def test_nested_parallel(self, same_inputs, kernel_type):
platform = default_backend()
rtol = RTOL if platform != 'tpu' else 0.05
rng = random.PRNGKey(0)
(input_key1,
input_key2,
input_key3,
input_key4,
mask_key,
mc_key) = random.split(rng, 6)
x1_1, x2_1 = _get_inputs(input_key1, same_inputs, (BATCH_SIZE, 5))
x1_2, x2_2 = _get_inputs(input_key2, same_inputs, (BATCH_SIZE, 2, 2, 2))
x1_3, x2_3 = _get_inputs(input_key3, same_inputs, (BATCH_SIZE, 2, 2, 3))
x1_4, x2_4 = _get_inputs(input_key4, same_inputs, (BATCH_SIZE, 3, 4))
m1_key, m2_key, m3_key, m4_key = random.split(mask_key, 4)
x1_1 = test_utils.mask(
x1_1, mask_constant=-1, mask_axis=(1,), key=m1_key, p=0.5)
x1_2 = test_utils.mask(
x1_2, mask_constant=-1, mask_axis=(2, 3,), key=m2_key, p=0.5)
if not same_inputs:
x2_3 = test_utils.mask(
x2_3, mask_constant=-1, mask_axis=(1, 3,), key=m3_key, p=0.5)
x2_4 = test_utils.mask(
x2_4, mask_constant=-1, mask_axis=(2,), key=m4_key, p=0.5)
x1 = (((x1_1, x1_2), x1_3), x1_4)
x2 = (((x2_1, x2_2), x2_3), x2_4) if not same_inputs else None
N_in = 2 ** 7
# We only include dropout on non-TPU backends, because it takes large N to
# converge on TPU.
dropout_or_id = stax.Dropout(0.9) if platform != 'tpu' else stax.Identity()
init_fn, apply_fn, kernel_fn = stax.parallel(
stax.parallel(
stax.parallel(stax.Dense(N_in),
stax.serial(stax.Conv(N_in + 1, (2, 2)),
stax.Flatten())),
stax.serial(stax.Conv(N_in + 2, (2, 2)),
dropout_or_id,
stax.GlobalAvgPool())),
stax.Conv(N_in + 3, (2,)))
kernel_fn_empirical = nt.monte_carlo_kernel_fn(
init_fn, apply_fn, mc_key, N_SAMPLES, implementation=2,
vmap_axes=(((((0, 0), 0), 0), (((0, 0), 0), 0), {})
if platform == 'tpu' else None)
)
test_utils.assert_close_matrices(
self,
kernel_fn(x1, x2, get=kernel_type, mask_constant=-1),
kernel_fn_empirical(x1, x2, get=kernel_type, mask_constant=-1),
rtol)
def _get_net(W_std, b_std, filter_shape, is_conv, use_pooling, is_res, padding,
phi, strides, width, is_ntk, proj_into_2d, layer_norm,
parameterization, use_dropout):
fc = partial(stax.Dense,
W_std=W_std,
b_std=b_std,
parameterization=parameterization)
conv = partial(stax.Conv,
filter_shape=filter_shape,
strides=strides,
padding=padding,
W_std=W_std,
b_std=b_std,
parameterization=parameterization)
affine = conv(width) if is_conv else fc(width)
rate = np.onp.random.uniform(0.5, 0.9)
dropout = stax.Dropout(rate, mode='train')
ave_pool = stax.AvgPool((2, 3), None,
'SAME' if padding == 'SAME' else 'CIRCULAR')
ave_pool_or_identity = ave_pool if use_pooling else stax.Identity()
dropout_or_identity = dropout if use_dropout else stax.Identity()
layer_norm_or_identity = (stax.Identity() if layer_norm is None else
stax.LayerNorm(axis=layer_norm))
res_unit = stax.serial(ave_pool_or_identity, phi, dropout_or_identity,
affine)
if is_res:
block = stax.serial(affine, stax.FanOut(2),
stax.parallel(stax.Identity(), res_unit),
stax.FanInSum(), layer_norm_or_identity)
else:
block = stax.serial(affine, res_unit, layer_norm_or_identity)
if proj_into_2d == 'FLAT':
proj_layer = stax.Flatten()
elif proj_into_2d == 'POOL':
proj_layer = stax.GlobalAvgPool()
elif proj_into_2d.startswith('ATTN'):
n_heads = int(np.sqrt(width))
n_chan_val = int(np.round(float(width) / n_heads))
fixed = proj_into_2d == 'ATTN_FIXED'
proj_layer = stax.serial(
stax.GlobalSelfAttention(width,
n_chan_key=width,
n_chan_val=n_chan_val,
n_heads=n_heads,
fixed=fixed,
W_key_std=W_std,
W_value_std=W_std,
W_query_std=W_std,
W_out_std=1.0,
b_std=b_std), stax.Flatten())
else:
raise ValueError(proj_into_2d)
readout = stax.serial(proj_layer, fc(1 if is_ntk else width))
return stax.serial(block, readout)
def test_mask_conv(self, same_inputs, get, mask_axis, mask_constant,
concat, proj, p, n, transpose):
if isinstance(concat, int) and concat > n:
raise absltest.SkipTest('Concatenation axis out of bounds.')
