def __call__(self, x, y, nr_docs):
x, y, nr_docs = as_variable(x), as_variable(y), as_variable(nr_docs)
det_a = self.ranker.max(x)
sto_a = self.ranker.draw(x)
self.report_scores(det_a, y, nr_docs, x.dtype)
self.report_scores(sto_a, y, nr_docs, x.dtype, prefix='stochastic/')
return as_variable(y)
def listpl(x, t, nr_docs, α=10.0):
"""
The ListPL loss, a stochastic variant of ListMLE that in expectation
approximates the true ListNet loss.
:param x: The activation of the previous layer
:type x: chainer.Variable
:param t: The target labels
:type t: chainer.Variable
:param nr_docs: The number of documents per query
:type nr_docs: chainer.Variable
:param α: The temperature parameter of the plackett-luce
:type α: float
:return: The loss
:rtype: chainer.Variable
"""
t, nr_docs = as_variable(t), as_variable(nr_docs)
t = as_variable(t.data.astype(x.dtype))
t = cf.log_softmax(t * α)
indices = sample_without_replacement(t)
x_hat = select_items_per_row(x, indices)
# Compute MLE loss
per_sample_loss = -cf.sum(x_hat - logcumsumexp(x_hat), axis=1)
return cf.mean(per_sample_loss)
def listmle(x, t, nr_docs):
"""
The ListMLE loss as in Xia et al (2008), Listwise Approach to Learning to
Rank - Theory and Algorithm.
:param x: The activation of the previous layer
:type x: chainer.Variable
:param t: The target labels
:type t: chainer.Variable
:param nr_docs: The number of documents per query
:type nr_docs: chainer.Variable
:return: The loss
:rtype: chainer.Variable
"""
t, nr_docs = as_variable(t), as_variable(nr_docs)
# Get the ground truth by sorting activations by the relevance labels
indices = argsort(t, axis=1)
x_hat = select_items_per_row(x, cf.flip(indices, axis=1))
# Compute MLE loss
per_sample_loss = -cf.sum(x_hat - logcumsumexp(x_hat), axis=1)
return cf.mean(per_sample_loss)
def uniform(self, x):
pred = self._predict(x)
xp = cuda.get_array_module(pred)
log_p = F.log_softmax(
as_variable(xp.ones(pred.shape, dtype=pred.dtype)))
action = as_variable(sample_without_replacement(log_p))
return self._cut(action)
def test_bandify_update():
policy = ADFUCBPolicy(3)
mcb = MultiClassBandify(policy)
mcb.update_policy(policy)
# Generate minibatch
np.random.seed(42)
x = as_variable(
np.array([[[2.0, 1.0, 3.0], [-2.0, 4.0, -1.0]],
[[-3.0, 0.0, -1.0], [1.0, 0.0, 1.0]],
[[-1.0, 0.01, -1.0], [0.1, 2.0, 2.0]]]))
y = as_variable(np.array([0, 1, 1], dtype=np.int32))
# Drawing at this point should return the default action
expected = np.array([1, 0, 1])
assert_allclose(policy.draw(x).data, expected)
# Run multi class bandify chain
np.random.seed(42)
for _ in range(100):
_, a, log_p, r = mcb(x, y)
# Drawing at this point should be perfect
expected = np.array([0, 1, 1])
assert_allclose(policy.draw(x).data, expected)
def circular_correlation(self, left_x, right_x):
"""
Computes the circular correlation of two vectors a and b via their fast fourier transforms
In python code, ifft(np.conj(fft(a)) * fft(b)).real
:param left_x: ()
:param right_x:
(a - j * b) * (c + j * d) = (ac + bd) + j * (ad - bc)
:return:
"""
left_x_real = left_x
left_x_imag = chainer.as_variable(
self.xp.zeros_like(left_x_real, dtype=self.xp.float32))
left_x_fft_real, left_x_fft_imag = functions.fft(
(left_x_real, left_x_imag))
right_x_real = right_x
right_x_imag = chainer.as_variable(
self.xp.zeros_like(right_x_real, dtype=self.xp.float32))
right_x_fft_real, right_x_fft_imag = functions.fft(
(right_x_real, right_x_imag))
