def __call__(self, X):
XY = X.dot(X.T)
x2 = at.sum(X**2, axis=1).dimshuffle(0, "x")
X2e = at.repeat(x2, X.shape[0], axis=1)
H = X2e + X2e.T - 2.0 * XY
V = at.sort(H.flatten())
length = V.shape[0]
# median distance
m = at.switch(
at.eq((length % 2), 0),
# if even vector
at.mean(V[((length // 2) - 1):((length // 2) + 1)]),
# if odd vector
V[length // 2],
)
h = 0.5 * m / at.log(floatX(H.shape[0]) + floatX(1))
# RBF
Kxy = at.exp(-H / h / 2.0)
# Derivative
dxkxy = -at.dot(Kxy, X)
sumkxy = at.sum(Kxy, axis=-1, keepdims=True)
dxkxy = at.add(dxkxy, at.mul(X, sumkxy)) / h
return Kxy, dxkxy
def test_GpuCrossentropySoftmaxArgmax1HotWithBias():
# This is basic test for GpuCrossentropySoftmaxArgmax1HotWithBias
# We check that we loop when their is too much threads
n_in = 1000
batch_size = 4097
n_out = 1250
if not isinstance(mode_with_gpu, aesara.compile.DebugMode):
n_in = 4098
n_out = 4099
y = tt.lvector("y")
b = tt.fvector("b")
# we precompute the dot with big shape before to allow the test of
# GpuCrossentropySoftmax1HotWithBiasDx to don't fail with the error
# (the launch timed out and was terminated) on GPU card not
# powerful enough. We need the big shape to check for corner
# case.
dot_result = tt.fmatrix("dot_result")
# Seed numpy.random with config.unittests.rseed
utt.seed_rng()
xx = np.asarray(np.random.rand(batch_size, n_in), dtype=np.float32)
yy = np.ones((batch_size, ), dtype="int32")
b_values = np.zeros((n_out, ), dtype="float32")
W_values = np.asarray(np.random.rand(n_in, n_out), dtype="float32")
dot_value = np.asarray(np.dot(xx, W_values), dtype="float32")
del W_values
p_y_given_x = tt.nnet.softmax(dot_result + b)
y_pred = tt.argmax(p_y_given_x, axis=-1)
loss = -tt.mean(tt.log(p_y_given_x)[tt.arange(y.shape[0]), y])
dW = tt.grad(loss, dot_result)
classify = aesara.function(inputs=[y, b, dot_result],
outputs=[loss, y_pred, dW],
mode=mode_without_gpu)
classify_gpu = aesara.function(inputs=[y, b, dot_result],
outputs=[loss, y_pred, dW],
mode=mode_with_gpu)
assert any([
isinstance(node.op, tt.nnet.CrossentropySoftmaxArgmax1HotWithBias)
for node in classify.maker.fgraph.toposort()
])
assert any([
isinstance(node.op, GpuCrossentropySoftmaxArgmax1HotWithBias)
for node in classify_gpu.maker.fgraph.toposort()
])
out = classify(yy, b_values, dot_value)
gout = classify_gpu(yy, b_values, dot_value)
assert len(out) == len(gout) == 3
utt.assert_allclose(out[0], gout[0])
utt.assert_allclose(out[2], gout[2], atol=3e-6)
utt.assert_allclose(out[1], gout[1])
def setup_gpu_op(self,
activations,
labels,
input_length,
compute_grad=True):
gpu_ctc_cost = gpu_ctc(activations, labels, input_length)
outputs = [gpu_ctc_cost]
if compute_grad:
# Symbolic gradient of CTC cost
gpu_ctc_grad = tt.grad(tt.mean(gpu_ctc_cost), activations)
outputs += [gpu_ctc_grad]
return aesara.function([], outputs, mode=mode_with_gpu)
def run_ctc(self, activations, labels, input_length, expected_costs,
expected_grads):
# Create symbolic variables
t_activations = aesara.shared(activations, name="activations")
t_activation_times = aesara.shared(input_length,
name="activation_times")
t_labels = aesara.shared(labels, name="labels")
t_cost = ctc(t_activations, t_labels, t_activation_times)
# Symbolic gradient of CTC cost
t_grad = tt.grad(tt.mean(t_cost), t_activations)
# Compile symbolic functions
train = aesara.function([], [t_cost, t_grad])
cost, grad = train()
utt.assert_allclose(expected_grads / cost.shape[0], grad)
utt.assert_allclose(expected_costs, cost)
self.check_grads_disabled(t_activations, t_labels, t_activation_times)
def get_moment(cls, rv, *sim_inputs):
# Take the mean of 10 draws
multiple_sim = rv.owner.op(*sim_inputs,
size=at.concatenate([[10], rv.shape]))
return at.mean(multiple_sim, axis=0)
import time
import numpy as np
import aesara
from aesara import tensor as tt
from aesara.ifelse import ifelse
a, b = tt.scalars("a", "b")
x, y = tt.matrices("x", "y")
z_switch = tt.switch(tt.lt(a, b), tt.mean(x), tt.mean(y))
z_lazy = ifelse(tt.lt(a, b), tt.mean(x), tt.mean(y))
f_switch = aesara.function([a, b, x, y], z_switch)
f_lazyifelse = aesara.function([a, b, x, y], z_lazy)
val1 = 0.0
val2 = 1.0
big_mat1 = np.ones((10000, 1000))
big_mat2 = np.ones((10000, 1000))
n_times = 10
tic = time.clock()
for i in range(n_times):
f_switch(val1, val2, big_mat1, big_mat2)
print("time spent evaluating both values %f sec" % (time.clock() - tic))
tic = time.clock()
