after which to start annealing on a 1/t schedule
nchains: number of negative chains
sml_gibbs_steps: number of gibbs steps to take per update
"""
Model.__init__(self)
Block.__init__(self)
if init_bias_vis_marginals is not None:
assert init_bias_vis is None
X = init_bias_vis_marginals.X
assert X.min() >= 0.0
assert X.max() <= 1.0
marginals = X.mean(axis=0)
#rescale the marginals a bit to avoid NaNs
init_bias_vis = inverse_sigmoid_numpy(.01 + .98 * marginals)
if init_bias_vis is None:
init_bias_vis = 0.0
if rng is None:
# TODO: global rng configuration stuff.
rng = numpy.random.RandomState(1001)
self.rng = rng
if vis_space is None:
#if we don't specify things in terms of spaces and a transformer,
#assume dense matrix multiplication and work off of nvis, nhid
assert hid_space is None
assert transformer is None or isinstance(transformer,MatrixMul)
assert nvis is not None
assert nhid is not None
if transformer is None:
if random_patches_src is None:
W = rng.uniform(-irange, irange, (nvis, nhid))
else:
if hasattr(random_patches_src, '__array__'):
W = irange * random_patches_src.T
assert W.shape == (nvis, nhid)
else:
#assert type(irange) == type(0.01)
#assert irange == 0.01
W = irange * random_patches_src.get_batch_design(nhid).T
self.transformer = MatrixMul( sharedX(
W,
name='W',
borrow=True
)
)
else:
self.transformer = transformer
self.vis_space = VectorSpace(nvis)
self.hid_space = VectorSpace(nhid)
else:
assert hid_space is not None
assert transformer is not None
assert nvis is None
assert nhid is None
self.vis_space = vis_space
self.hid_space = hid_space
self.transformer = transformer
try:
b_vis = self.vis_space.get_origin()
b_vis += init_bias_vis
except ValueError:
raise ValueError("bad shape or value for init_bias_vis")
self.bias_vis = sharedX(b_vis, name='bias_vis', borrow=True)
try:
b_hid = self.hid_space.get_origin()
b_hid += init_bias_hid
except ValueError:
raise ValueError('bad shape or value for init_bias_hid')
self.bias_hid = sharedX(b_hid, name='bias_hid', borrow=True)
self.random_patches_src = random_patches_src
self.register_names_to_del(['random_patches_src'])
self.__dict__.update(nhid=nhid, nvis=nvis)
self._params = safe_union(self.transformer.get_params(), [self.bias_vis, self.bias_hid])
self.base_lr = base_lr
self.anneal_start = anneal_start
self.nchains = nchains
self.sml_gibbs_steps = sml_gibbs_steps
def get_input_dim(self):
if not isinstance(self.vis_space, VectorSpace):
raise TypeError("Can't describe "+str(type(self.vis_space))+" as a dimensionality number.")
return self.vis_space.dim
def get_output_dim(self):
if not isinstance(self.hid_space, VectorSpace):
raise TypeError("Can't describe "+str(type(self.hid_space))+" as a dimensionality number.")
return self.hid_space.dim
def get_input_space(self):
return self.vis_space
def get_output_space(self):
return self.hid_space
def get_params(self):
return [param for param in self._params]
def get_weights(self, borrow=False):
weights ,= self.transformer.get_params()
return weights.get_value(borrow=borrow)
def get_weights_topo(self):
return self.transformer.get_weights_topo()
def get_weights_format(self):
return ['v', 'h']
def get_monitoring_channels(self, data):
V = data
theano_rng = RandomStreams(42)
#TODO: re-enable this in the case where self.transformer
#is a matrix multiply
#norms = theano_norms(self.weights)
H = self.mean_h_given_v(V)
h = H.mean(axis=0)
return { 'bias_hid_min' : T.min(self.bias_hid),
'bias_hid_mean' : T.mean(self.bias_hid),
'bias_hid_max' : T.max(self.bias_hid),
'bias_vis_min' : T.min(self.bias_vis),
'bias_vis_mean' : T.mean(self.bias_vis),
'bias_vis_max': T.max(self.bias_vis),
'h_min' : T.min(h),
'h_mean': T.mean(h),
'h_max' : T.max(h),
#'W_min' : T.min(self.weights),
#'W_max' : T.max(self.weights),
#'W_norms_min' : T.min(norms),
#'W_norms_max' : T.max(norms),
#'W_norms_mean' : T.mean(norms),
'reconstruction_error' : self.reconstruction_error(V, theano_rng) }
def get_monitoring_data_specs(self):
"""
Get the data_specs describing the data for get_monitoring_channel.
