def test_window_flip_coverage():
ddata = DummyDataset()
topo = ddata.get_topological_view()
ref_win = [set() for _ in xrange(4)]
for b in xrange(topo.shape[-1]):
for i in xrange(3):
for j in xrange(3):
window = topo[:, i:i + 3, j:j + 3, b]
assert_equal((3, 3), window.shape[1:])
ref_win[b].add(_hash_array(window))
ref_win[b].add(_hash_array(window[:, :, ::-1]))
actual_win = [set() for _ in xrange(4)]
wf = WindowAndFlipC01B(window_shape=(3, 3), randomize=[ddata])
wf.setup(None, ddata, None)
curr_topo = ddata.get_topological_view()
assert_equal((2, 3, 3, 4), curr_topo.shape)
for b in xrange(topo.shape[-1]):
hashed = _hash_array(curr_topo[..., b])
assert_contains(ref_win[b], hashed)
actual_win[b].add(hashed)
while not all(len(a) == len(b) for a, b in zip(ref_win, actual_win)):
prev_topo = curr_topo.copy()
wf.on_monitor(None, ddata, None)
curr_topo = ddata.get_topological_view()
assert_(not (prev_topo == curr_topo).all())
for b in xrange(topo.shape[-1]):
hashed = _hash_array(curr_topo[..., b])
assert hashed in ref_win[b]
actual_win[b].add(hashed)
def check_window_flip_coverage_B01C(flip):
ddata = DummyDataset(axes=('b', 0, 1, 'c'))
topo = ddata.get_topological_view()
ref_win = [set() for _ in xrange(4)]
for b in xrange(topo.shape[0]):
for i in xrange(3):
for j in xrange(3):
window = topo[b, i:i + 3, j:j + 3, :]
assert_equal((3, 3), window.shape[:-1])
ref_win[b].add(_hash_array(window))
if flip:
ref_win[b].add(_hash_array(window[:, ::-1, :]))
actual_win = [set() for _ in xrange(4)]
wf = WindowAndFlip(window_shape=(3, 3), randomize=[ddata], flip=flip)
wf.setup(None, ddata, None)
curr_topo = ddata.get_topological_view()
assert_equal((4, 3, 3, 2), curr_topo.shape)
for b in xrange(topo.shape[0]):
hashed = _hash_array(curr_topo[b, ...])
assert_contains(ref_win[b], hashed)
actual_win[b].add(hashed)
while not all(len(a) == len(b) for a, b in zip(ref_win, actual_win)):
prev_topo = curr_topo.copy()
wf.on_monitor(None, ddata, None)
curr_topo = ddata.get_topological_view()
assert_(not (prev_topo == curr_topo).all())
for b in xrange(topo.shape[0]):
hashed = _hash_array(curr_topo[b, ...])
assert_contains(ref_win[b], hashed)
actual_win[b].add(hashed)
def check_padding(axes):
padding = 3
ddata = DummyDataset()
topo = ddata.get_topological_view()
wf_cls = WindowAndFlip
wf = wf_cls(window_shape=(5, 5), randomize=[ddata],
pad_randomized=padding)
wf.setup(None, None, None)
new_topo = ddata.get_topological_view()
assert_equal(topo.shape, new_topo.shape)
saw_padding = dict([((direction, amount), False) for direction, amount
in itertools.product(['l', 'b', 'r', 't'],
xrange(padding))])
iters = 0
while not all(saw_padding.values()) and iters < 50:
for image in new_topo.swapaxes(0, 3):
for i in xrange(padding):
if (image[:i] == 0).all():
saw_padding['t', i] = True
if (image[-i:] == 0).all():
saw_padding['b', i] = True
if (image[:, -i:] == 0).all():
saw_padding['r', i] = True
if (image[:, :i] == 0).all():
saw_padding['l', i] = True
wf.on_monitor(None, None, None)