test_utils.skip_test(self)
if default_backend() == 'gpu' and n > 3:
raise absltest.SkipTest('>=4D-CNN is not supported on GPUs.')
width = 256
n_samples = 256
tol = 0.03
key = random.PRNGKey(1)
spatial_shape = ((1, 2, 3, 2, 1) if transpose else (15, 8, 9))[:n]
filter_shape = ((2, 3, 1, 2, 1) if transpose else (7, 2, 3))[:n]
strides = (2, 1, 3, 2, 3)[:n]
spatial_spec = 'HWDZX'[:n]
dimension_numbers = ('N' + spatial_spec + 'C', 'OI' + spatial_spec,
'N' + spatial_spec + 'C')
x1 = np.cos(random.normal(key, (2, ) + spatial_shape + (2, )))
x1 = test_utils.mask(x1, mask_constant, mask_axis, key, p)
if same_inputs:
x2 = None
else:
x2 = np.cos(random.normal(key, (4, ) + spatial_shape + (2, )))
x2 = test_utils.mask(x2, mask_constant, mask_axis, key, p)
def get_attn():
return stax.GlobalSelfAttention(
n_chan_out=width,
n_chan_key=width,
n_chan_val=int(np.round(float(width) / int(np.sqrt(width)))),
n_heads=int(np.sqrt(width)),
) if proj == 'avg' else stax.Identity()
conv = stax.ConvTranspose if transpose else stax.Conv
nn = stax.serial(
stax.FanOut(3),
stax.parallel(
stax.serial(
conv(dimension_numbers=dimension_numbers,
out_chan=width,
strides=strides,
filter_shape=filter_shape,
padding='CIRCULAR',
W_std=1.5,
b_std=0.2),
stax.LayerNorm(axis=(1, -1)),
stax.Abs(),
stax.DotGeneral(rhs=0.9),
conv(dimension_numbers=dimension_numbers,
out_chan=width,
strides=strides,
filter_shape=filter_shape,
padding='VALID',
W_std=1.2,
b_std=0.1),
),
stax.serial(
conv(dimension_numbers=dimension_numbers,
out_chan=width,
strides=strides,
filter_shape=filter_shape,
padding='SAME',
W_std=0.1,
b_std=0.3),
stax.Relu(),
stax.Dropout(0.7),
conv(dimension_numbers=dimension_numbers,
out_chan=width,
strides=strides,
filter_shape=filter_shape,
padding='VALID',
W_std=0.9,
b_std=1.),
),
stax.serial(
get_attn(),
conv(dimension_numbers=dimension_numbers,
out_chan=width,
strides=strides,
filter_shape=filter_shape,
padding='CIRCULAR',
W_std=1.,
b_std=0.1),
stax.Erf(),
stax.Dropout(0.2),
stax.DotGeneral(rhs=0.7),
conv(dimension_numbers=dimension_numbers,
out_chan=width,
strides=strides,
filter_shape=filter_shape,
padding='VALID',
W_std=1.,
b_std=0.1),
)),
(stax.FanInSum() if concat is None else stax.FanInConcat(concat)),
get_attn(),
{
'avg': stax.GlobalAvgPool(),
'sum': stax.GlobalSumPool(),
'flatten': stax.Flatten(),
}[proj],
)
if get == 'nngp':
init_fn, apply_fn, kernel_fn = stax.serial(
nn, stax.Dense(width, 1., 0.))