This implementation returns specification corresponding to unlabeled
inputs.
"""
return (self.get_input_space(), self.get_input_source())
def ml_gradients(self, pos_v, neg_v):
"""
Get the contrastive gradients given positive and negative phase
visible units.
Parameters
----------
pos_v : tensor_like
Theano symbolic representing a minibatch on the visible units,
with the first dimension indexing training examples and the second
indexing data dimensions (usually actual training data).
neg_v : tensor_like
Theano symbolic representing a minibatch on the visible units,
with the first dimension indexing training examples and the second
indexing data dimensions (usually reconstructions of the data or
sampler particles from a persistent Markov chain).
Returns
-------
grads : list
List of Theano symbolic variables representing gradients with
respect to model parameters, in the same order as returned by
`params()`.
Notes
-----
`pos_v` and `neg_v` need not have the same first dimension, i.e.
minibatch size.
"""
# taking the mean over each term independently allows for different
# mini-batch sizes in the positive and negative phase.
ml_cost = (self.free_energy_given_v(pos_v).mean() -
self.free_energy_given_v(neg_v).mean())
grads = tensor.grad(ml_cost, self.get_params(),
consider_constant=[pos_v, neg_v])
return grads
def train_batch(self, dataset, batch_size):
""" A default learning rule based on SML """
self.learn_mini_batch(dataset.get_batch_design(batch_size))
return True
def learn_mini_batch(self, X):
""" A default learning rule based on SML """
if not hasattr(self, 'learn_func'):
self.redo_theano()
rval = self.learn_func(X)
return rval
def redo_theano(self):
""" Compiles the theano function for the default learning rule """
init_names = dir(self)
minibatch = tensor.matrix()
optimizer = _SGDOptimizer(self, self.base_lr, self.anneal_start)
sampler = sampler = BlockGibbsSampler(self, 0.5 + np.zeros((self.nchains, self.get_input_dim())), self.rng,
steps= self.sml_gibbs_steps)
updates = training_updates(visible_batch=minibatch, model=self,
sampler=sampler, optimizer=optimizer)
self.learn_func = theano.function([minibatch], updates=updates)
final_names = dir(self)
self.register_names_to_del([name for name in final_names if name not in init_names])
def gibbs_step_for_v(self, v, rng):
"""
Do a round of block Gibbs sampling given visible configuration
Parameters
----------
v : tensor_like
Theano symbolic representing the hidden unit states for a batch of
training examples (or negative phase particles), with the first
dimension indexing training examples and the second indexing data
dimensions.
rng : RandomStreams object
Random number generator to use for sampling the hidden and visible
units.
Returns
-------
v_sample : tensor_like
Theano symbolic representing the new visible unit state after one
round of Gibbs sampling.
