def test_grad():
"""
Test Op's gradient w.r.t top_down against theano graph implementation
"""
rng = np.random.RandomState([2012,7,19])
batch_size_list = [1]
channels = 16
rows_list = [2, 24]
pool_rows_list = [2, 3]
# TODO theano graph version fails with pool shape 1,1,
# try it with python version
for batch_size in batch_size_list:
for rows, pool_rows in zip(rows_list, pool_rows_list):
cols = rows
pool_cols = pool_rows
zv = rng.randn(channels, rows, cols,
batch_size).astype(config.floatX)
tv = rng.randn(channels, rows / pool_rows, cols / \
pool_cols, batch_size).astype(config.floatX)
z = T.tensor4()
t = T.tensor4()
# gpu op
p, h = prob_max_pool_c01b(z, (pool_rows, pool_cols), top_down = t)
gh_t = T.grad(h.sum(), t)
gp_t = T.grad(p.sum(), t)
gh_z = T.grad(h.sum(), z)
gp_z = T.grad(p.sum(), z)
gph_z = T.grad(p.sum() + h.sum(), z)
gph_t = T.grad(p.sum() + h.sum(), t)
func = function([z, t], [gh_t, gp_t, gh_z, gp_z, gph_z, gph_t],
mode = mode_with_gpu)
op_rval = func(zv, tv)
# theano graph
p, h = max_pool_c01b(z, (pool_rows, pool_cols) , top_down = t)
gh_t = T.grad(h.sum(), t)
gp_t = T.grad(p.sum(), t)
gh_z = T.grad(h.sum(), z)
gp_z = T.grad(p.sum(), z)
gph_z = T.grad(p.sum() + h.sum(), z)
gph_t = T.grad(p.sum() + h.sum(), t)
func = function([z, t], [gh_t, gp_t, gh_z, gp_z, gph_z, gph_t],
mode = mode_without_gpu)
th_rval = func(zv, tv)
for op, th in zip (op_rval, th_rval):
assert np.allclose(op, th, rtol=1e-04, atol=1e-06)
def test_top_down_grad_correctness():
"""
Test Op's gradient w.r.t top_down against theano graph implementation
"""
rng = np.random.RandomState([2012, 7, 19])
batch_size_list = [128]
channels = 16
rows_list = [2, 8, 20]
pool_rows_list = [2, 4, 5]
# TODO theano graph version fails with pool shape 1,1,
# try it with python version
# TODO the results doesn't match for (30,30), (3, 3)
# check verify grad
for batch_size in batch_size_list:
for rows, pool_rows in zip(rows_list, pool_rows_list):
cols = rows
pool_cols = pool_rows
zv = rng.randn(channels, rows, cols,
batch_size).astype(config.floatX)
tv = rng.randn(channels, rows / pool_rows, cols / pool_cols,
batch_size).astype(config.floatX)
z = T.tensor4()
t = T.tensor4()
# gpu op
p, h = prob_max_pool_c01b(z, (pool_rows, pool_cols), top_down=t)
gt = T.grad(h.sum() + p.sum(), t)
gt = T.grad(h.sum() + p.sum(), t)
func = function([z, t], gt, mode=mode_with_gpu)
op_gt = func(zv, tv)
# theano graph
p, h = max_pool_c01b(z, (pool_rows, pool_cols), top_down=t)
gt = T.grad(h.sum() + p.sum(), t)
func = function([z, t], gt, mode=mode_without_gpu)
th_gt = func(zv, tv)
print batch_size, rows, pool_rows
assert np.allclose(op_gt, th_gt, rtol=1e-04, atol=1e-06)
def test_top_down_grad_correctness():
"""
Test Op's gradient w.r.t top_down against theano graph implementation
"""
rng = np.random.RandomState([2012,7,19])
batch_size_list = [128]
channels = 16
rows_list = [2, 8, 20]
pool_rows_list = [2, 4, 5]
# TODO theano graph version fails with pool shape 1,1,
# try it with python version
# TODO the results doesn't match for (30,30), (3, 3)
# check verify grad
for batch_size in batch_size_list:
for rows, pool_rows in zip(rows_list, pool_rows_list):
cols = rows
pool_cols = pool_rows
zv = rng.randn(channels, rows, cols, batch_size).astype(config.floatX)
tv = rng.randn(channels, rows / pool_rows, cols / pool_cols, batch_size).astype(config.floatX)
z = T.tensor4()
t = T.tensor4()
# gpu op
p, h = prob_max_pool_c01b(z, (pool_rows, pool_cols), top_down = t)
gt = T.grad(h.sum() + p.sum(), t)