elif get == 'ntk':
init_fn, apply_fn, kernel_fn = stax.serial(nn,
stax.Dense(1, 1., 0.))
else:
raise ValueError(get)
kernel_fn_mc = nt.monte_carlo_kernel_fn(
init_fn,
apply_fn,
key,
n_samples,
device_count=0 if concat in (0, -n) else -1,
implementation=_DEFAULT_TESTING_NTK_IMPLEMENTATION,
vmap_axes=None if concat in (0, -n) else 0,
)
kernel_fn = jit(kernel_fn, static_argnames='get')
exact = kernel_fn(x1, x2, get, mask_constant=mask_constant)
empirical = kernel_fn_mc(x1, x2, get=get, mask_constant=mask_constant)
test_utils.assert_close_matrices(self, empirical, exact, tol)
def _get_net(W_std, b_std, filter_shape, is_conv, use_pooling, is_res, padding,
phi, strides, width, is_ntk, proj_into_2d, pool_type, layer_norm,
parameterization, s, use_dropout):
if is_conv:
# Select a random filter order.
default_filter_spec = 'HW'
filter_specs = [''.join(p) for p in itertools.permutations('HWIO')]
filter_spec = prandom.choice(filter_specs)
filter_shape = tuple(filter_shape[default_filter_spec.index(c)]
for c in filter_spec if c in default_filter_spec)
strides = tuple(strides[default_filter_spec.index(c)]
for c in filter_spec if c in default_filter_spec)
# Select the activation order.
default_spec = 'NHWC'
if default_backend() == 'tpu':
# Keep batch dimension leading for TPU for batching to work.
specs = ['N' + ''.join(p) for p in itertools.permutations('CHW')]
else:
specs = [''.join(p) for p in itertools.permutations('NCHW')]
spec = prandom.choice(specs)
input_shape = tuple(INPUT_SHAPE[default_spec.index(c)] for c in spec)
else:
input_shape = (INPUT_SHAPE[0], onp.prod(INPUT_SHAPE[1:]))
if default_backend() == 'tpu':
spec = 'NC'
else:
spec = prandom.choice(['NC', 'CN'])
if spec.index('N') == 1:
input_shape = input_shape[::-1]
filter_spec = None
dimension_numbers = (spec, filter_spec, spec)
batch_axis, channel_axis = spec.index('N'), spec.index('C')
spec_fc = ''.join(c for c in spec if c in ('N', 'C'))
batch_axis_fc, channel_axis_fc = spec_fc.index('N'), spec_fc.index('C')
if not is_conv:
batch_axis = batch_axis_fc
channel_axis = channel_axis_fc
if layer_norm:
layer_norm = tuple(spec.index(c) for c in layer_norm)
def fc(out_dim, s):
return stax.Dense(
out_dim=out_dim,
W_std=W_std,
b_std=b_std,
parameterization=parameterization,
s=s,
batch_axis=batch_axis_fc,
channel_axis=channel_axis_fc
)
def conv(out_chan, s):
return stax.Conv(
out_chan=out_chan,
filter_shape=filter_shape,
strides=strides,
padding=padding,
W_std=W_std,
b_std=b_std,
dimension_numbers=dimension_numbers,
parameterization=parameterization,
s=s
)
affine = conv(width, (s, s)) if is_conv else fc(width, (s, s))
affine_bottom = conv(width, (1, s)) if is_conv else fc(width, (1, s))
rate = onp.random.uniform(0.5, 0.9)
dropout = stax.Dropout(rate, mode='train')
if pool_type == 'AVG':
pool_fn = stax.AvgPool
global_pool_fn = stax.GlobalAvgPool
elif pool_type == 'SUM':
pool_fn = stax.SumPool
global_pool_fn = stax.GlobalSumPool
else:
raise ValueError(pool_type)
if use_pooling:
pool_or_identity = pool_fn((2, 3),
None,
'SAME' if padding == 'SAME' else 'CIRCULAR',
batch_axis=batch_axis,
channel_axis=channel_axis)
else:
pool_or_identity = stax.Identity()
dropout_or_identity = dropout if use_dropout else stax.Identity()
layer_norm_or_identity = (stax.Identity() if layer_norm is None else