locals : dict
Contains the following auxiliary state as keys (all symbolics
except shape tuples):
* `h_mean`: the returned value from `mean_h_given_v`
* `h_mean_shape`: shape tuple indicating the size of `h_mean` and
`h_sample`
* `h_sample`: the stochastically sampled hidden units
* `v_mean_shape`: shape tuple indicating the shape of `v_mean` and
`v_sample`
* `v_mean`: the returned value from `mean_v_given_h`
* `v_sample`: the stochastically sampled visible units
"""
h_mean = self.mean_h_given_v(v)
assert h_mean.type.dtype == v.type.dtype
# For binary hidden units
# TODO: factor further to extend to other kinds of hidden units
# (e.g. spike-and-slab)
h_sample = rng.binomial(size = h_mean.shape, n = 1 , p = h_mean, dtype=h_mean.type.dtype)
assert h_sample.type.dtype == v.type.dtype
# v_mean is always based on h_sample, not h_mean, because we don't
# want h transmitting more than one bit of information per unit.
v_mean = self.mean_v_given_h(h_sample)
assert v_mean.type.dtype == v.type.dtype
v_sample = self.sample_visibles([v_mean], v_mean.shape, rng)
assert v_sample.type.dtype == v.type.dtype
return v_sample, locals()
def sample_visibles(self, params, shape, rng):
"""
Stochastically sample the visible units given hidden unit
configurations for a set of training examples.
Parameters
----------
params : list
List of the necessary parameters to sample :math:`p(v|h)`. In the
case of a binary-binary RBM this is a single-element list
containing the symbolic representing :math:`p(v|h)`, as returned
by `mean_v_given_h`.
Returns
-------
vprime : tensor_like
Theano symbolic representing stochastic samples from :math:`p(v|h)`
"""
v_mean = params[0]
return as_floatX(rng.uniform(size=shape) < v_mean)
def input_to_h_from_v(self, v):
"""
Compute the affine function (linear map plus bias) that serves as
input to the hidden layer in an RBM.
Parameters
----------
v : tensor_like or list of tensor_likes
Theano symbolic (or list thereof) representing the one or several
minibatches on the visible units, with the first dimension indexing
training examples and the second indexing data dimensions.
Returns
-------
a : tensor_like or list of tensor_likes
Theano symbolic (or list thereof) representing the input to each
hidden unit for each training example.
"""
if isinstance(v, tensor.Variable):
return self.bias_hid + self.transformer.lmul(v)
else:
return [self.input_to_h_from_v(vis) for vis in v]
def input_to_v_from_h(self, h):
"""
Compute the affine function (linear map plus bias) that serves as
input to the visible layer in an RBM.
Parameters
----------
h : tensor_like or list of tensor_likes
Theano symbolic (or list thereof) representing the one or several
minibatches on the hidden units, with the first dimension indexing
training examples and the second indexing data dimensions.
Returns
-------
a : tensor_like or list of tensor_likes
Theano symbolic (or list thereof) representing the input to each
visible unit for each row of h.
"""
if isinstance(h, tensor.Variable):
return self.bias_vis + self.transformer.lmul_T(h)
else:
return [self.input_to_v_from_h(hid) for hid in h]
def upward_pass(self, v):
"""
wrapper around mean_h_given_v method. Called when RBM is accessed
by mlp.HiddenLayer.
"""
return self.mean_h_given_v(v)
def mean_h_given_v(self, v):
"""
Compute the mean activation of the hidden units given visible unit
configurations for a set of training examples.
Parameters
----------
v : tensor_like or list of tensor_likes
Theano symbolic (or list thereof) representing the hidden unit
states for a batch (or several) of training examples, with the
first dimension indexing training examples and the second indexing
data dimensions.
Returns
-------
h : tensor_like or list of tensor_likes
Theano symbolic (or list thereof) representing the mean
(deterministic) hidden unit activations given the visible units.
"""
if isinstance(v, tensor.Variable):
return nnet.sigmoid(self.input_to_h_from_v(v))
else:
return [self.mean_h_given_v(vis) for vis in v]
def mean_v_given_h(self, h):
"""
Compute the mean activation of the visibles given hidden unit
configurations for a set of training examples.
Parameters
----------
h : tensor_like or list of tensor_likes
Theano symbolic (or list thereof) representing the hidden unit
states for a batch (or several) of training examples, with the
first dimension indexing training examples and the second indexing
hidden units.
Returns
-------
vprime : tensor_like or list of tensor_likes
Theano symbolic (or list thereof) representing the mean
(deterministic) reconstruction of the visible units given the
hidden units.
