def local_lrnGrad_mkl(node):
if not mkl_available():
return
if not isinstance(node.op, mkl_lrn.AbstractLRNGrad):
return
if node.inputs[0].type.ndim != 4:
return
try:
x, gz, = node.inputs
x_u2i = U2ILRN(alpha=node.op.alpha,
beta=node.op.beta,
k=node.op.k,
n=node.op.n)(x)
lrnOut = mkl_lrn.LRN(alpha=node.op.alpha,
beta=node.op.beta,
k=node.op.k,
n=node.op.n)(x_u2i)
gz_u2i = I2UGrad()(lrnOut, gz)
lrnGradOut = mkl_lrn.LRNGrad(alpha=node.op.alpha,
beta=node.op.beta,
k=node.op.k,
n=node.op.n)(x_u2i, gz_u2i)
gx_i2u = U2IGrad()(x, lrnGradOut)
rval = gx_i2u
return [rval]
except Exception as e:
msg = ('Failed to apply local opt to Op %s. '
'Exception message: %s\n') % (node.op, str(e))
_logger.warning(msg)
return
def test_lrn_float64(self):
old_floatX = theano.config.floatX
theano.config.floatX = 'float64'
x = tensor.dtensor4('x')
x_internal = U2ILRN()(x)
z_internal = mkl_lrn.LRN()(x_internal)
z = I2U()(z_internal)
f = theano.function([x], z, mode=mode_with_mkl)
imval = numpy.random.rand(4, 2, 4, 4).astype(theano.config.floatX)
f(imval)
assert f(imval).dtype == 'float64'
theano.config.floatX = old_floatX
def test_lrn_grad_float32(self):
old_floatX = theano.config.floatX
theano.config.floatX = 'float32'
x = tensor.ftensor4('x')
x_internal = U2ILRN()(x)
z_internal = mkl_lrn.LRN()(x_internal)
z = I2U()(z_internal)
z_sum = tensor.sum(z)
g = tensor.grad(z_sum, [x])
f = theano.function([x], g, mode=mode_with_mkl)
imval = numpy.random.rand(4, 2, 4, 4).astype(theano.config.floatX)
f(imval)
assert f(imval)[0].dtype == 'float32'
theano.config.floatX = old_floatX
def test_mkl_lrn_value(self):
shape = [(2, 15, 3, 4), (256, 256, 27, 27)] # NCHW
n = 5
k = 2
alpha = 0.0001
beta = 0.75
x = tensor.dtensor4('x')
x_internal = U2ILRN()(x)
z_internal = mkl_lrn.LRN(alpha, beta, k, n)(x_internal)
z = I2U()(z_internal)
fz = theano.function([x], z, mode=mode_with_mkl)
# for shape[0]
input_data = numpy.random.rand(*shape[0]).astype(theano.config.floatX)
t = self.ground_truth_normalizer(input_data,
k=k,
n=n,
alpha=alpha,
beta=beta)
assert (fz(input_data)).shape == t.shape
assert numpy.allclose(fz(input_data), t)