def test_take_along_axis_grad(self, shape, axis, samples):
if axis < 0:
_axis = len(shape) + axis
else:
_axis = axis
# Setup the theano function
t_arr, t_indices = self.get_input_tensors(shape)
t_out2 = theano.grad(
tt.sum(self._output_tensor(t_arr**2, t_indices, axis)),
t_arr,
)
func = theano.function([t_arr, t_indices], [t_out2])
# Test that the gradient gives the same output as what is expected
arr, indices = self.get_input_values(shape, axis, samples)
expected_grad = np.zeros_like(arr)
slicer = [slice(None)] * len(shape)
for i in range(indices.shape[axis]):
slicer[axis] = i
inds = indices[slicer].reshape(shape[:_axis] + (1, ) +
shape[_axis + 1:])
inds = _make_along_axis_idx(shape, inds, _axis)
expected_grad[inds] += 1
expected_grad *= 2 * arr
out = func(arr, indices)[0]
assert np.allclose(out, expected_grad)
def apply(self, f):
# f: kernel function for KSD f(histogram) -> (k(x,.), \nabla_x k(x,.))
X = self.approx.histogram
t = self.approx.normalizing_constant
dlogpdx = theano.scan(
fn=lambda zg: theano.grad(self.logp_norm(zg), zg),
sequences=[X])[0] # bottleneck
Kxy, dxkxy = f(X)
# scaling factor
# not needed for Kxy as we already scaled dlogpdx
dxkxy /= t
n = X.shape[0].astype('float32') / t
svgd_grad = (tt.dot(Kxy, dlogpdx) + dxkxy) / n
return -1 * svgd_grad # gradient
def apply(self, f):
# f: kernel function for KSD f(histogram) -> (k(x,.), \nabla_x k(x,.))
X = self.approx.histogram
t = self.approx.normalizing_constant
dlogpdx = theano.scan(
fn=lambda zg: theano.grad(self.logp_norm(zg), zg),
sequences=[X]
)[0] # bottleneck
Kxy, dxkxy = f(X)
# scaling factor
# not needed for Kxy as we already scaled dlogpdx
dxkxy /= t
n = X.shape[0].astype('float32') / t
svgd_grad = (tt.dot(Kxy, dlogpdx) + dxkxy) / n
return -1 * svgd_grad # gradient
def dlogp(self):
return theano.scan(
fn=lambda zg: theano.grad(self.approx.logp_norm(zg), zg),
sequences=[self.input_matrix])[0]
def train_model(learning_rate=0.1,
n_epochs=1000,
batch_size=20,
n_hidden=500):
# BUILD MODEL #
print("...building model")
rng = numpy.random.RandomState(1234)
classifier = DeepVS(
rng=rng,
input=x,
n_in=42,
n_hidden=n_hidden,
n_out=3
)
cost = (
classifier.negative_log_likelihood(y)
+ L1_reg * classifier.L1
+ L2_reg * classifier.L2_sqr
)
test_model = theano.function(
inputs=[index],
outputs=classifier.errors(y),
givens={
x: test_set_x[index * batch_size:(index + 1) * batch_size],
y: test_set_y[index * batch_size:(index + 1) * batch_size]
}
)
validate_model = theano.function(
inputs=[index],
outputs=classifier.errors(y),
givens={
x: valid_set_x[index * batch_size:(index + 1) * batch_size],
y: valid_set_y[index * batch_size:(index + 1) * batch_size]
}
)
# compute the gradient of cost with respect to theta (sorted in params)
# the resulting gradients will be stored in a list gparams
gparams = [T.grad(cost, param) for param in classifier.params]
updates = [
(param, param - learning_rate * gparam)
for param, gparam in zip(classifier.params, gparams)
]
# compiling a Theano function `train_model` that returns the cost, but
# in the same time updates the parameter of the model based on the rules
# defined in `updates`
train_model = theano.function(
inputs=[index],
outputs=cost,
updates=updates,
givens={
x: train_set_x[index * batch_size: (index + 1) * batch_size],
y: train_set_y[index * batch_size: (index + 1) * batch_size]
}
)