"""
if isinstance(h, tensor.Variable):
return nnet.sigmoid(self.input_to_v_from_h(h))
else:
return [self.mean_v_given_h(hid) for hid in h]
def free_energy_given_v(self, v):
"""
Calculate the free energy of a visible unit configuration by
marginalizing over the hidden units.
Parameters
----------
v : tensor_like
Theano symbolic representing the hidden unit states for a batch of
training examples, with the first dimension indexing training
examples and the second indexing data dimensions.
Returns
-------
f : tensor_like
1-dimensional tensor (vector) representing the free energy
associated with each row of v.
"""
sigmoid_arg = self.input_to_h_from_v(v)
return (-tensor.dot(v, self.bias_vis) -
nnet.softplus(sigmoid_arg).sum(axis=1))
def free_energy(self, V):
return self.free_energy_given_v(V)
def free_energy_given_h(self, h):
"""
Calculate the free energy of a hidden unit configuration by
marginalizing over the visible units.
Parameters
----------
h : tensor_like
Theano symbolic representing the hidden unit states, with the
first dimension indexing training examples and the second
indexing data dimensions.
Returns
-------
f : tensor_like
1-dimensional tensor (vector) representing the free energy
associated with each row of v.
"""
sigmoid_arg = self.input_to_v_from_h(h)
return (-tensor.dot(h, self.bias_hid) -
nnet.softplus(sigmoid_arg).sum(axis=1))
def __call__(self, v):
"""
Forward propagate (symbolic) input through this module, obtaining
a representation to pass on to layers above.
This just aliases the `mean_h_given_v()` function for syntactic
sugar/convenience.
"""
return self.mean_h_given_v(v)
def reconstruction_error(self, v, rng):
"""
Compute the mean-squared error (mean over examples, sum over units)
across a minibatch after a Gibbs
step starting from the training data.
Parameters
----------
v : tensor_like
Theano symbolic representing the hidden unit states for a batch of
training examples, with the first dimension indexing training
examples and the second indexing data dimensions.
rng : RandomStreams object
Random number generator to use for sampling the hidden and visible
units.
Returns
-------
mse : tensor_like
0-dimensional tensor (essentially a scalar) indicating the mean
reconstruction error across the minibatch.
Notes
-----
The reconstruction used to assess error samples only the hidden
units. For the visible units, it uses the conditional mean.
No sampling of the visible units is done, to reduce noise in the estimate.
"""
sample, _locals = self.gibbs_step_for_v(v, rng)
return ((_locals['v_mean'] - v) ** 2).sum(axis=1).mean()
class StereoViewConverter(object):
"""
Converts stereo image data between two formats:
#. A dense design matrix, one stereo pair per row (`VectorSpace`)
#. An image pair (`CompositeSpace` of two `Conv2DSpace`)
The arguments describe how the data is laid out in the design matrix.
Parameters
----------
shape: tuple
A tuple of 4 ints, describing the shape of each datum. This is the size
of each axis in `<axes>`, excluding the `b` axis.