prod_fft_real = left_x_fft_real * right_x_fft_real + left_x_fft_imag * right_x_fft_imag
prod_fft_imag = left_x_fft_real * right_x_fft_imag - left_x_fft_imag * right_x_fft_real
ifft_real, _ = functions.ifft((prod_fft_real, prod_fft_imag))
return ifft_real
def ranknet(x, t, nr_docs):
"""
The RankNet loss as in Burges et al (2005), Learning to Rank using Gradient
Descent
:param x: The activation of the previous layer
:type x: chainer.Variable
:param t: The target labels
:type t: chainer.Variable
:return: The RankNet loss
:rtype: chainer.Variable
"""
x, t, nr_docs = as_variable(x), as_variable(t), as_variable(nr_docs)
t = as_variable(t.data.astype(x.dtype))
x_ij = _tiled_diff(x)
t_ij = _tiled_diff(t)
p_t_ij = cf.sigmoid(t_ij)
# This loss is a simplified sigmoid cross entropy described in the paper
c_ij = -p_t_ij * x_ij + cf.log(1.0 + cf.exp(x_ij))
loss = cf.mean(c_ij)
return loss
def ndcg(ranking, relevance_scores, nr_docs=None, k=0, exp=True):
"""
Computes the nDCG@k for given list of true relevance labels
(relevance_labels) and given permutation of documents (permutation)
:param ranking: The ranking of the documents
:type ranking: chainer.Variable
:param relevance_scores: The ground truth relevance labels
:type relevance_scores: chainer.Variable
:param nr_docs: A vector of the nr_docs per row
:type nr_docs: chainer.Variable
:param k: The cut-off point (if set to 0, it does not cut-off, if set to
smaller than 0, it computes all possible cut-offs and returns an
array)
:type k: int
:param exp: Set to true to use the exponential variant of nDCG which has a
stronger emphasis on retrieving relevant documents
:type exp: bool
:return: The nDCG@k value
:rtype: chainer.Variable
"""
xp = cuda.get_array_module(relevance_scores)
optimal_ranking = as_variable(xp.fliplr(xp.argsort(relevance_scores.data,
axis=1)))
_dcg = dcg(ranking, relevance_scores, nr_docs, k, exp).data
_idcg = dcg(optimal_ranking, relevance_scores, nr_docs, k, exp).data
_idcg[_idcg == 0.0] = 1.0
return as_variable(_dcg / _idcg)
def multibox_loss(mb_locs, mb_confs, gt_mb_locs, gt_mb_labels, k, comm=None):
mb_locs = chainer.as_variable(mb_locs)
mb_confs = chainer.as_variable(mb_confs)
gt_mb_locs = chainer.as_variable(gt_mb_locs)
gt_mb_labels = chainer.as_variable(gt_mb_labels)
xp = chainer.backends.cuda.get_array_module(gt_mb_labels.array)
with chainer.backends.cuda.get_device_from_array(gt_mb_labels.array):
positive = gt_mb_labels.array > 0
n_positive = positive.sum()
if comm:
n_positive = comm.allreduce_obj(n_positive) / comm.size
if n_positive == 0:
z = chainer.Variable(xp.zeros((), dtype=np.float32))
return z, z
loc_loss = F.huber_loss(mb_locs, gt_mb_locs, 1, reduce='no')
loc_loss = F.sum(loc_loss, axis=-1)
loc_loss *= positive.astype(loc_loss.dtype)
loc_loss = F.sum(loc_loss) / n_positive
conf_loss = _elementwise_softmax_cross_entropy(mb_confs, gt_mb_labels)
hard_negative = _hard_negative(conf_loss.array, positive, k)
conf_loss *= xp.logical_or(positive,
hard_negative).astype(conf_loss.dtype)
conf_loss = F.sum(conf_loss) / n_positive
return loc_loss, conf_loss
def test_update():
policy = ThompsonPolicy(4, 6)
# Generate minibatch
np.random.seed(42)
x = as_variable(
np.array([[1.0, 2.0, 3.0, 3.0, -2.0, -1.0],
[2.0, 3.0, 1.0, -1.0, -3.0, -2.0],
[-1.0, -2.0, -1.0, 1.0, 3.0, 1.0],
[-1.0, -2.0, 1.0, 1.0, 3.0, 1.0]]))
y = as_variable(np.array([2, 1, 0, 3]))
# Drawing at this point should return the default action
expected = np.array([0, 1, 2, 3])
