拡張型

!pip install -q tf_nightly
import tensorflow as tf
import numpy as np
from typing import Tuple, List, Mapping, Union, Optional
import tempfile

2022-12-14 20:32:52.787665: E tensorflow/tsl/lib/monitoring/collection_registry.cc:81] Cannot register 2 metrics with the same name: /tensorflow/core/bfc_allocator_delay

class TensorGraph(tf.experimental.ExtensionType):
  """A collection of labeled nodes connected by weighted edges."""
  edge_weights: tf.Tensor               # shape=[num_nodes, num_nodes]
  node_labels: Mapping[str, tf.Tensor]  # shape=[num_nodes]; dtype=any

class MaskedTensor(tf.experimental.ExtensionType):
  """A tensor paired with a boolean mask, indicating which values are valid."""
  values: tf.Tensor
  mask: tf.Tensor       # shape=values.shape; false for missing/invalid values.

class CSRSparseMatrix(tf.experimental.ExtensionType):
  """Compressed sparse row matrix (https://en.wikipedia.org/wiki/Sparse_matrix)."""
  values: tf.Tensor     # shape=[num_nonzero]; dtype=any
  col_index: tf.Tensor  # shape=[num_nonzero]; dtype=int64
  row_index: tf.Tensor  # shape=[num_rows+1]; dtype=int64

class MaskedTensor(tf.experimental.ExtensionType):
  values: tf.Tensor
  mask: tf.Tensor

  def replace_mask(self, new_mask):
      self.values.shape.assert_is_compatible_with(new_mask.shape)
      return MaskedTensor(self.values, new_mask)

class MaskedTensor(tf.experimental.ExtensionType):
  values: tf.Tensor
  mask: tf.Tensor

# Constructor takes one parameter for each field.
mt = MaskedTensor(values=[[1, 2, 3], [4, 5, 6]],
                  mask=[[True, True, False], [True, False, True]])

# Fields are type-checked and converted to the declared types.
# For example, `mt.values` is converted to a Tensor.
print(mt.values)

tf.Tensor(
[[1 2 3]
 [4 5 6]], shape=(2, 3), dtype=int32)

try:
  MaskedTensor([1, 2, 3], None)
except TypeError as e:
  print(f"Got expected TypeError: {e}")

Got expected TypeError: mask: expected a Tensor, got 'NoneType'

class Pencil(tf.experimental.ExtensionType):
  color: str = "black"
  has_erasor: bool = True
  length: tf.Tensor = 1.0

Pencil()

Pencil(color='black', has_erasor=True, length=<tf.Tensor: shape=(), dtype=float32, numpy=1.0>)

Pencil(length=0.5, color="blue")

Pencil(color='blue', has_erasor=True, length=<tf.Tensor: shape=(), dtype=float32, numpy=0.5>)

print(MaskedTensor(values=[1, 2, 3], mask=[True, True, False]))

MaskedTensor(values=<tf.Tensor: shape=(3,), dtype=int32, numpy=array([1, 2, 3], dtype=int32)>, mask=<tf.Tensor: shape=(3,), dtype=bool, numpy=array([ True,  True, False])>)

a = MaskedTensor([1, 2], [True, False])
b = MaskedTensor([[3, 4], [5, 6]], [[False, True], [True, True]])
print(f"a == a: {a==a}")
print(f"a == b: {a==b}")
print(f"a == a.values: {a==a.values}")

a == a: True
a == b: False
a == a.values: False

class MaskedTensor(tf.experimental.ExtensionType):
  """A tensor paired with a boolean mask, indicating which values are valid."""
  values: tf.Tensor
  mask: tf.Tensor
  def __validate__(self):
    self.values.shape.assert_is_compatible_with(self.mask.shape)
    assert self.mask.dtype.is_bool, 'mask.dtype must be bool'

try:
  MaskedTensor([1, 2, 3], [0, 1, 0])  # Wrong `dtype` for mask.
except AssertionError as e:
  print(f"Got expected AssertionError: {e}")