axes : tuple
Tuple of the following elements in any order:
* 'b' : batch axis
* 's' : stereo axis
* 0 : image axis 0 (row)
* 1 : image axis 1 (column)
* 'c' : channel axis
"""
def __init__(self, shape, axes=None):
shape = tuple(shape)
if not all(isinstance(s, int) for s in shape):
raise TypeError("Shape must be a tuple/list of ints")
if len(shape) != 4:
raise ValueError("Shape array needs to be of length 4, got %s." %
shape)
datum_axes = list(axes)
datum_axes.remove('b')
if shape[datum_axes.index('s')] != 2:
raise ValueError("Expected 's' axis to have size 2, got %d.\n"
" axes: %s\n"
" shape: %s" %
(shape[datum_axes.index('s')],
axes,
shape))
self.shape = shape
self.set_axes(axes)
def make_conv2d_space(shape, axes):
shape_axes = list(axes)
shape_axes.remove('b')
image_shape = tuple(shape[shape_axes.index(axis)]
for axis in (0, 1))
conv2d_axes = list(axes)
conv2d_axes.remove('s')
return Conv2DSpace(shape=image_shape,
num_channels=shape[shape_axes.index('c')],
axes=conv2d_axes)
conv2d_space = make_conv2d_space(shape, axes)
self.topo_space = CompositeSpace((conv2d_space, conv2d_space))
self.storage_space = VectorSpace(dim=numpy.prod(shape))
def get_formatted_batch(self, batch, space):
"""
.. todo::
WRITEME
"""
return self.storage_space.np_format_as(batch, space)
def design_mat_to_topo_view(self, design_mat):
"""
Called by DenseDesignMatrix.get_formatted_view(), get_batch_topo()
"""
return self.storage_space.np_format_as(design_mat, self.topo_space)
def design_mat_to_weights_view(self, design_mat):
"""
Called by DenseDesignMatrix.get_weights_view()
"""
return self.design_mat_to_topo_view(design_mat)
def topo_view_to_design_mat(self, topo_batch):
"""
Used by `DenseDesignMatrix.set_topological_view()` and
`DenseDesignMatrix.get_design_mat()`.
"""
return self.topo_space.np_format_as(topo_batch, self.storage_space)
def view_shape(self):
"""
.. todo::
WRITEME
"""
return self.shape
def weights_view_shape(self):
"""
.. todo::
WRITEME
"""
return self.view_shape()
def set_axes(self, axes):
"""
.. todo::
WRITEME
"""
axes = tuple(axes)
if len(axes) != 5:
raise ValueError("Axes must have 5 elements; got %s" % str(axes))
for required_axis in ('b', 's', 0, 1, 'c'):
if required_axis not in axes:
raise ValueError("Axes must contain 'b', 's', 0, 1, and 'c'. "
"Got %s." % str(axes))
if axes.index('b') != 0:
raise ValueError("The 'b' axis must come first (axes = %s)." %
str(axes))
def get_batchless_axes(axes):
axes = list(axes)
axes.remove('b')
return tuple(axes)
if hasattr(self, 'axes'):
# Reorders the shape vector to match the new axis ordering.
assert hasattr(self, 'shape')
old_axes = get_batchless_axes(self.axes)
new_axes = get_batchless_axes(axes)
new_shape = tuple(self.shape[old_axes.index(a)] for a in new_axes)
self.shape = new_shape
self.axes = axes
class Softmax(Layer):
def __init__(self, n_classes, layer_name, irange = None,
istdev = None,
sparse_init = None, W_lr_scale = None,
b_lr_scale = None, max_row_norm = None):
"""
"""
if isinstance(W_lr_scale, str):
W_lr_scale = float(W_lr_scale)
self.__dict__.update(locals())
del self.self
assert isinstance(n_classes, int)
self.output_space = VectorSpace(n_classes)
self.b = sharedX( np.zeros((n_classes,)), name = 'softmax_b')
def get_lr_scalers(self):
rval = OrderedDict()
if self.W_lr_scale is not None:
assert isinstance(self.W_lr_scale, float)
rval[self.W] = self.W_lr_scale
if not hasattr(self, 'b_lr_scale'):
self.b_lr_scale = None
if self.b_lr_scale is not None:
assert isinstance(self.b_lr_scale, float)
rval[self.b] = self.b_lr_scale
return rval
def get_monitoring_channels_from_state(self, state, target=None):
mx = state.max(axis=1)
rval = OrderedDict([