assert_allclose(policy.draw(x).data, expected)
# Perfect update
log_p = as_variable(np.zeros(y.shape))
for _ in range(100):
a = as_variable(np.random.randint(4, size=y.shape))
r = (1.0 * (a.data == y.data))
policy.update(x, a, log_p, as_variable(r))
# Drawing at this point should be perfect
expected = np.array([2, 1, 0, 3])
assert_allclose(policy.draw(x).data, expected)
def forward(self, inputs):
xp = cuda.get_array_module(*inputs)
ranking, relevance_labels = inputs
# Computing nDCG on empty array should just return 0.0
if ranking.shape[1] == 0:
return xp.zeros(ranking.shape[0]),
# Top-k cutoff
last = ranking.shape[1]
if self.k > 0:
last = min(self.k, last)
# For the rankings, compute the relevance labels in order
relevance = select_items_per_row(as_variable(relevance_labels),
as_variable(ranking))
relevance = relevance[:, :last].data.astype(dtype=xp.float32)
# Compute numerator of DCG formula
if self.exp:
numerator = (2.0 ** relevance) - 1.0
else:
numerator = relevance
# Compute denominator of DCG formula
arange = xp.broadcast_to(2.0 + xp.arange(relevance.shape[1]),
relevance.shape)
denominator = xp.log2(arange)
if self.k >= 0:
return xp.asarray(xp.sum(numerator / denominator, axis=1)),
else:
return xp.asarray(xp.cumsum(numerator / denominator, axis=1)),
def add(self, r_hat):
xp = cuda.get_array_module(r_hat)
weights = self.alpha**(xp.arange(r_hat.shape[0], 0, -1.0) - 1.0)
summation = (1 - self.alpha) * F.sum(as_variable(weights) * r_hat,
axis=0)
# Compute exponential moving variance
batch = (1 - self.alpha) * F.cumsum(
as_variable(weights) * r_hat,
axis=0) + (self.alpha**r_hat.shape[0]) * self.biased_mean
batch = batch.data / (1 - self.alpha**(
self.n + 1.0 + xp.arange(0.0, r_hat.shape[0], 1.0)))
batch_s = xp.roll(batch, 1, axis=0)
batch_s[0] = self.mean
diff = r_hat - batch_s
diff = self.alpha * (diff**2)
var = (1 - self.alpha) * F.sum(as_variable(weights) * diff, axis=0)
self.var = (self.alpha**r_hat.shape[0]) * self.var + var.data
# Compute exponential moving average
self.biased_mean = (self.alpha**
r_hat.shape[0]) * self.biased_mean + summation.data
self.n += r_hat.shape[0]
self.bias_correction_factor *= self.alpha**r_hat.shape[0]
self.mean = self.biased_mean / (1 - self.bias_correction_factor)
return self.mean
def get_longlinear(batchsize):
model = LongLinear()
x = np.random.uniform(size=(batchsize, 1000)).astype('f')
x = chainer.as_variable(x)
t = np.random.uniform(size=(batchsize, 1000)).astype('f')
t = chainer.as_variable(t)
return [x, t], model
def test_unpad_basic():
x = as_variable(np.array([[3, 2, 0, 1], [0, 1, 3, 2]]))
print(x)
nr_docs = as_variable(np.array([2, 3]))
y = unpad(x, nr_docs)
expected = as_variable(np.array([[0, 1, 3, 2], [0, 1, 2, 3]], 'f'))
assert_allclose(y.data, expected.data)
def single_backward(fn,
indices,
retained_inputs,
o_grads,
i_grads,
outputs=None):
input_vars = [chainer.as_variable(z) for z in retained_inputs]
fn.inputs = tuple([z.node for z in input_vars])
if outputs is not None:
output_vars = [chainer.as_variable(z) for z in outputs]
fn.outputs = tuple([weakref.ref(z.node) for z in output_vars])
fn._retained_output_data = tuple(outputs)
i_grads = fn.backward_accumulate(indices, o_grads, i_grads)
if len(i_grads) is not len(indices):
i_grads_ = []
for i in indices:
i_grads_.append(i_grads[i])
i_grads = i_grads_
input_vars = None
fn.inputs = None
output_vars = None
fn.outputs = None
fn._retained_output_data = None
def unwrap(v):