Got expected AssertionError: mask.dtype must be bool

try:
  MaskedTensor([1, 2, 3], [True, False])  # shapes don't match.
except ValueError as e:
  print(f"Got expected ValueError: {e}")

Got expected ValueError: Shapes (3,) and (2,) are incompatible

mt = MaskedTensor([1, 2, 3], [True, False, True])

try:
  mt.mask = [True, True, True]
except AttributeError as e:
  print(f"Got expected AttributeError: {e}")

Got expected AttributeError: Cannot mutate attribute `mask` outside the custom constructor of ExtensionType.

try:
  mt.mask[0] = False
except TypeError as e:
  print(f"Got expected TypeError: {e}")

Got expected TypeError: 'tensorflow.python.framework.ops.EagerTensor' object does not support item assignment

try:
  del mt.mask
except AttributeError as e:
  print(f"Got expected AttributeError: {e}")

Got expected AttributeError: Cannot mutate attribute `mask` outside the custom constructor of ExtensionType.

class Player(tf.experimental.ExtensionType):
  name: tf.Tensor
  attributes: Mapping[str, tf.Tensor]

anne = Player("Anne", {"height": 8.3, "speed": 28.1})
anne_spec = tf.type_spec_from_value(anne)
print(anne_spec.name)  # Records `dtype` and `shape`, but not the string value.
print(anne_spec.attributes)  # Records keys and TensorSpecs for values.

TensorSpec(shape=(), dtype=tf.string, name=None)
ImmutableDict({'height': TensorSpec(shape=(), dtype=tf.float32, name=None), 'speed': TensorSpec(shape=(), dtype=tf.float32, name=None)})

spec1 = Player.Spec(name=tf.TensorSpec([], tf.float32), attributes={})
spec2 = tf.type_spec_from_value(anne)

@tf.function
def anonymize_player(player):
  print("<<TRACING>>")
  return Player("<anonymous>", player.attributes)

# Function gets traced (first time the function has been called):
anonymize_player(Player("Anne", {"height": 8.3, "speed": 28.1}))

<<TRACING>>
Player(name=<tf.Tensor: shape=(), dtype=string, numpy=b'<anonymous>'>, attributes=ImmutableDict({'height': <tf.Tensor: shape=(), dtype=float32, numpy=8.3>, 'speed': <tf.Tensor: shape=(), dtype=float32, numpy=28.1>}))

# Function does NOT get traced (same TypeSpec: just tensor values changed)
anonymize_player(Player("Bart", {"height": 8.1, "speed": 25.3}))

Player(name=<tf.Tensor: shape=(), dtype=string, numpy=b'<anonymous>'>, attributes=ImmutableDict({'height': <tf.Tensor: shape=(), dtype=float32, numpy=8.1>, 'speed': <tf.Tensor: shape=(), dtype=float32, numpy=25.3>}))

# Function gets traced (new TypeSpec: keys for attributes changed):
anonymize_player(Player("Chuck", {"height": 11.0, "jump": 5.3}))

<<TRACING>>
Player(name=<tf.Tensor: shape=(), dtype=string, numpy=b'<anonymous>'>, attributes=ImmutableDict({'height': <tf.Tensor: shape=(), dtype=float32, numpy=11.0>, 'jump': <tf.Tensor: shape=(), dtype=float32, numpy=5.3>}))

class MaskedTensor(tf.experimental.ExtensionType):
  """A tensor paired with a boolean mask, indicating which values are valid."""
  values: tf.Tensor
  mask: tf.Tensor       # shape=values.shape; false for invalid values.