('mean_max_class' , mx.mean()),
('max_max_class' , mx.max()),
('min_max_class' , mx.min())
])
if target is not None:
y_hat = T.argmax(state, axis=1)
y = T.argmax(target, axis=1)
misclass = T.neq(y, y_hat).mean()
misclass = T.cast(misclass, config.floatX)
rval['misclass'] = misclass
return rval
def set_input_space(self, space):
self.input_space = space
if not isinstance(space, Space):
raise TypeError("Expected Space, got "+
str(space)+" of type "+str(type(space)))
self.input_dim = space.get_total_dimension()
self.needs_reformat = not isinstance(space, VectorSpace)
self.desired_space = VectorSpace(self.input_dim)
if not self.needs_reformat:
assert self.desired_space == self.input_space
rng = self.mlp.rng
if self.irange is not None:
assert self.istdev is None
assert self.sparse_init is None
W = rng.uniform(-self.irange,self.irange, (self.input_dim,self.n_classes))
elif self.istdev is not None:
assert self.sparse_init is None
W = rng.randn(self.input_dim, self.n_classes) * self.istdev
else:
assert self.sparse_init is not None
W = np.zeros((self.input_dim, self.n_classes))
for i in xrange(self.n_classes):
for j in xrange(self.sparse_init):
idx = rng.randint(0, self.input_dim)
while W[idx, i] != 0.:
idx = rng.randint(0, self.input_dim)
W[idx, i] = rng.randn()
self.W = sharedX(W, 'softmax_W' )
self._params = [ self.b, self.W ]
def get_weights_topo(self):
if not isinstance(self.input_space, Conv2DSpace):
raise NotImplementedError()
desired = self.W.get_value().T
ipt = self.desired_space.format_as(desired, self.input_space)
rval = Conv2DSpace.convert_numpy(ipt, self.input_space.axes, ('b', 0, 1, 'c'))
return rval
def get_weights(self):
if not isinstance(self.input_space, VectorSpace):
raise NotImplementedError()
return self.W.get_value()
def set_weights(self, weights):
self.W.set_value(weights)
def set_biases(self, biases):
self.b.set_value(biases)
def get_biases(self):
return self.b.get_value()
def get_weights_format(self):
return ('v', 'h')
def fprop(self, state_below):
self.input_space.validate(state_below)
if self.needs_reformat:
state_below = self.input_space.format_as(state_below, self.desired_space)
for value in get_debug_values(state_below):
if value.shape[0] != self.mlp.batch_size:
raise ValueError("state_below should have batch size "+str(self.dbm.batch_size)+" but has "+str(value.shape[0]))
self.desired_space.validate(state_below)
assert self.W.ndim == 2
assert state_below.ndim == 2
b = self.b
Z = T.dot(state_below, self.W) + b
rval = T.nnet.softmax(Z)
for value in get_debug_values(rval):
assert value.shape[0] == self.mlp.batch_size
return rval
def cost(self, Y, Y_hat):
"""
Y must be one-hot binary. Y_hat is a softmax estimate.
of Y. Returns negative log probability of Y under the Y_hat
distribution.
"""
assert hasattr(Y_hat, 'owner')
owner = Y_hat.owner
assert owner is not None
op = owner.op
if isinstance(op, Print):
assert len(owner.inputs) == 1
Y_hat, = owner.inputs
owner = Y_hat.owner
op = owner.op
assert isinstance(op, T.nnet.Softmax)
z ,= owner.inputs
assert z.ndim == 2
z = z - z.max(axis=1).dimshuffle(0, 'x')
log_prob = z - T.log(T.exp(z).sum(axis=1).dimshuffle(0, 'x'))
# we use sum and not mean because this is really one variable per row
log_prob_of = (Y * log_prob).sum(axis=1)
assert log_prob_of.ndim == 1
rval = log_prob_of.mean()
return - rval
def get_weight_decay(self, coeff):
if isinstance(coeff, str):
coeff = float(coeff)
assert isinstance(coeff, float) or hasattr(coeff, 'dtype')
return coeff * T.sqr(self.W).sum()
def censor_updates(self, updates):
if self.max_row_norm is not None:
W = self.W
if W in updates:
updated_W = updates[W]
row_norms = T.sqrt(T.sum(T.sqr(updated_W), axis=1))