if isinstance(v, chainer.Variable):
return v.data
else:
return v
return (unwrap(v) for v in i_grads)
def test_unpad_single_column():
x = as_variable(np.array([[1], [2], [4], [0], [3]]))
print(x)
nr_docs = as_variable(np.array([1, 1, 1, 1, 1]))
y = unpad(x, nr_docs)
expected = as_variable(np.array([[1], [2], [4], [0], [3]], 'f'))
assert_allclose(y.data, expected.data)
def test_unpad_single_row():
x = as_variable(np.array([[1, 2, 4, 0, 3]]))
print(x)
nr_docs = as_variable(np.array([3]))
y = unpad(x, nr_docs)
expected = as_variable(np.array([[1, 2, 0, 4, 3]], 'f'))
assert_allclose(y.data, expected.data)
def __call__(self, x, adj):
# x (batch_size, ): atom id array
h = chainer.as_variable(x)
h = self.embed_model.embedding(h)
# add gaussian noise
if chainer.config.train:
h += (self.xp.random.randn(*h.shape) * self.word_channel_stds *
self.hyperparams.feature_noise_scale)
adj = chainer.as_variable(adj)
sum_log_det_jacobian_x = chainer.as_variable(
self.xp.zeros([h.shape[0]], dtype=self.xp.float32))
sum_log_det_jacobian_adj = chainer.as_variable(
self.xp.zeros([h.shape[0]], dtype=self.xp.float32))
# forward step for channel-coupling layers
for i in range(self.hyperparams.num_coupling["feature"]):
h, log_det_jacobians = self.clinks[i](h, adj)
sum_log_det_jacobian_x += log_det_jacobians
# add uniform noise to adjacency tensors
if chainer.config.train:
adj += self.xp.random.uniform(0, 0.9, adj.shape)
# forward step for adjacency-coupling layers
for i in range(self.hyperparams.num_coupling["feature"],
len(self.clinks)):
adj, log_det_jacobians = self.clinks[i](adj)
sum_log_det_jacobian_adj += log_det_jacobians
adj = F.reshape(adj, (adj.shape[0], -1))
h = F.reshape(h, (h.shape[0], -1))
out = [h, adj]
return out, [sum_log_det_jacobian_x, sum_log_det_jacobian_adj]
def get_diamond(batchsize):
model = DiamondClassifier(784, 10)
x = np.random.uniform(size=(batchsize, 784)).astype('f')
x = chainer.as_variable(x)
t = np.random.randint(size=(batchsize,), low=0, high=10)\
.astype(np.int32)
t = chainer.as_variable(t)
return [x, t], model
def _noise_generate(self, batchsize):
vis_noise = xp.random.uniform(-1, 1,
(batchsize, 128)).astype(xp.float32)
zvis = chainer.as_variable(vis_noise)
ztag = self._get_fake_tag_batch(batchsize, self.dim, threshold=0.75)
ztag = chainer.as_variable(ztag)
return zvis, ztag
def get_convrelu(batchsize):
model = ConvReLUClassifier()
x = np.random.uniform(size=(batchsize, 10000)).astype('f')
x = chainer.as_variable(x)
t = np.random.randint(size=(batchsize, ), low=0, high=10).astype(np.int32)
t = chainer.as_variable(t)
return [x, t], model
def test_inv_select_items_none():
idx = as_variable(np.array([[0, 1, 2, 3], [0, 1, 2, 3]], dtype=np.int32))
val = as_variable(np.array([[0.5, 3.14, 0.0, -9.9], [1.0, -1.0, 1.0,
4.0]]))
out = inverse_select_items_per_row(val, idx)
assert_allclose(out.data, np.array([[], []], dtype=np.int32))
def _tiles(x):
xp = cuda.get_array_module(x)
x = x.data
x_i = xp.reshape(x, (x.shape[0], x.shape[1], 1))
x_j = xp.reshape(x, (x.shape[0], 1, x.shape[1]))
x_i = xp.broadcast_to(x_i, (x.shape[0], x.shape[1], x.shape[1]))
x_j = xp.broadcast_to(x_j, (x.shape[0], x.shape[1], x.shape[1]))
return as_variable(x_i), as_variable(x_j)
def test_select_items_identity():
idx = as_variable(np.array([[0, 1, 2, 3], [0, 1, 2, 3]]))
val = as_variable(np.array([[0.5, 3.14, 0.0, -9.9], [1.0, -1.0, 1.0,
4.0]]))
out = select_items_per_row(val, idx)
assert_allclose(out.data, val.data)