  def __repr__(self):
    return masked_tensor_str(self.values, self.mask)

def masked_tensor_str(values, mask):
  if isinstance(values, tf.Tensor):
    if hasattr(values, 'numpy') and hasattr(mask, 'numpy'):
      return f'<MaskedTensor {masked_tensor_str(values.numpy(), mask.numpy())}>'
    else:
      return f'MaskedTensor(values={values}, mask={mask})'
  if len(values.shape) == 1:
    items = [repr(v) if m else '_' for (v, m) in zip(values, mask)]
  else:
    items = [masked_tensor_str(v, m) for (v, m) in zip(values, mask)]
  return '[%s]' % ', '.join(items)

mt = MaskedTensor(values=[[1, 2, 3], [4, 5, 6]],
                  mask=[[True, True, False], [True, False, True]])
print(mt)

<MaskedTensor [[1, 2, _], [4, _, 6]]>

class MaskedTensor(tf.experimental.ExtensionType):
  values: tf.Tensor
  mask: tf.Tensor

  def with_default(self, default):
    return tf.where(self.mask, self.values, default)

MaskedTensor([1, 2, 3], [True, False, True]).with_default(0)

<tf.Tensor: shape=(3,), dtype=int32, numpy=array([1, 0, 3], dtype=int32)>

class MaskedTensor(tf.experimental.ExtensionType):
  values: tf.Tensor
  mask: tf.Tensor

  def __repr__(self):
    return masked_tensor_str(self.values, self.mask)

  @staticmethod
  def from_tensor_and_value_to_mask(values, value_to_mask):
    return MaskedTensor(values, values != value_to_mask)

x = tf.constant([[1, 0, 2], [3, 0, 0]])
MaskedTensor.from_tensor_and_value_to_mask(x, 0)

<MaskedTensor [[1, _, 2], [3, _, _]]>

class MaskedTensor(tf.experimental.ExtensionType):
  values: tf.Tensor
  mask: tf.Tensor

  @property
  def dtype(self):
    return self.values.dtype

MaskedTensor([1, 2, 3], [True, False, True]).dtype

tf.int32

class Toy(tf.experimental.ExtensionType):
  name: str
  price: tf.Tensor
  def __init__(self, name, price, discount=0):
    self.name = name
    self.price = price * (1 - discount)

print(Toy("ball", 5.0, discount=0.2))  # On sale -- 20% off!

Toy(name='ball', price=<tf.Tensor: shape=(), dtype=float32, numpy=4.0>)

class Toy(tf.experimental.ExtensionType):
  name: str
  price: tf.Tensor

  @staticmethod
  def new_toy_with_discount(name, price, discount):
    return Toy(name, price * (1 - discount))

print(Toy.new_toy_with_discount("ball", 5.0, discount=0.2))

Toy(name='ball', price=<tf.Tensor: shape=(), dtype=float32, numpy=4.0>)

class MaskedTensor(tf.experimental.ExtensionType):
  values: tf.Tensor
  mask: tf.Tensor

  def __repr__(self):
    return masked_tensor_str(self.values, self.mask)

  def __eq__(self, other):
    result = tf.math.equal(self.values, other.values)
    result = result | ~(self.mask & other.mask)
    return tf.reduce_all(result)

x = MaskedTensor([1, 2, 3, 4], [True, True, False, True])
y = MaskedTensor([5, 2, 0, 4], [False, True, False, True])
print(x == y)

tf.Tensor(True, shape=(), dtype=bool)

class Node(tf.experimental.ExtensionType):
  value: tf.Tensor
  children: Tuple["Node", ...] = ()

Node(3, [Node(5), Node(2)])

Node(value=<tf.Tensor: shape=(), dtype=int32, numpy=3>, children=(Node(value=<tf.Tensor: shape=(), dtype=int32, numpy=5>, children=()), Node(value=<tf.Tensor: shape=(), dtype=int32, numpy=2>, children=())))

class TensorGraph(tf.experimental.ExtensionType):
  num_nodes: tf.Tensor
  edge_src: tf.Tensor   # edge_src[e] = index of src node for edge e.
  edge_dst: tf.Tensor   # edge_dst[e] = index of dst node for edge e.

class TensorGraphWithNodeFeature(TensorGraph):
  node_features: tf.Tensor  # node_features[n] = feature value for node n.