new_topo = ddata.get_topological_view()
iters += 1
def check_padding(axes, use_old_c01b=False):
padding = 3
ddata = DummyDataset()
topo = ddata.get_topological_view()
if use_old_c01b:
wf_cls = WindowAndFlipC01B
else:
wf_cls = WindowAndFlip
wf = wf_cls(window_shape=(5, 5), randomize=[ddata], pad_randomized=padding)
wf.setup(None, None, None)
new_topo = ddata.get_topological_view()
assert_equal(topo.shape, new_topo.shape)
saw_padding = dict([((direction, amount), False)
for direction, amount in itertools.product(
['l', 'b', 'r', 't'], xrange(padding))])
iters = 0
while not all(saw_padding.values()) and iters < 50:
for image in new_topo.swapaxes(0, 3):
for i in xrange(padding):
if (image[:i] == 0).all():
saw_padding['t', i] = True
if (image[-i:] == 0).all():
saw_padding['b', i] = True
if (image[:, -i:] == 0).all():
saw_padding['r', i] = True
if (image[:, :i] == 0).all():
saw_padding['l', i] = True
wf.on_monitor(None, None, None)
new_topo = ddata.get_topological_view()
iters += 1
def test_conditional_initialize_parameters():
"""
Conditional.initialize_parameters does the following:
* Set its input_space and ndim attributes
* Calls its MLP's set_mlp method
* Sets its MLP's input_space
* Validates its MLP
* Sets its params and param names
"""
mlp = MLP(layers=[Linear(layer_name='h', dim=5, irange=0.01,
max_col_norm=0.01)])
conditional = DummyConditional(mlp=mlp, name='conditional')
vae = DummyVAE()
conditional.set_vae(vae)
input_space = VectorSpace(dim=5)
conditional.initialize_parameters(input_space=input_space, ndim=5)
testing.assert_same_object(input_space, conditional.input_space)
testing.assert_equal(conditional.ndim, 5)
testing.assert_same_object(mlp.get_mlp(), conditional)
testing.assert_same_object(mlp.input_space, input_space)
mlp_params = mlp.get_params()
conditional_params = conditional.get_params()
assert all([mp in conditional_params for mp in mlp_params])
assert all([cp in mlp_params for cp in conditional_params])
def check_window_flip_coverage_B01C(flip):
ddata = DummyDataset(axes=('b', 0, 1, 'c'))
topo = ddata.get_topological_view()
ref_win = [set() for _ in xrange(4)]
for b in xrange(topo.shape[0]):
for i in xrange(3):
for j in xrange(3):
window = topo[b, i:i + 3, j:j + 3, :]
assert_equal((3, 3), window.shape[:-1])
ref_win[b].add(_hash_array(window))
if flip:
ref_win[b].add(_hash_array(window[:, ::-1, :]))
actual_win = [set() for _ in xrange(4)]
wf = WindowAndFlip(window_shape=(3, 3), randomize=[ddata], flip=flip)
wf.setup(None, ddata, None)
curr_topo = ddata.get_topological_view()
assert_equal((4, 3, 3, 2), curr_topo.shape)
for b in xrange(topo.shape[0]):
hashed = _hash_array(curr_topo[b, ...])
assert_contains(ref_win[b], hashed)
actual_win[b].add(hashed)
while not all(len(a) == len(b) for a, b in zip(ref_win, actual_win)):
prev_topo = curr_topo.copy()
wf.on_monitor(None, ddata, None)
curr_topo = ddata.get_topological_view()
assert_(not (prev_topo == curr_topo).all())
for b in xrange(topo.shape[0]):
hashed = _hash_array(curr_topo[b, ...])
