def test_generic_compute_Lx():
## now compare against theano version
vv = T.matrix()
gg = T.matrix()
hh = T.matrix()
aa = T.vector()
bb = T.vector()
cc = T.vector()
xxw_mat = T.matrix()
xxv_mat = T.matrix()
xxw = T.vector()
xxv = T.vector()
xxa = T.vector()
xxb = T.vector()
xxc = T.vector()
# test compute_Lx
LLx = natural.generic_compute_Lx([vv, gg, hh],
[xxw_mat, xxv_mat],
[xxa, xxb, xxc])
f = theano.function([vv, gg, hh, xxw_mat, xxv_mat, xxa, xxb, xxc], LLx)
t1 = time.time()
rvals = f(v, g, h, xw_mat, xv_mat, xa, xb, xc)
print 'Elapsed: ', time.time() - t1
numpy.testing.assert_almost_equal(Lx_w, rvals[0], decimal=3)
numpy.testing.assert_almost_equal(Lx_v, rvals[1], decimal=3)
numpy.testing.assert_almost_equal(Lx_a, rvals[2], decimal=3)
numpy.testing.assert_almost_equal(Lx_b, rvals[3], decimal=3)
numpy.testing.assert_almost_equal(Lx_c, rvals[4], decimal=3)
def test_generic_compute_Lx():
## now compare against theano version
vv = T.matrix()
gg = T.matrix()
hh = T.matrix()
qq = T.matrix()
aa = T.vector()
bb = T.vector()
cc = T.vector()
dd = T.vector()
xxw_mat = T.matrix()
xxv_mat = T.matrix()
xxz_mat = T.matrix()
xxw = T.vector()
xxv = T.vector()
xxa = T.vector()
xxb = T.vector()
xxc = T.vector()
xxd = T.vector()
# test compute_Lx
LLx = natural.generic_compute_Lx([vv, gg, hh, qq],
[xxw_mat, xxv_mat, xxz_mat],
[xxa, xxb, xxc, xxd])
f = theano.function(
[vv, gg, hh, qq, xxw_mat, xxv_mat, xxz_mat, xxa, xxb, xxc, xxd], LLx)
rvals = f(v, g, h, q, xw_mat, xv_mat, xz_mat, xa, xb, xc, xd)
numpy.testing.assert_almost_equal(Lx_w, rvals[0], decimal=3)
numpy.testing.assert_almost_equal(Lx_v, rvals[1], decimal=3)
numpy.testing.assert_almost_equal(Lx_z, rvals[2], decimal=3)
numpy.testing.assert_almost_equal(Lx_a, rvals[3], decimal=3)
numpy.testing.assert_almost_equal(Lx_b, rvals[4], decimal=3)
numpy.testing.assert_almost_equal(Lx_c, rvals[5], decimal=3)
numpy.testing.assert_almost_equal(Lx_d, rvals[6], decimal=3)
def test_generic_compute_Lx():
## now compare against theano version
vv = T.matrix()
gg = T.matrix()
hh = T.matrix()
aa = T.vector()
bb = T.vector()
cc = T.vector()
xxw_mat = T.matrix()
xxv_mat = T.matrix()
xxw = T.vector()
xxv = T.vector()
xxa = T.vector()
xxb = T.vector()
xxc = T.vector()
# test compute_Lx
LLx = natural.generic_compute_Lx([vv, gg, hh], [xxw_mat, xxv_mat],
[xxa, xxb, xxc])
f = theano.function([vv, gg, hh, xxw_mat, xxv_mat, xxa, xxb, xxc], LLx)
t1 = time.time()
rvals = f(v, g, h, xw_mat, xv_mat, xa, xb, xc)
print 'Elapsed: ', time.time() - t1
numpy.testing.assert_almost_equal(Lx_w, rvals[0], decimal=3)
numpy.testing.assert_almost_equal(Lx_v, rvals[1], decimal=3)
numpy.testing.assert_almost_equal(Lx_a, rvals[2], decimal=3)
numpy.testing.assert_almost_equal(Lx_b, rvals[3], decimal=3)
numpy.testing.assert_almost_equal(Lx_c, rvals[4], decimal=3)
def test_generic_compute_Lx():
## now compare against theano version
vv = T.matrix()
gg = T.matrix()
hh = T.matrix()
qq = T.matrix()
aa = T.vector()
bb = T.vector()
cc = T.vector()
dd = T.vector()
xxw_mat = T.matrix()
xxv_mat = T.matrix()
xxz_mat = T.matrix()
xxw = T.vector()
xxv = T.vector()
xxa = T.vector()
xxb = T.vector()
xxc = T.vector()
xxd = T.vector()
# test compute_Lx
LLx = natural.generic_compute_Lx([vv, gg, hh, qq],
[xxw_mat, xxv_mat, xxz_mat],
[xxa, xxb, xxc, xxd])
f = theano.function([vv, gg, hh, qq, xxw_mat, xxv_mat, xxz_mat, xxa, xxb, xxc, xxd], LLx)
rvals = f(v, g, h, q, xw_mat, xv_mat, xz_mat, xa, xb, xc, xd)
numpy.testing.assert_almost_equal(Lx_w, rvals[0], decimal=3)