  def propagate_features(self, weight=1.0) -> 'TensorGraphWithNodeFeature':
    updates = tf.gather(self.node_features, self.edge_src) * weight
    new_node_features = tf.tensor_scatter_nd_add(
        self.node_features, tf.expand_dims(self.edge_dst, 1), updates)
    return TensorGraphWithNodeFeature(
        self.num_nodes, self.edge_src, self.edge_dst, new_node_features)

g = TensorGraphWithNodeFeature(  # Edges: 0->1, 4->3, 2->2, 2->1
    num_nodes=5, edge_src=[0, 4, 2, 2], edge_dst=[1, 3, 2, 1],
    node_features=[10.0, 0.0, 2.0, 5.0, -1.0, 0.0])

print("Original features:", g.node_features)
print("After propagating:", g.propagate_features().node_features)

Original features: tf.Tensor([10.  0.  2.  5. -1.  0.], shape=(6,), dtype=float32)
After propagating: tf.Tensor([10. 12.  4.  4. -1.  0.], shape=(6,), dtype=float32)

class MaskedTensor(tf.experimental.ExtensionType):
  values: tf.Tensor
  mask: tf.Tensor

  shape = property(lambda self: self.values.shape)
  dtype = property(lambda self: self.values.dtype)

  def __repr__(self):
    return masked_tensor_str(self.values, self.mask)

  def with_values(self, new_values):
    return MaskedTensor(new_values, self.mask)

  class Spec:
    def __init__(self, shape, dtype=tf.float32):
      self.values = tf.TensorSpec(shape, dtype)
      self.mask = tf.TensorSpec(shape, tf.bool)

    def __repr__(self):
      return f"MaskedTensor.Spec(shape={self.shape}, dtype={self.dtype})"

    shape = property(lambda self: self.values.shape)
    dtype = property(lambda self: self.values.dtype)

@tf.experimental.dispatch_for_api(tf.stack)
def masked_stack(values: List[MaskedTensor], axis = 0):
  return MaskedTensor(tf.stack([v.values for v in values], axis),
                      tf.stack([v.mask for v in values], axis))

x = MaskedTensor([1, 2, 3], [True, True, False])
y = MaskedTensor([4, 5, 6], [False, True, True])
tf.stack([x, y])

<MaskedTensor [[1, 2, _], [_, 5, 6]]>

tf.experimental.unregister_dispatch_for(masked_stack)

def convert_to_masked_tensor(x):
  if isinstance(x, MaskedTensor):
    return x
  else:
    return MaskedTensor(x, tf.ones_like(x, tf.bool))

@tf.experimental.dispatch_for_api(tf.stack)
def masked_stack_v2(values: List[Union[MaskedTensor, tf.Tensor]], axis = 0):
  values = [convert_to_masked_tensor(v) for v in values]
  return MaskedTensor(tf.stack([v.values for v in values], axis),
                      tf.stack([v.mask for v in values], axis))
x = MaskedTensor([1, 2, 3], [True, True, False])
y = tf.constant([4, 5, 6])
tf.stack([x, y, x])

<MaskedTensor [[1, 2, _], [4, 5, 6], [1, 2, _]]>

@tf.experimental.dispatch_for_unary_elementwise_apis(MaskedTensor)
 def masked_tensor_unary_elementwise_api_handler(api_func, x):
   return MaskedTensor(api_func(x.values), x.mask)

x = MaskedTensor([1, -2, -3], [True, False, True])
 print(tf.abs(x))

<MaskedTensor [1, _, 3]>

print(tf.ones_like(x, dtype=tf.float32))

<MaskedTensor [1.0, _, 1.0]>

@tf.experimental.dispatch_for_binary_elementwise_apis(MaskedTensor, MaskedTensor)
def masked_tensor_binary_elementwise_api_handler(api_func, x, y):
  return MaskedTensor(api_func(x.values, y.values), x.mask & y.mask)

x = MaskedTensor([1, -2, -3], [True, False, True])
y = MaskedTensor([[4], [5]], [[True], [False]])
tf.math.add(x, y)