def get_resnet152(batchsize):
model = L.ResNet152Layers(pretrained_model=None)
model = Wrapper(model, 'fc6')
x = np.random.uniform(size=(batchsize, 3, 224, 224)).astype('f')
x = chainer.as_variable(x)
t = np.random.randint(size=(batchsize, ), low=0,
high=1000).astype(np.int32)
t = chainer.as_variable(t)
return [x, t], model
def get_pspnet(batchsize):
model = pspnet()
x = np.random.uniform(size=(batchsize, 3, 713, 713)).astype('f')
x = chainer.as_variable(x)
t = np.random.randint(size=(batchsize, 713, 713,), low=0, high=10)\
.astype(np.int32)
t = chainer.as_variable(t)
return [x, t], model
def __call__(self, xs, ys, nr_docs):
prediction = self.ranker(xs)
loss = self.loss_fn(prediction, ys, nr_docs)
report({"loss": loss})
xp = cuda.get_array_module(prediction)
ranking = as_variable(xp.fliplr(xp.argsort(prediction.data, axis=1)))
ndcg_score = ndcg(ranking, as_variable(ys), as_variable(nr_docs))
report({"ndcg": F.mean(ndcg_score)})
return loss
def get_unet(batchsize):
model = UNET()
x = np.random.uniform(size=(batchsize, 1, 572, 572)).astype('f')
x = chainer.as_variable(x)
t = np.random.randint(size=(batchsize, 388, 388), low=0,
high=2).astype(np.int32)
t = chainer.as_variable(t)
return [x, t], model
def get_segnet(batchsize):
model = SegNetBasic(n_class=17)
model = PixelwiseSoftmaxClassifier(model, class_weight=np.ones(17))
x = np.random.uniform(size=(batchsize, 3, 1024, 1024)).astype('f')
x = chainer.as_variable(x)
t = np.random.randint(size=(batchsize, 1024, 1024), low=0, high=10)\
.astype(np.int32)
t = chainer.as_variable(t)
return [x, t], model
def forward(self, inputs):
loc, scale = inputs
xp = backend.get_array_module(loc)
eps = xp.random.randn(*loc.shape)
z = as_variable(loc).array + eps * as_variable(scale).array
self.retain_inputs((0, 1))
self.retain_outputs((0, ))
return xp.array(z, dtype='f'),
def __init__(self, loc, **kwargs):
scale_tril = None
if kwargs:
scale_tril, = argument.parse_kwargs(
kwargs, ('scale_tril', scale_tril))
if scale_tril is None:
raise ValueError("`scale_tril` must have a value.")
self.loc = chainer.as_variable(loc)
self.scale_tril = chainer.as_variable(scale_tril)
self.d = self.scale_tril.shape[-1]
def __init__(self, p=None, logit=None):
super(Bernoulli, self).__init__()
if not (p is None) ^ (logit is None):
raise ValueError(
"Either `p` or `logit` (not both) must have a value.")
with chainer.using_config('enable_backprop', True):
if p is None:
self.logit = chainer.as_variable(logit)
self.p = sigmoid.sigmoid(self.logit)
else:
self.p = chainer.as_variable(p)
self.logit = exponential.log(self.p) \
- logarithm_1p.log1p(-self.p)
def broadcast(*args):
"""Broadcast given variables.
Args:
args (:class:`~chainer.Variable` or :class:`numpy.ndarray` \
or :class:`cupy.ndarray`):
Input variables to be broadcasted. Each dimension of the shapes \
of the input variables must have the same size.
Returns:
~chainer.Variable: :class:`~chainer.Variable` or tuple of \
:class:`~chainer.Variable` objects which are broadcasted \
from given arguments.
.. admonition:: Example
>>> x = np.random.uniform(0, 1, (3, 2)).astype(np.float32)
>>> y = F.broadcast(x)
>>> np.all(x == y.data)
True
>>> z = np.random.uniform(0, 1, (3, 2)).astype(np.float32)
>>> y, w = F.broadcast(x, z)
>>> np.all(x == y.data) & np.all(z == w.data)
True
"""
if len(args) == 1:
return chainer.as_variable(args[0])
return Broadcast().apply(args)
def cast(x, typ):
"""Cast an input variable to a given type.