numpy.testing.assert_almost_equal(Lx_v, rvals[1], decimal=3)
numpy.testing.assert_almost_equal(Lx_z, rvals[2], decimal=3)
numpy.testing.assert_almost_equal(Lx_a, rvals[3], decimal=3)
numpy.testing.assert_almost_equal(Lx_b, rvals[4], decimal=3)
numpy.testing.assert_almost_equal(Lx_c, rvals[5], decimal=3)
numpy.testing.assert_almost_equal(Lx_d, rvals[6], decimal=3)
def test_runtime():
### theano implementation ###
energies = - T.sum(T.dot(symb['v'], symb['W']) * symb['g'], axis=1) \
- T.sum(T.dot(symb['g'], symb['V']) * symb['h'], axis=1) \
- T.dot(symb['v'], symb['a']) \
- T.dot(symb['g'], symb['b']) \
- T.dot(symb['h'], symb['c'])
# Fisher Implementation
symb_params = [symb['W'], symb['V'], symb['a'], symb['b'], symb['c']]
symb_x = [symb['x_W'], symb['x_V'], symb['x_a'], symb['x_b'], symb['x_c']]
f_inputs = [symb['v'], symb['g'], symb['h']] + symb_params + symb_x
fisher_Lx = fisher.compute_Lx(energies, symb_params, symb_x)
fisher_func = theano.function(f_inputs, fisher_Lx)
samples = [symb['v'], symb['g'], symb['h']]
symb_weights = [symb['x_W'], symb['x_V']]
symb_biases = [symb['x_a'], symb['x_b'], symb['x_c']]
f_inputs = [symb['v'], symb['g'], symb['h']] + symb_weights + symb_biases
natural_Lx = natural.generic_compute_Lx(samples, symb_weights, symb_biases)
natural_func = theano.function(f_inputs, natural_Lx)
t1 = time.time()
fisher_rval = fisher_func(vals['v'], vals['g'], vals['h'], vals['W'],
vals['V'], vals['a'], vals['b'], vals['c'],
vals['x_W'], vals['x_V'], vals['x_a'],
vals['x_b'], vals['x_c'])
print 'Fisher runtime (s): ', time.time() - t1
t1 = time.time()
nat_rval = natural_func(vals['v'], vals['g'], vals['h'], vals['x_W'],
vals['x_V'], vals['x_a'], vals['x_b'], vals['x_c'])
print 'Natural runtime (s): ', time.time() - t1
### make sure the two return the same thing ###
for (rval1, rval2) in zip(fisher_rval, nat_rval):
numpy.testing.assert_almost_equal(rval1, rval2, decimal=2)
def test_runtime():
### theano implementation ###
energies = - T.sum(T.dot(symb['v'], symb['W']) * symb['g'], axis=1) \
- T.sum(T.dot(symb['g'], symb['V']) * symb['h'], axis=1) \
- T.dot(symb['v'], symb['a']) \
- T.dot(symb['g'], symb['b']) \
- T.dot(symb['h'], symb['c'])
# Fisher Implementation
symb_params = [symb['W'], symb['V'], symb['a'], symb['b'], symb['c']]
symb_x = [symb['x_W'], symb['x_V'], symb['x_a'], symb['x_b'], symb['x_c']]
f_inputs = [symb['v'], symb['g'], symb['h']] + symb_params + symb_x
fisher_Lx = fisher.compute_Lx(energies, symb_params, symb_x)
fisher_func = theano.function(f_inputs, fisher_Lx)
samples = [symb['v'], symb['g'], symb['h']]
symb_weights = [symb['x_W'], symb['x_V']]
symb_biases = [symb['x_a'], symb['x_b'], symb['x_c']]
f_inputs = [symb['v'], symb['g'], symb['h']] + symb_weights + symb_biases
natural_Lx = natural.generic_compute_Lx(samples, symb_weights, symb_biases)
natural_func = theano.function(f_inputs, natural_Lx)
t1 = time.time()
fisher_rval = fisher_func(vals['v'], vals['g'], vals['h'],
vals['W'], vals['V'], vals['a'], vals['b'], vals['c'],
vals['x_W'], vals['x_V'], vals['x_a'], vals['x_b'], vals['x_c'])
print 'Fisher runtime (s): ', time.time() - t1
t1 = time.time()
nat_rval = natural_func(vals['v'], vals['g'], vals['h'],
vals['x_W'], vals['x_V'], vals['x_a'], vals['x_b'], vals['x_c'])
print 'Natural runtime (s): ', time.time() - t1
### make sure the two return the same thing ###
for (rval1, rval2) in zip(fisher_rval, nat_rval):
numpy.testing.assert_almost_equal(rval1, rval2, decimal=2)