stax.LayerNorm(axis=layer_norm,
batch_axis=batch_axis,
channel_axis=channel_axis))
res_unit = stax.serial(dropout_or_identity, affine, pool_or_identity)
if is_res:
block = stax.serial(
affine_bottom,
stax.FanOut(2),
stax.parallel(stax.Identity(),
res_unit),
stax.FanInSum(),
layer_norm_or_identity,
phi)
else:
block = stax.serial(
affine_bottom,
res_unit,
layer_norm_or_identity,
phi)
if proj_into_2d == 'FLAT':
proj_layer = stax.Flatten(batch_axis, batch_axis_fc)
elif proj_into_2d == 'POOL':
proj_layer = global_pool_fn(batch_axis, channel_axis)
elif proj_into_2d.startswith('ATTN'):
n_heads = int(np.sqrt(width))
n_chan_val = int(np.round(float(width) / n_heads))
proj_layer = stax.serial(
stax.GlobalSelfAttention(
n_chan_out=width,
n_chan_key=width,
n_chan_val=n_chan_val,
n_heads=n_heads,
linear_scaling=True,
W_key_std=W_std,
W_value_std=W_std,
W_query_std=W_std,
W_out_std=1.0,
b_std=b_std,
batch_axis=batch_axis,
channel_axis=channel_axis),
stax.Flatten(batch_axis, batch_axis_fc))
else:
raise ValueError(proj_into_2d)
readout = stax.serial(proj_layer,
fc(1 if is_ntk else width, (s, 1 if is_ntk else s)))
device_count = -1 if spec.index('N') == 0 else 0
net = stax.serial(block, readout)
return net, input_shape, device_count, channel_axis_fc
def bann_model(
W_std,
b_std,
first_layer_width,
second_layer_width,
subNN_num,
keep_rate,
activation,
parameterization
):
"""Construct fully connected NN model and infinite width NTK & NNGP kernel
function.
Args:
W_std (float): Weight standard deviation.
b_std (float): Bias standard deviation.
first_layer_width (int): First Hidden layer width.
second_layer_width (int): Second Hidden layer width.
subNN_num (int) : Number of sub neural networks in the architecture
keep_rate (float): 1 - Dropout rate.
activation (string): Activation function string, 'erf' or 'relu'.
parameterization (string): Parameterization string, 'ntk' or 'standard'.
Returns:
`(init_fn, apply_fn, kernel_fn)`
"""
act = activation_fn(activation)
# multi-task learning
# Computational Skeleton Block
CSB = stax.serial(
stax.FanOut(subNN_num),
stax.parallel(
stax.serial(
Dense(first_layer_width, W_std, b_std, parameterization=parameterization), act(),
Dense(second_layer_width, W_std, b_std, parameterization=parameterization), act(),
stax.Dropout(keep_rate)
),
stax.serial(
Dense(first_layer_width, W_std, b_std, parameterization=parameterization), act(),
Dense(2 * second_layer_width, W_std, b_std, parameterization=parameterization), act(),
stax.Dropout(keep_rate)
),
stax.serial(
Dense(first_layer_width, W_std, b_std, parameterization=parameterization), act(),
Dense(3 * second_layer_width, W_std, b_std, parameterization=parameterization), act(),
stax.Dropout(keep_rate)
),
stax.serial(
Dense(first_layer_width, W_std, b_std, parameterization=parameterization), act(),
Dense(4 * second_layer_width, W_std, b_std, parameterization=parameterization), act(),
stax.Dropout(keep_rate)
),
stax.serial(
Dense(first_layer_width, W_std, b_std, parameterization=parameterization), act(),
Dense(5 * second_layer_width, W_std, b_std, parameterization=parameterization), act(),
stax.Dropout(keep_rate)
),
stax.serial(
Dense(first_layer_width, W_std, b_std, parameterization=parameterization), act(),