Args:
x (:class:`~chainer.Variable` or :ref:`ndarray`):
Input variable to be casted. A \
:math:`(s_1, s_2, ..., s_N)`-shaped array.
typ (:class:`str` of dtype or :class:`numpy.dtype`):
Typecode or data type to cast.
Returns:
~chainer.Variable: Variable holding a casted array.
.. admonition:: Example
>>> x = np.arange(0, 3, dtype=np.float64)
>>> x.dtype
dtype('float64')
>>> y = F.cast(x, np.float32)
>>> y.dtype
dtype('float32')
>>> y = F.cast(x, 'float16')
>>> y.dtype
dtype('float16')
"""
if x.dtype == typ:
if not chainer.config.enable_backprop:
return chainer.as_variable(x)
return Cast(typ).apply((x,))[0]
def sign(x):
"""Elementwise sign function.
For a given input :math:`x`, this function returns :math:`sgn(x)`
defined as
.. math::
sgn(x) = \\left \\{ \\begin{array}{cc}
-1 & {\\rm if~x < 0} \\\\
0 & {\\rm if~x = 0} \\\\
1 & {\\rm if~x > 0} \\\\
\\end{array} \\right.
.. note::
The gradient of this function is ``None`` everywhere and therefore
unchains the computational graph.
Args:
x (~chainer.Variable): Input variable for which the sign is computed.
Returns:
~chainer.Variable: Output variable.
"""
if isinstance(x, chainer.variable.Variable):
x = x.array
xp = backend.get_array_module(x)
return chainer.as_variable(utils.force_array(xp.sign(x)))
def broadcast_to(x, shape):
"""Broadcast a given variable to a given shape.
Args:
x (:class:`~chainer.Variable` or :class:`numpy.ndarray` or \
:class:`cupy.ndarray`):
Input variable be broadcasted. A \
:math:`(s_1, s_2, ..., s_N)`-shaped float array.
shape (tuple): Tuple of :class:`int` of the shape of the \
output variable.
Returns:
~chainer.Variable: Output variable broadcasted to the given shape.
.. admonition:: Example
>>> x = np.arange(0, 3)
>>> x
array([0, 1, 2])
>>> y = F.broadcast_to(x, (3, 3))
>>> y.data
array([[0, 1, 2],
[0, 1, 2],
[0, 1, 2]])
"""
if x.shape == shape:
return chainer.as_variable(x)
y, = BroadcastTo(shape).apply((x,))
return y
def sum_to(x, shape):
"""Sum elements along axes to output an array of a given shape.
Args:
x (:class:`~chainer.Variable` or :ref:`ndarray`): Input variable.
shape (tuple of int): The target shape.
Returns:
~chainer.Variable: Output variable of shape ``shape``.
.. admonition:: Example
>>> x = np.array([[1., 2., 3.], [4., 5., 6.]])
>>> x
array([[1., 2., 3.],
[4., 5., 6.]])
>>> y = F.sum_to(x, (1, 3))
>>> y
variable([[5., 7., 9.]])
>>> z = F.sum_to(x, (2, 1))
>>> z
variable([[ 6.],
[15.]])
"""
if x.shape == shape:
return chainer.as_variable(x)
y, = SumTo(shape).apply((x,))
return y
def prob(self, x):
x = chainer.as_variable(x)
if self._is_gpu:
valid = cuda.cupy.bitwise_or(x.array == 0, x.array == 1)
else:
valid = numpy.bitwise_or(x.array == 0, x.array == 1)
ret = x * self.p + (1 - x) * (1 - self.p)
return ret * valid
def __init__(self, p=None, **kwargs):
logit = None
if kwargs:
logit, = argument.parse_kwargs(
kwargs, ('logit', logit))
if not (p is None) ^ (logit is None):
raise ValueError(
"Either `p` or `logit` (not both) must have a value.")
