def test_setter_tree(self):
witness = []
x = {"a": jnp.ones([]), "b": jnp.zeros([123])}
y = jax.tree_map(lambda x: x + 1, x)
def my_setter(next_setter, value, ctx):
self.assertIs(value, x)
self.assertEqual(ctx.original_shape, {"a": (), "b": (123, )})
self.assertEqual(ctx.original_dtype, {
"a": jnp.float32,
"b": jnp.float32
})
self.assertEqual(ctx.full_name, "~/x")
self.assertEqual(ctx.name, "x")
self.assertIsNone(ctx.module)
witness.append(None)
del next_setter
return y
with base.new_context():
with base.custom_setter(my_setter):
base.set_state("x", x)
x = base.get_state("x")
self.assertIs(x, y)
self.assertNotEmpty(witness)
def test_do_not_store(self):
def my_creator(next_creator, shape, dtype, init, context):
del next_creator, shape, dtype, init, context
return base.DO_NOT_STORE
def my_getter(next_getter, value, context):
assert value is base.DO_NOT_STORE
return next_getter(
context.original_init(context.original_shape,
context.original_dtype))
def my_setter(next_setter, value, context):
del next_setter, value, context
return base.DO_NOT_STORE
with base.new_context() as ctx:
with base.custom_creator(my_creator, state=True), \
base.custom_getter(my_getter, state=True), \
base.custom_setter(my_setter):
self.assertEqual(base.get_parameter("w", [], init=jnp.ones), 1)
self.assertEqual(base.get_state("s1", [], init=jnp.ones), 1)
base.set_state("s2", jnp.ones([]))
self.assertEmpty(ctx.collect_params())
self.assertEmpty(ctx.collect_state())
def __call__(self, value, update_stats=True, error_on_non_matrix=False):
"""Performs Spectral Normalization and returns the new value.
Args:
value: The array-like object for which you would like to perform an
spectral normalization on.
update_stats: A boolean defaulting to True. Regardless of this arg, this
function will return the normalized input. When
`update_stats` is True, the internal state of this object will also be
updated to reflect the input value. When `update_stats` is False the
internal stats will remain unchanged.
error_on_non_matrix: Spectral normalization is only defined on matrices.
By default, this module will return scalars unchanged and flatten
higher-order tensors in their leading dimensions. Setting this flag to
True will instead throw errors in those cases.
Returns:
The input value normalized by it's first singular value.
Raises:
ValueError: If `error_on_non_matrix` is True and `value` has ndims > 2.
"""
value = jnp.asarray(value)
value_shape = value.shape
# Handle scalars.
if value.ndim <= 1:
raise ValueError("Spectral normalization is not well defined for "
"scalar or vector inputs.")
# Handle higher-order tensors.
elif value.ndim > 2:
if error_on_non_matrix:
raise ValueError(
"Input is {}D but error_on_non_matrix is True".format(
value.ndim))
else:
value = jnp.reshape(value, [-1, value.shape[-1]])
u0 = base.get_state("u0",
shape=[1, value.shape[-1]],
dtype=value.dtype,
init=initializers.RandomNormal())
# Power iteration for the weight's singular value.
for _ in range(self._n_steps):
v0 = _l2_normalize(jnp.matmul(u0, value.transpose([1, 0])),
eps=self._eps)
u0 = _l2_normalize(jnp.matmul(v0, value), eps=self._eps)
u0 = jax.lax.stop_gradient(u0)
v0 = jax.lax.stop_gradient(v0)
sigma = jnp.matmul(jnp.matmul(v0, value), jnp.transpose(u0))[0, 0]
value /= sigma
value_bar = value.reshape(value_shape)
if update_stats:
base.set_state("u0", u0)
base.set_state("sigma", sigma)
return value_bar
def test_stateful(self):
with base.new_context() as ctx:
for _ in range(10):
count = base.get_state("count", (), jnp.int32, jnp.zeros)
base.set_state("count", count + 1)
self.assertEqual(ctx.collect_initial_state(), {"~": {"count": 0}})
self.assertEqual(ctx.collect_state(), {"~": {"count": 10}})
def __call__(self, x):
assert x.ndim == 0
p = base.get_parameter("p", [],
jnp.int32,
init=lambda *_: jnp.array(2))
y = x**p
base.set_state("y", y)
return y
def test_context_copies_input(self):
before = {"~": {"w": jnp.array(1.)}}
with base.new_context(params=before, state=before) as ctx:
base.get_parameter("w", [], init=jnp.ones)
base.set_state("w", jnp.array(2.))