Dense(6 * second_layer_width, W_std, b_std, parameterization=parameterization), act(),
stax.Dropout(keep_rate)
),
stax.serial(
Dense(first_layer_width, W_std, b_std, parameterization=parameterization), act(),
Dense(7 * second_layer_width, W_std, b_std, parameterization=parameterization), act(),
stax.Dropout(keep_rate)
),
stax.serial(
Dense(first_layer_width, W_std, b_std, parameterization=parameterization), act(),
Dense(8 * second_layer_width, W_std, b_std, parameterization=parameterization), act(),
stax.Dropout(keep_rate)
),
stax.serial(
Dense(first_layer_width, W_std, b_std, parameterization=parameterization), act(),
Dense(9 * second_layer_width, W_std, b_std, parameterization=parameterization), act(),
stax.Dropout(keep_rate)
),
stax.serial(
Dense(first_layer_width, W_std, b_std, parameterization=parameterization), act(),
Dense(10 * second_layer_width, W_std, b_std, parameterization=parameterization), act(),
stax.Dropout(keep_rate)
)
),
stax.FanInConcat()
)
Additive = stax.serial(
stax.FanOut(2),
stax.parallel(
stax.serial(
CSB,
stax.Dropout(keep_rate)
),
stax.serial(
CSB,
stax.Dropout(keep_rate)
)
),
stax.FanInConcat()
)
init_fn, apply_fn, kernel_fn = stax.serial(
Additive,
Dense(1, W_std, b_std, parameterization=parameterization)
)
apply_fn = jit(apply_fn)
return init_fn, apply_fn, kernel_fn
def _get_net(W_std, b_std, filter_shape, is_conv, use_pooling, is_res, padding,
phi, strides, width, is_ntk, proj_into_2d, pool_type, layer_norm,
parameterization, use_dropout):
if is_conv:
# Select a random dimension order.
default_spec = 'NHWC'
if xla_bridge.get_backend().platform == 'tpu':
# Keep batch dimension leading for TPU for batching to work.
specs = ['NHWC', 'NHCW', 'NCHW']
else:
specs = ['NHWC', 'NHCW', 'NCHW', 'CHWN', 'CHNW', 'CNHW']
spec = prandom.choice(specs)
input_shape = tuple(INPUT_SHAPE[default_spec.index(c)] for c in spec)
if layer_norm:
layer_norm = tuple(spec.index(c) for c in layer_norm)
else:
# Only `NC` dimension order is supported and is enforced by layers.
spec = None
input_shape = INPUT_SHAPE
if layer_norm:
layer_norm = prandom.choice([(1,), (-1,)])
dimension_numbers = (spec, 'HWIO', spec)
fc = partial(
stax.Dense, W_std=W_std, b_std=b_std, parameterization=parameterization)
def conv(out_chan): return stax.GeneralConv(
dimension_numbers=dimension_numbers,
out_chan=out_chan,
filter_shape=filter_shape,
strides=strides,
padding=padding,
W_std=W_std,
b_std=b_std,
parameterization=parameterization
)
affine = conv(width) if is_conv else fc(width)
spec = dimension_numbers[-1]
rate = np.onp.random.uniform(0.5, 0.9)
dropout = stax.Dropout(rate, mode='train')
if pool_type == 'AVG':
pool_fn = stax.AvgPool
globalPool_fn = stax.GlobalAvgPool
elif pool_type == 'SUM':
pool_fn = stax.SumPool
globalPool_fn = stax.GlobalSumPool
if use_pooling:
pool_or_identity = pool_fn((2, 3),
None,
'SAME' if padding == 'SAME' else 'CIRCULAR',
spec=spec)
else:
pool_or_identity = stax.Identity()
dropout_or_identity = dropout if use_dropout else stax.Identity()
layer_norm_or_identity = (stax.Identity() if layer_norm is None
else stax.LayerNorm(axis=layer_norm, spec=spec))
res_unit = stax.serial(pool_or_identity, phi, dropout_or_identity, affine)
if is_res:
block = stax.serial(
affine,
stax.FanOut(2),
stax.parallel(stax.Identity(),
res_unit),