with chainer.using_config('enable_backprop', True):
if p is None:
logit = chainer.as_variable(logit)
self.__log_p = log_softmax.log_softmax(logit, axis=-1)
self.__p = exponential.exp(self.__log_p)
else:
self.__p = chainer.as_variable(p)
self.__log_p = exponential.log(self.__p)
def forward(self, input, target, mask):
input = chainer.as_variable(input)
target = chainer.as_variable(target)
mask = chainer.as_variable(mask)
output = self.predictor(input)
output = output * mask
target = target * mask
d_fake = self.discriminator(input, output)
d_real = self.discriminator(input, target)
loss = {
'predictor': self._loss_predictor(self.predictor, output, target, d_fake),
'discriminator': self._loss_discriminator(self.discriminator, d_real, d_fake),
}
return loss
def log_prob(self, x):
x = chainer.as_variable(x)
logp = self._log_alpha \
+ self.alpha * self._log_scale \
- (self.alpha + 1) * exponential.log(x)
xp = logp.xp
return where.where(
utils.force_array(x.data >= self.scale.data),
logp, xp.array(-xp.inf, logp.dtype))
def __init__(self, loc, scale=None, **kwargs):
super(Normal, self).__init__()
log_scale = None
if kwargs:
log_scale, = argument.parse_kwargs(
kwargs, ('log_scale', log_scale))
if not (scale is None) ^ (log_scale is None):
raise ValueError(
"Either `scale` or `log_scale` (not both) must have a value.")
self.loc = chainer.as_variable(loc)
with chainer.using_config('enable_backprop', True):
if scale is None:
self.__log_scale = chainer.as_variable(log_scale)
self.__scale = exponential.exp(self.log_scale)
else:
self.__scale = chainer.as_variable(scale)
self.__log_scale = exponential.log(self.scale)
def log_prob(self, x):
x = chainer.as_variable(x)
logp = (self.a - 1) * exponential.log(x) \
+ (self.b - 1) * exponential.log(1 - x) \
- _lbeta(self.a, self.b)
xp = logp.xp
return where.where(
utils.force_array((x.array >= 0) & (x.array <= 1)),
logp, xp.array(-xp.inf, logp.dtype))
def dropout(x, ratio=.5, **kwargs):
"""dropout(x, ratio=.5)
Drops elements of input variable randomly.
This function drops input elements randomly with probability ``ratio`` and
scales the remaining elements by factor ``1 / (1 - ratio)``. In testing
mode, it does nothing and just returns ``x``.
.. warning::
``train`` argument is not supported anymore since v2.
Instead, use ``chainer.using_config('train', boolean)``.
See :func:`chainer.using_config`.
Args:
x (:class:`~chainer.Variable` or :class:`numpy.ndarray` or \
:class:`cupy.ndarray`):
Input variable. A :math:`(s_1, s_2, ..., s_N)` -shaped float array.
ratio (float):
Dropout ratio. The ``ratio`` must be ``0.0 <= ratio < 1.0``.
Returns:
~chainer.Variable: Output variable.
See the paper by G. Hinton: `Improving neural networks by preventing \
co-adaptation of feature detectors <https://arxiv.org/abs/1207.0580>`_.
.. admonition:: Example
>>> x = np.array([[-1, 0], [2, -3], [-2, 1]], np.float32)
>>> with chainer.using_config('train', True):
... y = F.dropout(x)
>>> y.data
array([[-2., 0.],
[ 4., -6.],
[-0., 2.]], dtype=float32)
>>> with chainer.using_config('train', True):
... y = F.dropout(x, ratio=0.0) \
# dropout returns original input if ratio=0.0
>>> (x == y.data).all()
True
>>> with chainer.using_config('train', False):
... y = F.dropout(x) \
# dropout in test mode returns original input
>>> (x == y.data).all()
True
"""
argument.check_unexpected_kwargs(
kwargs, train='train argument is not supported anymore. '
'Use chainer.using_config')
argument.assert_kwargs_empty(kwargs)
if configuration.config.train:
return Dropout(ratio).apply((x,))[0]
return chainer.as_variable(x)
def prob(self, x):
x = chainer.as_variable(x)
prob = x * self.p + (1 - x) * (1 - self.p)
if self.binary_check:
if self._is_gpu:
valid = cuda.cupy.bitwise_or(x.array == 0, x.array == 1)
else:
valid = numpy.bitwise_or(x.array == 0, x.array == 1)
prob *= valid
return prob
def __init__(self, **kwargs):
low, high, loc, scale = None, None, None, None
if kwargs:
low, high, loc, scale = argument.parse_kwargs(
kwargs, ('low', low), ('high', high), ('loc', loc),
('scale', scale))
if not (low is None or high is None) ^ (loc is None or scale is None):
raise ValueError(
"Either `low, high` or `loc, scale` (not both) must have a "
"value.")
with chainer.using_config('enable_backprop', True):
if low is None:
self.__loc = chainer.as_variable(loc)
self.__scale = chainer.as_variable(scale)
self.__low = self.__loc
self.__high = self.__loc + self.__scale
else:
self.__low = chainer.as_variable(low)
self.__high = chainer.as_variable(high)
self.__loc = self.__low
self.__scale = self.__high - self.__low
def reshape(x, shape):
"""Reshapes an input variable without copy.