self.assertEqual(ctx.collect_params(), {"~": {"w": jnp.array(1.)}})
self.assertIsNot(ctx.collect_initial_state(), before)
self.assertEqual(ctx.collect_initial_state(), before)
self.assertEqual(ctx.collect_state(), {"~": {"w": jnp.array(2.)}})
self.assertEqual(before, {"~": {"w": jnp.array(1.)}})
def test_difference_update_state(self):
base.get_state("a", [], init=jnp.zeros)
base.get_state("b", [], init=jnp.zeros)
before = stateful.internal_state()
base.set_state("b", jnp.ones([]))
after = stateful.internal_state()
diff = stateful.difference(before, after)
self.assertEmpty(diff.params)
self.assertEqual(diff.state, {"~": {"a": None,
"b": base.StatePair(0., 1.)}})
self.assertIsNone(diff.rng)
def test_set_then_get(self):
with base.new_context() as ctx:
base.set_state("i", 1)
base.get_state("i")
self.assertEqual(ctx.collect_initial_state(), {"~": {"i": 1}})
for _ in range(10):
with ctx:
base.set_state("i", 1)
y = base.get_state("i")
self.assertEqual(y, 1)
self.assertEqual(ctx.collect_initial_state(), {"~": {"i": 1}})
def test_without_state_raises_if_state_used_on_apply(self):
f = lambda: base.set_state("~", 1)
f = transform.without_state(transform.transform_with_state(f))
rng = jax.random.PRNGKey(42)
with self.assertRaisesRegex(ValueError, "use.*transform_with_state"):
params = f.init(rng)
f.apply(params, rng)
def test_setter_array(self):
witness = []
x = jnp.ones([])
y = x + 1
def my_setter(next_setter, value, context):
self.assertIs(value, x)
self.assertEqual(context.original_shape, value.shape)
self.assertEqual(context.original_dtype, value.dtype)
self.assertEqual(context.full_name, "~/x")
self.assertEqual(context.name, "x")
self.assertIsNone(context.module)
witness.append(None)
del next_setter
return y
with base.new_context():
with base.custom_setter(my_setter):
base.set_state("x", x)
x = base.get_state("x")
self.assertIs(x, y)
self.assertNotEmpty(witness)
def __call__(self, value, update_stats=True):
"""Updates the EMA and returns the new value.
Args:
value: The array-like object for which you would like to perform an
exponential decay on.
update_stats: A Boolean, whether to update the internal state
of this object to reflect the input value. When `update_stats` is False
the internal stats will remain unchanged.
Returns:
The exponentially weighted average of the input value.
"""
if not isinstance(value, jnp.ndarray):
value = jnp.asarray(value)
counter = base.get_state(
"counter", (),
jnp.int32,
init=initializers.Constant(-self._warmup_length))
counter += 1
decay = jax.lax.convert_element_type(self._decay, value.dtype)
if self._warmup_length > 0:
decay = self._cond(counter <= 0, 0.0, decay, value.dtype)
one = jnp.ones([], value.dtype)
hidden = base.get_state("hidden",
value.shape,
value.dtype,
init=jnp.zeros)
hidden = hidden * decay + value * (one - decay)
average = hidden
if self._zero_debias:
average /= (one - jnp.power(decay, counter))
if update_stats:
base.set_state("counter", counter)
base.set_state("hidden", hidden)
base.set_state("average", average)
return average
def __call__(self, value, update_stats=True):
"""Updates the EMA and returns the new value.
Args:
value: The array-like object for which you would like to perform an
exponential decay on.
update_stats: A Boolean, whether to update the internal state
of this object to reflect the input value. When `update_stats` is False
the internal stats will remain unchanged.
Returns:
The exponentially weighted average of the input value.
"""
value = jnp.asarray(value) # Ensure value has a dtype.
prev_counter = base.get_state(
"counter",
shape=(),
dtype=jnp.int32,
init=initializers.Constant(-self._warmup_length))
prev_hidden = base.get_state("hidden",
shape=value.shape,
dtype=value.dtype,
init=jnp.zeros)
decay = jnp.asarray(self._decay).astype(value.dtype)
counter = prev_counter + 1
decay = self._cond(jnp.less_equal(counter, 0), 0.0, decay, value.dtype)
hidden = prev_hidden * decay + value * (1 - decay)
if self._zero_debias:
average = hidden / (1. - jnp.power(decay, counter))
else:
average = hidden
if update_stats:
base.set_state("counter", counter)
base.set_state("hidden", hidden)
base.set_state("average", average)
return average
def __call__(self, x):
y = self.op(x)
base.set_state("count", self.count + 1)
return y
def net():
base.set_state("i", 1)
return base.get_state("i")
def test_new_state_in_apply(self):
with base.new_context(params={}, state={}) as ctx:
base.set_state("count", 1)
self.assertEqual(ctx.collect_initial_state(), {"~": {"count": 1}})
self.assertEqual(ctx.collect_state(), {"~": {"count": 1}})
state=True)
identity_carry = lambda f: lambda carry, x: (carry, f(x))
ignore_index = lambda f: lambda i, x: f(x)
def with_rng_example():
with base.with_rng(jax.random.PRNGKey(42)):