stax.FanInSum(),
layer_norm_or_identity)
else:
block = stax.serial(
affine,
res_unit,
layer_norm_or_identity)
if proj_into_2d == 'FLAT':
proj_layer = stax.Flatten(spec=spec)
elif proj_into_2d == 'POOL':
proj_layer = globalPool_fn(spec=spec)
elif proj_into_2d.startswith('ATTN'):
n_heads = int(np.sqrt(width))
n_chan_val = int(np.round(float(width) / n_heads))
fixed = proj_into_2d == 'ATTN_FIXED'
proj_layer = stax.serial(
stax.GlobalSelfAttention(
n_chan_out=width,
n_chan_key=width,
n_chan_val=n_chan_val,
n_heads=n_heads,
fixed=fixed,
W_key_std=W_std,
W_value_std=W_std,
W_query_std=W_std,
W_out_std=1.0,
b_std=b_std,
spec=spec), stax.Flatten(spec=spec))
else:
raise ValueError(proj_into_2d)
readout = stax.serial(proj_layer, fc(1 if is_ntk else width))
return stax.serial(block, readout), input_shape
def test_kwargs(self, do_batch, mode):
rng = random.PRNGKey(1)
x_train = random.normal(rng, (8, 7, 10))
x_test = random.normal(rng, (4, 7, 10))
y_train = random.normal(rng, (8, 1))
rng_train, rng_test = random.split(rng, 2)
pattern_train = random.normal(rng, (8, 7, 7))
pattern_test = random.normal(rng, (4, 7, 7))
init_fn, apply_fn, kernel_fn = stax.serial(
stax.Dense(8),
stax.Relu(),
stax.Dropout(rate=0.4),
stax.Aggregate(),
stax.GlobalAvgPool(),
stax.Dense(1)
)
kw_dd = dict(pattern=(pattern_train, pattern_train))
kw_td = dict(pattern=(pattern_test, pattern_train))
kw_tt = dict(pattern=(pattern_test, pattern_test))
if mode == 'mc':
kernel_fn = monte_carlo_kernel_fn(init_fn, apply_fn, rng, 2,
batch_size=2 if do_batch else 0)
elif mode == 'empirical':
kernel_fn = empirical_kernel_fn(apply_fn)
if do_batch:
raise absltest.SkipTest('Batching of empirical kernel is not '
'implemented with keyword arguments.')
for kw in (kw_dd, kw_td, kw_tt):
kw.update(dict(params=init_fn(rng, x_train.shape)[1],
get=('nngp', 'ntk')))
kw_dd.update(dict(rng=(rng_train, None)))
kw_td.update(dict(rng=(rng_test, rng_train)))
kw_tt.update(dict(rng=(rng_test, None)))
elif mode == 'analytic':
if do_batch:
kernel_fn = batch.batch(kernel_fn, batch_size=2)
else:
raise ValueError(mode)
k_dd = kernel_fn(x_train, None, **kw_dd)
k_td = kernel_fn(x_test, x_train, **kw_td)
k_tt = kernel_fn(x_test, None, **kw_tt)
# Infinite time NNGP/NTK.
predict_fn_gp = predict.gp_inference(k_dd, y_train)
out_gp = predict_fn_gp(k_test_train=k_td, nngp_test_test=k_tt.nngp)
if mode == 'empirical':
for kw in (kw_dd, kw_td, kw_tt):
kw.pop('get')
predict_fn_ensemble = predict.gradient_descent_mse_ensemble(kernel_fn,
x_train,
y_train,
**kw_dd)
out_ensemble = predict_fn_ensemble(x_test=x_test, compute_cov=True, **kw_tt)
self.assertAllClose(out_gp, out_ensemble)
# Finite time NTK test.
predict_fn_mse = predict.gradient_descent_mse(k_dd.ntk, y_train)
out_mse = predict_fn_mse(t=1.,
fx_train_0=None,
fx_test_0=0.,
k_test_train=k_td.ntk)
out_ensemble = predict_fn_ensemble(t=1.,
get='ntk',
x_test=x_test,
compute_cov=False,
**kw_tt)
self.assertAllClose(out_mse, out_ensemble)
# Finite time NNGP train.
predict_fn_mse = predict.gradient_descent_mse(k_dd.nngp, y_train)
out_mse = predict_fn_mse(t=2.,
fx_train_0=0.,
fx_test_0=None,
k_test_train=k_td.nngp)
out_ensemble = predict_fn_ensemble(t=2.,
get='nngp',
x_test=None,
compute_cov=False,
**kw_dd)
self.assertAllClose(out_mse, out_ensemble)