<MaskedTensor [[5, _, 1], [_, _, _]]>

class Network(tf.experimental.ExtensionType):  # This version is not batchable.
  work: tf.Tensor       # work[n] = work left to do at node n
  bandwidth: tf.Tensor  # bandwidth[n1, n2] = bandwidth from n1->n2

net1 = Network([5., 3, 8], [[0., 2, 0], [2, 0, 3], [0, 3, 0]])
net2 = Network([3., 4, 2], [[0., 2, 2], [2, 0, 2], [2, 2, 0]])

class Network(tf.experimental.BatchableExtensionType):
  shape: tf.TensorShape  # batch shape. A single network has shape=[].
  work: tf.Tensor        # work[*shape, n] = work left to do at node n
  bandwidth: tf.Tensor   # bandwidth[*shape, n1, n2] = bandwidth from n1->n2

  def __init__(self, work, bandwidth):
    self.work = tf.convert_to_tensor(work)
    self.bandwidth = tf.convert_to_tensor(bandwidth)
    work_batch_shape = self.work.shape[:-1]
    bandwidth_batch_shape = self.bandwidth.shape[:-2]
    self.shape = work_batch_shape.merge_with(bandwidth_batch_shape)

  def __repr__(self):
    return network_repr(self)

def network_repr(network):
  work = network.work
  bandwidth = network.bandwidth
  if hasattr(work, 'numpy'):
    work = ' '.join(str(work.numpy()).split())
  if hasattr(bandwidth, 'numpy'):
    bandwidth = ' '.join(str(bandwidth.numpy()).split())
  return (f"<Network shape={network.shape} work={work} bandwidth={bandwidth}>")

net1 = Network([5., 3, 8], [[0., 2, 0], [2, 0, 3], [0, 3, 0]])
net2 = Network([3., 4, 2], [[0., 2, 2], [2, 0, 2], [2, 2, 0]])
batch_of_networks = Network(
    work=tf.stack([net1.work, net2.work]),
    bandwidth=tf.stack([net1.bandwidth, net2.bandwidth]))
print(f"net1={net1}")
print(f"net2={net2}")
print(f"batch={batch_of_networks}")

net1=<Network shape=() work=[5. 3. 8.] bandwidth=[[0. 2. 0.] [2. 0. 3.] [0. 3. 0.]]>
net2=<Network shape=() work=[3. 4. 2.] bandwidth=[[0. 2. 2.] [2. 0. 2.] [2. 2. 0.]]>
batch=<Network shape=(2,) work=[[5. 3. 8.] [3. 4. 2.]] bandwidth=[[[0. 2. 0.] [2. 0. 3.] [0. 3. 0.]] [[0. 2. 2.] [2. 0. 2.] [2. 2. 0.]]]>

dataset = tf.data.Dataset.from_tensor_slices(batch_of_networks)
for i, network in enumerate(dataset):
  print(f"Batch element {i}: {network}")

Batch element 0: <Network shape=() work=[5. 3. 8.] bandwidth=[[0. 2. 0.] [2. 0. 3.] [0. 3. 0.]]>
Batch element 1: <Network shape=() work=[3. 4. 2.] bandwidth=[[0. 2. 2.] [2. 0. 2.] [2. 2. 0.]]>

def balance_work_greedy(network):
  delta = (tf.expand_dims(network.work, -1) - tf.expand_dims(network.work, -2))
  delta /= 4
  delta = tf.maximum(tf.minimum(delta, network.bandwidth), -network.bandwidth)
  new_work = network.work + tf.reduce_sum(delta, -1)
  return Network(new_work, network.bandwidth)

tf.map_fn(balance_work_greedy, batch_of_networks)