pass
# Methods in Haiku that mutate internal state.
SIDE_EFFECTING_FUNCTIONS = (
("get_parameter", lambda: base.get_parameter("w", [], init=jnp.zeros)),
("get_state", lambda: base.get_state("w", [], init=jnp.zeros)),
("set_state", lambda: base.set_state("w", 1)),
("next_rng_key", base.next_rng_key),
("next_rng_keys", lambda: base.next_rng_keys(2)),
("reserve_rng_keys", lambda: base.reserve_rng_keys(2)),
("with_rng", with_rng_example),
)
# JAX transforms and control flow that need to be aware of Haiku internal
# state to operate unsurprisingly.
# pylint: disable=g-long-lambda
JAX_PURE_EXPECTING_FNS = (
# Just-in-time compilation.
("jit", jax.jit),
("make_jaxpr", jax.make_jaxpr),
("eval_shape", lambda f: (lambda x: jax.eval_shape(f, x))),
def __call__(self, x):
y = x ** 2
base.set_state("count", self.count + 1)
return y
def apply():
s = base.get_state('s')
base.set_state('s', s + 1)
def __call__(self):
for _ in range(10):
count = base.get_state("count", (), jnp.int32, jnp.zeros)
base.set_state("count", count + 1)
return count
def inner():
w = base.get_state("w", [], init=jnp.zeros)
w += 1
base.set_state("w", w)
return w
def __call__(self, inputs, is_training):
"""Connects the module to some inputs.
Args:
inputs: Tensor, final dimension must be equal to embedding_dim. All other
leading dimensions will be flattened and treated as a large batch.
is_training: boolean, whether this connection is to training data. When
this is set to False, the internal moving average statistics will not be
updated.
Returns:
dict containing the following keys and values:
quantize: Tensor containing the quantized version of the input.
loss: Tensor containing the loss to optimize.
perplexity: Tensor containing the perplexity of the encodings.
encodings: Tensor containing the discrete encodings, ie which element
of the quantized space each input element was mapped to.
encoding_indices: Tensor containing the discrete encoding indices, ie
which element of the quantized space each input element was mapped to.
"""
flat_inputs = jnp.reshape(inputs, [-1, self.embedding_dim])
embeddings = self.embeddings
distances = (
jnp.sum(flat_inputs**2, 1, keepdims=True) -
2 * jnp.matmul(flat_inputs, embeddings) +
jnp.sum(embeddings**2, 0, keepdims=True))
encoding_indices = jnp.argmax(-distances, 1)
encodings = jax.nn.one_hot(encoding_indices,
self.num_embeddings,
dtype=distances.dtype)
# NB: if your code crashes with a reshape error on the line below about a
# Tensor containing the wrong number of values, then the most likely cause
# is that the input passed in does not have a final dimension equal to
# self.embedding_dim. Ideally we would catch this with an Assert but that
# creates various other problems related to device placement / TPUs.
encoding_indices = jnp.reshape(encoding_indices, inputs.shape[:-1])
quantized = self.quantize(encoding_indices)
e_latent_loss = jnp.mean((jax.lax.stop_gradient(quantized) - inputs)**2)
if is_training:
updated_ema_cluster_size = self.ema_cluster_size(
jnp.sum(encodings, axis=0))
dw = jnp.matmul(flat_inputs.T, encodings)
updated_ema_dw = self.ema_dw(dw)
n = jnp.sum(updated_ema_cluster_size)
updated_ema_cluster_size = ((updated_ema_cluster_size + self.epsilon) /
(n + self.num_embeddings * self.epsilon) * n)
normalised_updated_ema_w = (
updated_ema_dw / jnp.reshape(updated_ema_cluster_size, [1, -1]))
base.set_state('embeddings', normalised_updated_ema_w)
loss = self.commitment_cost * e_latent_loss
else:
loss = self.commitment_cost * e_latent_loss
# Straight Through Estimator
quantized = inputs + jax.lax.stop_gradient(quantized - inputs)
avg_probs = jnp.mean(encodings, 0)
perplexity = jnp.exp(-jnp.sum(avg_probs * jnp.log(avg_probs + 1e-10)))
return {
'quantize': quantized,
'loss': loss,
'perplexity': perplexity,
'encodings': encodings,
'encoding_indices': encoding_indices,
'distances': distances,
}
def init():
s = base.get_state('s', [], init=jnp.zeros)
base.set_state('s', s + 1)