assert_contains(ref_win[b], hashed)
actual_win[b].add(hashed)
def test_WindowAndFlipC01B_axes_guard():
ddata = DummyDataset(axes=('b', 0, 1, 'c'))
raised_error = False
try:
wf = WindowAndFlipC01B(window_shape=(3, 3), randomize=[ddata])
except ValueError:
raised_error = True
assert_equal(raised_error, True)
def test_WindowAndFlipC01B_axes_guard():
ddata = DummyDataset(axes=('b', 0, 1, 'c'))
raised_error = False
try:
wf = WindowAndFlipC01B(window_shape=(3, 3), randomize=[ddata])
except ValueError:
raised_error = True
assert_equal(raised_error, True)
def test_conditional_returns_lr_scalers():
"""
Conditional.get_lr_scalers calls its MLP's get_lr_scalers method
"""
mlp = MLP(layers=[Linear(layer_name="h", dim=5, irange=0.01, W_lr_scale=0.01)])
conditional = DummyConditional(mlp=mlp, name="conditional")
vae = DummyVAE()
conditional.set_vae(vae)
conditional.initialize_parameters(input_space=VectorSpace(dim=5), ndim=5)
testing.assert_equal(conditional.get_lr_scalers(), mlp.get_lr_scalers())
def test_conditional_returns_lr_scalers():
"""
Conditional.get_lr_scalers calls its MLP's get_lr_scalers method
"""
mlp = MLP(layers=[Linear(layer_name='h', dim=5, irange=0.01,
W_lr_scale=0.01)])
conditional = DummyConditional(mlp=mlp, name='conditional')
vae = DummyVAE()
conditional.set_vae(vae)
conditional.initialize_parameters(input_space=VectorSpace(dim=5), ndim=5)
testing.assert_equal(conditional.get_lr_scalers(), mlp.get_lr_scalers())
def test_prior_set_vae():
"""
Prior.set_vae adds a reference to the vae and adopts the vae's rng
and batch_size attributes
"""
prior = DummyPrior()
vae = DummyVAE()
prior.set_vae(vae)
testing.assert_same_object(prior.vae, vae)
testing.assert_same_object(prior.rng, vae.rng)
testing.assert_equal(prior.batch_size, vae.batch_size)
def test_prior_set_vae():
"""
Prior.set_vae adds a reference to the vae and adopts the vae's rng
and batch_size attributes
"""
prior = DummyPrior()
vae = DummyVAE()
prior.set_vae(vae)
testing.assert_same_object(prior.vae, vae)
testing.assert_same_object(prior.rng, vae.rng)
testing.assert_equal(prior.batch_size, vae.batch_size)
def test_conditional_modify_updates():
"""
Conditional.modify_updates calls its MLP's modify_updates method
"""
mlp = MLP(layers=[Linear(layer_name="h", dim=5, irange=0.01, max_col_norm=0.01)])
conditional = DummyConditional(mlp=mlp, name="conditional")
vae = DummyVAE()
conditional.set_vae(vae)
conditional.initialize_parameters(input_space=VectorSpace(dim=5), ndim=5)
updates = OrderedDict(zip(mlp.get_params(), mlp.get_params()))
testing.assert_equal(conditional.modify_updates(updates), mlp.modify_updates(updates))
def test_conditional_set_vae():
"""
Conditional.set_vae adds a reference to the vae and adopts the vae's rng
and batch_size attributes
"""
mlp = MLP(layers=[Linear(layer_name='h', dim=5, irange=0.01)])
conditional = DummyConditional(mlp=mlp, name='conditional')
vae = DummyVAE()
conditional.set_vae(vae)
testing.assert_same_object(conditional.vae, vae)
testing.assert_same_object(conditional.rng, vae.rng)
testing.assert_equal(conditional.batch_size, vae.batch_size)
def test_conditional_set_vae():
"""
Conditional.set_vae adds a reference to the vae and adopts the vae's rng
and batch_size attributes
"""
mlp = MLP(layers=[Linear(layer_name="h", dim=5, irange=0.01)])
conditional = DummyConditional(mlp=mlp, name="conditional")
vae = DummyVAE()
conditional.set_vae(vae)
testing.assert_same_object(conditional.vae, vae)
testing.assert_same_object(conditional.rng, vae.rng)
testing.assert_equal(conditional.batch_size, vae.batch_size)
def test_conditional_modify_updates():
"""
Conditional.modify_updates calls its MLP's modify_updates method
"""
mlp = MLP(layers=[Linear(layer_name='h', dim=5, irange=0.01,
max_col_norm=0.01)])
conditional = DummyConditional(mlp=mlp, name='conditional')
vae = DummyVAE()
conditional.set_vae(vae)
conditional.initialize_parameters(input_space=VectorSpace(dim=5), ndim=5)
updates = OrderedDict(zip(mlp.get_params(), mlp.get_params()))
testing.assert_equal(conditional.modify_updates(updates),
mlp.modify_updates(updates))
def check_window_flip_coverage_C01B(flip, use_old_c01b=False):
# 4 5x5x2 images (stored in a 2x5x5x4 tensor)
ddata = DummyDataset(axes=('c', 0, 1, 'b'))
topo = ddata.get_topological_view()