desired_norms = T.clip(row_norms, 0, self.max_row_norm)
updates[W] = updated_W * (desired_norms / (1e-7 + row_norms)).dimshuffle(0, 'x')
def test_np_format_as_vector2vector():
vector_space_initial = VectorSpace(dim=8*8*3, sparse=False)
vector_space_final = VectorSpace(dim=8*8*3, sparse=False)
data = np.arange(5*8*8*3).reshape(5, 8*8*3)
rval = vector_space_initial.np_format_as(data, vector_space_final)
assert np.all(rval == data)
class ConditionalGenerator(Generator):
def __init__(self, mlp, input_condition_space, condition_distribution, noise_dim=100, *args, **kwargs):
super(ConditionalGenerator, self).__init__(mlp, *args, **kwargs)
self.noise_dim = noise_dim
self.noise_space = VectorSpace(dim=self.noise_dim)
self.condition_space = input_condition_space
self.condition_distribution = condition_distribution
self.input_space = CompositeSpace([self.noise_space, self.condition_space])
self.mlp.set_input_space(self.input_space)
def sample_and_noise(
self, conditional_data, default_input_include_prob=1.0, default_input_scale=1.0, all_g_layers=False
):
"""
Retrieve a sample (and the noise used to generate the sample)
conditioned on some input data.
Parameters
----------
conditional_data: member of self.condition_space
A minibatch of conditional data to feedforward.
default_input_include_prob: float
WRITEME
default_input_scale: float
WRITEME
all_g_layers: boolean
If true, return all generator layers in `other_layers` slot
of this method's return value. (Otherwise returns `None` in
this slot.)
Returns
-------
net_output: 3-tuple
Tuple of the form `(sample, noise, other_layers)`.
"""
if isinstance(conditional_data, int):
conditional_data = self.condition_distribution.sample(conditional_data)
num_samples = conditional_data.shape[0]
noise = self.get_noise((num_samples, self.noise_dim))
# TODO necessary?
formatted_noise = self.noise_space.format_as(noise, self.noise_space)
# Build inputs: concatenate noise with conditional data
inputs = (formatted_noise, conditional_data)
# Feedforward
# if all_g_layers:
# rval = self.mlp.dropout_fprop(inputs, default_input_include_prob=default_input_include_prob,
# default_input_scale=default_input_scale, return_all=all_g_layers)
# other_layers, rval = rval[:-1], rval[-1]
# else:
rval = self.mlp.dropout_fprop(
inputs, default_input_include_prob=default_input_include_prob, default_input_scale=default_input_scale
)
# other_layers = None
return rval, formatted_noise, conditional_data, None # , other_layers
def sample(self, conditional_data, **kwargs):
sample, _, _, _ = self.sample_and_noise(conditional_data, **kwargs)
return sample
def get_monitoring_channels(self, data):
if data is None:
m = 100
conditional_data = self.condition_distribution.sample(m)
else:
_, conditional_data = data
m = conditional_data.shape[0]
noise = self.get_noise((m, self.noise_dim))
rval = OrderedDict()
sampled_data = (noise, conditional_data)
try:
rval.update(self.mlp.get_monitoring_channels((sampled_data, None)))
except Exception:
warnings.warn("something went wrong with generator.mlp's monitoring channels")
if self.monitor_ll:
rval["ll"] = T.cast(self.ll(data, self.ll_n_samples, self.ll_sigma), theano.config.floatX).mean()
rval["nll"] = -rval["ll"]
return rval
def ll(self, data, n_samples, sigma):
real_data, conditional_data = data
sampled_data = self.sample(conditional_data)
output_space = self.mlp.get_output_space()
if "Conv2D" in str(output_space):
samples = output_space.convert(sampled_data, output_space.axes, ("b", 0, 1, "c"))
samples = samples.flatten(2)
data = output_space.convert(real_data, output_space.axes, ("b", 0, 1, "c"))
data = data.flatten(2)
parzen = theano_parzen(data, samples, sigma)
return parzen