Args:
x (:class:`~chainer.Variable` or :class:`numpy.ndarray` or \
:class:`cupy.ndarray`): Input variable.
shape (:class:`tuple` of :class:`int` s):
Expected shape of the output array. The number of elements which
the array of ``shape`` contains must be equal to that of input
array. One shape dimension can be -1. In this case, the value is
inferred from the length of the array and remaining dimensions.
Returns:
~chainer.Variable:
Variable that holds a reshaped version of the input variable.
.. seealso:: :func:`numpy.reshape`, :func:`cupy.reshape`
.. admonition:: Example
>>> x = np.array([[1, 2, 3, 4], [5, 6, 7, 8]])
>>> y = F.reshape(x, (8,))
>>> y.shape
(8,)
>>> y.data
array([1, 2, 3, 4, 5, 6, 7, 8])
>>> y = F.reshape(x, (4, -1)) # the shape of output is inferred
>>> y.shape
(4, 2)
>>> y.data
array([[1, 2],
[3, 4],
[5, 6],
[7, 8]])
>>> y = F.reshape(x, (4, 3)) \
# the shape of input and output are not consistent
Traceback (most recent call last):
...
chainer.utils.type_check.InvalidType:
Invalid operation is performed in: Reshape (Forward)
Expect: prod(in_types[0].shape) == prod((4, 3))
Actual: 8 != 12
"""
if x.shape == shape:
return chainer.as_variable(x)
y, = Reshape(shape).apply((x,))
return y
def ceil(x):
"""Elementwise ceil function.
.. math::
y_i = \\lceil x_i \\rceil
Args:
x (~chainer.Variable): Input variable.
Returns:
~chainer.Variable: Output variable.
"""
if isinstance(x, chainer.variable.Variable):
x = x.data
xp = backend.get_array_module(x)
return chainer.as_variable(utils.force_array(xp.ceil(x), x.dtype))
def floor(x):
"""Elementwise floor function.
.. math::
y_i = \\lfloor x_i \\rfloor
Args:
x (:class:`~chainer.Variable` or :ref:`ndarray`): Input variable.
Returns:
~chainer.Variable: Output variable.
"""
if isinstance(x, chainer.variable.Variable):
x = x.array
xp = backend.get_array_module(x)
return chainer.as_variable(utils.force_array(xp.floor(x), x.dtype))
def floor(x):
"""Elementwise floor function.
.. math::
y_i = \\lfloor x_i \\rfloor
Args:
x (~chainer.Variable): Input variable.
Returns:
~chainer.Variable: Output variable.
"""
if isinstance(x, chainer.variable.Variable):
x = x.data
xp = cuda.get_array_module(x)
return chainer.as_variable(utils.force_array(xp.floor(x), x.dtype))
def fix(x):
"""Elementwise fix function.
.. math::
y_i = \\lfix x_i \\rfix
Args:
x (~chainer.Variable): Input variable.
Returns:
~chainer.Variable: Output variable.
"""
if isinstance(x, chainer.variable.Variable):
x = x.array
xp = backend.get_array_module(x)
return chainer.as_variable(utils.force_array(xp.fix(x), x.dtype))
def __init__(self, a, b):
super(Beta, self).__init__()
self.__a = chainer.as_variable(a)
self.__b = chainer.as_variable(b)
def p(self):
if self.__p is not None:
return chainer.as_variable(self.__p)
else:
return sigmoid.sigmoid(self.logit)
def b(self):
return chainer.as_variable(self.__b)
def logit(self):
if self.__logit is not None:
return chainer.as_variable(self.__logit)
else:
return exponential.log(self.p) - logarithm_1p.log1p(-self.p)
def a(self):
return chainer.as_variable(self.__a)
def __init__(self, mu, sigma):
self.__mu = chainer.as_variable(mu)
self.__sigma = chainer.as_variable(sigma)
def loc(self):
return chainer.as_variable(self.__loc)
def scale(self):
return chainer.as_variable(self.__scale)
def __init__(self, alpha):
self.__alpha = chainer.as_variable(alpha)