# ref_win[b]: a set of hashes, computed from all possible 3x3 windows of
# topo[..., b].
ref_win = [set() for _ in xrange(4)]
for b in xrange(topo.shape[-1]):
# get all possible 3x3 windows within the 5x5 images.
for i in xrange(3):
for j in xrange(3):
window = topo[:, i:i + 3, j:j + 3, b]
assert_equal((3, 3), window.shape[1:])
# Add a SHA1 digest of the window to set ref_win[b]
ref_win[b].add(_hash_array(window))
if flip:
# Also add the hash of the window with axis 1 flipped
ref_win[b].add(_hash_array(window[:, :, ::-1]))
actual_win = [set() for _ in xrange(4)]
if use_old_c01b:
wf_cls = WindowAndFlipC01B
else:
wf_cls = WindowAndFlip
# no zero-padding.
wf = wf_cls(window_shape=(3, 3), randomize=[ddata], flip=flip)
wf.setup(None, ddata, None) # ddata argument is ignored
curr_topo = ddata.get_topological_view()
assert_equal((2, 3, 3, 4), curr_topo.shape)
for b in xrange(topo.shape[-1]):
hashed = _hash_array(curr_topo[..., b])
assert_contains(ref_win[b], hashed)
actual_win[b].add(hashed)
while not all(len(a) == len(b) for a, b in zip(ref_win, actual_win)):
prev_topo = curr_topo.copy()
wf.on_monitor(None, ddata, None)
curr_topo = ddata.get_topological_view()
assert_(not (prev_topo == curr_topo).all())
for b in xrange(topo.shape[-1]):
hashed = _hash_array(curr_topo[..., b])
assert_contains(ref_win[b], hashed)
actual_win[b].add(hashed)
def check_window_flip_coverage_C01B(flip, use_old_c01b=False):
# 4 5x5x2 images (stored in a 2x5x5x4 tensor)
ddata = DummyDataset(axes=('c', 0, 1, 'b'))
topo = ddata.get_topological_view()
# ref_win[b]: a set of hashes, computed from all possible 3x3 windows of
# topo[..., b].
ref_win = [set() for _ in xrange(4)]
for b in xrange(topo.shape[-1]):
# get all possible 3x3 windows within the 5x5 images.
for i in xrange(3):
for j in xrange(3):
window = topo[:, i:i + 3, j:j + 3, b]
assert_equal((3, 3), window.shape[1:])
# Add a SHA1 digest of the window to set ref_win[b]
ref_win[b].add(_hash_array(window))
if flip:
# Also add the hash of the window with axis 1 flipped
ref_win[b].add(_hash_array(window[:, :, ::-1]))
actual_win = [set() for _ in xrange(4)]
if use_old_c01b:
wf_cls = WindowAndFlipC01B
else:
wf_cls = WindowAndFlip
# no zero-padding.
wf = wf_cls(window_shape=(3, 3), randomize=[ddata], flip=flip)
wf.setup(None, ddata, None) # ddata argument is ignored
curr_topo = ddata.get_topological_view()
assert_equal((2, 3, 3, 4), curr_topo.shape)
for b in xrange(topo.shape[-1]):
hashed = _hash_array(curr_topo[..., b])
assert_contains(ref_win[b], hashed)
actual_win[b].add(hashed)
while not all(len(a) == len(b) for a, b in zip(ref_win, actual_win)):
prev_topo = curr_topo.copy()
wf.on_monitor(None, ddata, None)
curr_topo = ddata.get_topological_view()
assert_(not (prev_topo == curr_topo).all())
for b in xrange(topo.shape[-1]):
hashed = _hash_array(curr_topo[..., b])
assert_contains(ref_win[b], hashed)
actual_win[b].add(hashed)
def check_window_flip_coverage_C01B(flip, use_old_c01b=False):
ddata = DummyDataset(axes=('c', 0, 1, 'b'))
topo = ddata.get_topological_view()
ref_win = [set() for _ in xrange(4)]
for b in xrange(topo.shape[-1]):
for i in xrange(3):
for j in xrange(3):
window = topo[:, i:i + 3, j:j + 3, b]
assert_equal((3, 3), window.shape[1:])
ref_win[b].add(_hash_array(window))
if flip:
ref_win[b].add(_hash_array(window[:, :, ::-1]))
actual_win = [set() for _ in xrange(4)]
if use_old_c01b:
wf_cls = WindowAndFlipC01B
else:
wf_cls = WindowAndFlip
wf = wf_cls(window_shape=(3, 3), randomize=[ddata], flip=flip)
wf.setup(None, ddata, None)
curr_topo = ddata.get_topological_view()
assert_equal((2, 3, 3, 4), curr_topo.shape)
for b in xrange(topo.shape[-1]):
hashed = _hash_array(curr_topo[..., b])
assert_contains(ref_win[b], hashed)
actual_win[b].add(hashed)
while not all(len(a) == len(b) for a, b in zip(ref_win, actual_win)):
prev_topo = curr_topo.copy()
wf.on_monitor(None, ddata, None)
curr_topo = ddata.get_topological_view()
assert_(not (prev_topo == curr_topo).all())
for b in xrange(topo.shape[-1]):
hashed = _hash_array(curr_topo[..., b])
assert_contains(ref_win[b], hashed)
actual_win[b].add(hashed)
def check_window_flip_coverage_C01B(flip, use_old_c01b=False):
ddata = DummyDataset(axes=('c', 0, 1, 'b'))
topo = ddata.get_topological_view()
ref_win = [set() for _ in xrange(4)]
for b in xrange(topo.shape[-1]):
for i in xrange(3):
for j in xrange(3):
window = topo[:, i:i + 3, j:j + 3, b]
assert_equal((3, 3), window.shape[1:])
ref_win[b].add(_hash_array(window))
if flip:
ref_win[b].add(_hash_array(window[:, :, ::-1]))
actual_win = [set() for _ in xrange(4)]
if use_old_c01b:
wf_cls = WindowAndFlipC01B
else:
wf_cls = WindowAndFlip
wf = wf_cls(window_shape=(3, 3), randomize=[ddata], flip=flip)
wf.setup(None, ddata, None)
curr_topo = ddata.get_topological_view()
assert_equal((2, 3, 3, 4), curr_topo.shape)
for b in xrange(topo.shape[-1]):
hashed = _hash_array(curr_topo[..., b])
assert_contains(ref_win[b], hashed)
actual_win[b].add(hashed)
while not all(len(a) == len(b) for a, b in zip(ref_win, actual_win)):
prev_topo = curr_topo.copy()
wf.on_monitor(None, ddata, None)
curr_topo = ddata.get_topological_view()
assert_(not (prev_topo == curr_topo).all())
for b in xrange(topo.shape[-1]):
hashed = _hash_array(curr_topo[..., b])
assert_contains(ref_win[b], hashed)
actual_win[b].add(hashed)
def test_conditional_initialize_parameters():
"""
Conditional.initialize_parameters does the following:
* Set its input_space and ndim attributes
* Calls its MLP's set_mlp method
* Sets its MLP's input_space
* Validates its MLP
* Sets its params and param names
"""
mlp = MLP(layers=[Linear(layer_name="h", dim=5, irange=0.01, max_col_norm=0.01)])
conditional = DummyConditional(mlp=mlp, name="conditional")
vae = DummyVAE()
conditional.set_vae(vae)
input_space = VectorSpace(dim=5)
conditional.initialize_parameters(input_space=input_space, ndim=5)
testing.assert_same_object(input_space, conditional.input_space)
testing.assert_equal(conditional.ndim, 5)
testing.assert_same_object(mlp.get_mlp(), conditional)
testing.assert_same_object(mlp.input_space, input_space)
mlp_params = mlp.get_params()
conditional_params = conditional.get_params()
assert all([mp in conditional_params for mp in mlp_params])
assert all([cp in mlp_params for cp in conditional_params])