gt = T.grad(h.sum() + p.sum(), t)
func = function([z, t], gt, mode = mode_with_gpu)
op_gt = func(zv, tv)
# theano graph
p, h = max_pool_c01b(z, (pool_rows, pool_cols) , top_down = t)
gt = T.grad(h.sum() + p.sum(), t)
func = function([z, t], gt, mode = mode_without_gpu)
th_gt = func(zv, tv)
print batch_size, rows, pool_rows
assert np.allclose(op_gt, th_gt, rtol=1e-04, atol=1e-06)
def test_top_donw_correctness():
"""
Test the forward pass Op against theano graph implementation
"""
rng = np.random.RandomState([2012,7,19])
batch_size_list = [1]
channels = 16
rows_list = [2, 24]
pool_rows_list = [2, 3]
# TODO theano graph version fails with pool shape 1,1,
# try it with python version
for batch_size in batch_size_list:
for rows, pool_rows in zip(rows_list, pool_rows_list):
cols = rows
pool_cols = pool_rows
zv = rng.randn(channels, rows, cols, batch_size).astype(config.floatX)
tv = rng.randn(channels, rows / pool_rows, cols / pool_cols, batch_size).astype(config.floatX)
z = T.tensor4()
t = T.tensor4()
# gpu op
p, h = prob_max_pool_c01b(z, (pool_rows, pool_cols), top_down = t)
func = function([z, t], [p, h], mode = mode_with_gpu)
p_op, h_op = func(zv, tv)
# theano graph
p, h = max_pool_c01b(z, (pool_rows, pool_cols), top_down = t)
func = function([z, t], [p, h], mode = mode_without_gpu)
p_th, h_th = func(zv, tv)
assert np.allclose(p_op, p_th)
assert np.allclose(h_op, h_th)
def test_grad_correctness():
"""
Test Op's gradient against theano graph implementation
"""
rng = np.random.RandomState([2012, 7, 19])
batch_size_list = [1, 5, 128]
channels = 16
rows_list = [2, 8, 30]
pool_rows_list = [2, 4, 3]
#TODO theano graph version fails with pool shape 1,1,
# try it with python version
for batch_size in batch_size_list:
for rows, pool_rows in zip(rows_list, pool_rows_list):
cols = rows
pool_cols = pool_rows
zv = rng.randn(channels, rows, cols,
batch_size).astype(config.floatX)
z = T.tensor4()
# gpu op
p, h = prob_max_pool_c01b(z, (pool_rows, pool_cols))
gz = T.grad(h.sum() + p.sum(), z)
func = function([z], gz, mode=mode_with_gpu)
op_gz = func(zv)
# theano graph
p, h = max_pool_c01b(z, (pool_rows, pool_cols))
gz = T.grad(h.sum() + p.sum(), z)
func = function([z], gz, mode=mode_without_gpu)
th_gz = func(zv)
assert np.allclose(op_gz, th_gz, rtol=1e-04, atol=1e-06)
def test_grad_correctness():
"""
Test Op's gradient against theano graph implementation
"""
rng = np.random.RandomState([2012,7,19])
batch_size_list = [1, 5, 128]
channels = 16
rows_list = [2, 8, 30]
pool_rows_list = [2, 4, 3]
#TODO theano graph version fails with pool shape 1,1,
# try it with python version
for batch_size in batch_size_list:
for rows, pool_rows in zip(rows_list, pool_rows_list):
cols = rows
pool_cols = pool_rows
zv = rng.randn(channels, rows, cols, batch_size).astype(config.floatX)
z = T.tensor4()
# gpu op
p, h = prob_max_pool_c01b(z, (pool_rows, pool_cols) )
gz = T.grad(h.sum() + p.sum(), z)
func = function([z], gz, mode = mode_with_gpu)
op_gz = func(zv)
# theano graph
p, h = max_pool_c01b(z, (pool_rows, pool_cols) )
gz = T.grad(h.sum() + p.sum(), z)
func = function([z], gz, mode = mode_without_gpu)
th_gz = func(zv)
assert np.allclose(op_gz, th_gz, rtol=1e-04, atol=1e-06)
import numpy as np from pylearn2.utils import sharedX from pylearn2.expr.probabilistic_max_pooling import max_pool_c01b X = sharedX(np.zeros((16, 34, 34, 2))) P, H = max_pool_c01b(X, pool_shape=[2, 2]) obj = P.sum() + H.sum() from theano import tensor as T from theano import function g = T.grad(obj, X) f = function([], g)