<Network shape=(2,) work=[[5.5 1.25 9.25] [3. 4.75 1.25]] bandwidth=[[[0. 2. 0.] [2. 0. 3.] [0. 3. 0.]] [[0. 2. 2.] [2. 0. 2.] [2. 2. 0.]]]>

class Pastry(tf.experimental.ExtensionType):
  sweetness: tf.Tensor  # 2d embedding that encodes sweetness
  chewiness: tf.Tensor  # 2d embedding that encodes chewiness

@tf.function
def combine_pastry_features(x: Pastry):
  return (x.sweetness + x.chewiness) / 2

cookie = Pastry(sweetness=[1.2, 0.4], chewiness=[0.8, 0.2])
combine_pastry_features(cookie)

<tf.Tensor: shape=(2,), dtype=float32, numpy=array([1. , 0.3], dtype=float32)>

pastry_spec = Pastry.Spec(tf.TensorSpec([2]), tf.TensorSpec(2))

@tf.function(input_signature=[pastry_spec])
def increase_sweetness(x: Pastry, delta=1.0):
  return Pastry(x.sweetness + delta, x.chewiness)

increase_sweetness(cookie)

Pastry(sweetness=<tf.Tensor: shape=(2,), dtype=float32, numpy=array([2.2, 1.4], dtype=float32)>, chewiness=<tf.Tensor: shape=(2,), dtype=float32, numpy=array([0.8, 0.2], dtype=float32)>)

cf = combine_pastry_features.get_concrete_function(pastry_spec)
cf(cookie)

<tf.Tensor: shape=(2,), dtype=float32, numpy=array([1. , 0.3], dtype=float32)>

# Example: using tf.cond to select between two MaskedTensors. Note that the
# two MaskedTensors don't need to have the same shape.
a = MaskedTensor([1., 2, 3], [True, False, True])
b = MaskedTensor([22., 33, 108, 55], [True, True, True, False])
condition = tf.constant(True)
print(tf.cond(condition, lambda: a, lambda: b))

<MaskedTensor [1.0, _, 3.0]>

# Example: using tf.while_loop with MaskedTensor.
cond = lambda i, _: i < 10
def body(i, mt):
  return i + 1, mt.with_values(mt.values + 3 / 7)
print(tf.while_loop(cond, body, [0, b])[1])

<MaskedTensor [26.285717, 37.285698, 112.285736, _]>

@tf.function
def fn(x, b):
  if b:
    x = MaskedTensor(x, tf.less(x, 0))
  else:
    x = MaskedTensor(x, tf.greater(x, 0))
  for i in tf.range(5 if b else 7):
    x = x.with_values(x.values + 1 / 2)
  return x

print(fn(tf.constant([1., -2, 3]), tf.constant(True)))
print(fn(tf.constant([1., -2, 3]), tf.constant(False)))

<MaskedTensor [_, 0.5, _]>
<MaskedTensor [4.5, _, 6.5]>

class Network(tf.experimental.BatchableExtensionType):
  shape: tf.TensorShape  # batch shape. A single network has shape=[].
  work: tf.Tensor        # work[*shape, n] = work left to do at node n
  bandwidth: tf.Tensor   # bandwidth[*shape, n1, n2] = bandwidth from n1->n2

  def __init__(self, work, bandwidth):
    self.work = tf.convert_to_tensor(work)
    self.bandwidth = tf.convert_to_tensor(bandwidth)
    work_batch_shape = self.work.shape[:-1]
    bandwidth_batch_shape = self.bandwidth.shape[:-2]
    self.shape = work_batch_shape.merge_with(bandwidth_batch_shape)

  def __repr__(self):
    return network_repr(self)

single_network = Network(  # A single network with 4 nodes.
    work=[8.0, 5, 12, 2],
    bandwidth=[[0.0, 1, 2, 2], [1, 0, 0, 2], [2, 0, 0, 1], [2, 2, 1, 0]])

batch_of_networks = Network(  # Batch of 2 networks, each w/ 2 nodes.
    work=[[8.0, 5], [3, 2]],
    bandwidth=[[[0.0, 1], [1, 0]], [[0, 2], [2, 0]]])

class BalanceNetworkLayer(tf.keras.layers.Layer):
  """Layer that balances work between nodes in a network.

  Shifts work from more busy nodes to less busy nodes, constrained by bandwidth.
  """
  def call(self, inputs):
    # This function is defined above in the "Batchable `ExtensionType`s" section.
    return balance_work_greedy(inputs)

input_spec = Network.Spec(shape=None,
                          work=tf.TensorSpec(None, tf.float32),
                          bandwidth=tf.TensorSpec(None, tf.float32))
model = tf.keras.Sequential([
    tf.keras.layers.Input(type_spec=input_spec),
    BalanceNetworkLayer(),
    ])

model(single_network)

<Network shape=() work=[ 9.25 5. 14. -1.25] bandwidth=[[0. 1. 2. 2.] [1. 0. 0. 2.] [2. 0. 0. 1.] [2. 2. 1. 0.]]>

model(batch_of_networks)

<Network shape=(2,) work=[[8.75 4.25] [3.25 1.75]] bandwidth=[[[0. 1.] [1. 0.]] [[0. 2.] [2. 0.]]]>

class MaskedTensor(tf.experimental.BatchableExtensionType):
  # __name__ is required for serialization in SavedModel; see below for details.
  __name__ = 'extension_type_colab.MaskedTensor'

  values: tf.Tensor
  mask: tf.Tensor

  shape = property(lambda self: self.values.shape)
  dtype = property(lambda self: self.values.dtype)

  def with_default(self, default):
    return tf.where(self.mask, self.values, default)

  def __repr__(self):
    return masked_tensor_str(self.values, self.mask)

  class Spec:
    def __init__(self, shape, dtype=tf.float32):
      self.values = tf.TensorSpec(shape, dtype)
      self.mask = tf.TensorSpec(shape, tf.bool)

    shape = property(lambda self: self.values.shape)
    dtype = property(lambda self: self.values.dtype)

    def with_shape(self):
      return MaskedTensor.Spec(tf.TensorSpec(shape, self.values.dtype),
                               tf.TensorSpec(shape, self.mask.dtype))

@tf.experimental.dispatch_for_unary_elementwise_apis(MaskedTensor)
def unary_elementwise_op_handler(op, x):
 return MaskedTensor(op(x.values), x.mask)

@tf.experimental.dispatch_for_binary_elementwise_apis(
    Union[MaskedTensor, tf.Tensor],
    Union[MaskedTensor, tf.Tensor])
def binary_elementwise_op_handler(op, x, y):
  x = convert_to_masked_tensor(x)
  y = convert_to_masked_tensor(y)
  return MaskedTensor(op(x.values, y.values), x.mask & y.mask)

@tf.experimental.dispatch_for_api(tf.matmul)
def masked_matmul(a: MaskedTensor, b,
                  transpose_a=False, transpose_b=False,
                  adjoint_a=False, adjoint_b=False,
                  a_is_sparse=False, b_is_sparse=False,
                  output_type=None):
  if isinstance(a, MaskedTensor):
    a = a.with_default(0)
  if isinstance(b, MaskedTensor):
    b = b.with_default(0)
  return tf.matmul(a, b, transpose_a, transpose_b, adjoint_a,
                   adjoint_b, a_is_sparse, b_is_sparse, output_type)

input_spec = MaskedTensor.Spec([None, 2], tf.float32)

masked_tensor_model = tf.keras.Sequential([
    tf.keras.layers.Input(type_spec=input_spec),
    tf.keras.layers.Dense(16, activation="relu"),
    tf.keras.layers.Dense(1)])
masked_tensor_model.compile(loss='binary_crossentropy', optimizer='rmsprop')

a = MaskedTensor([[1., 2], [3, 4], [5, 6]],
                  [[True, False], [False, True], [True, True]])
masked_tensor_model.fit(a, tf.constant([[1], [0], [1]]), epochs=3)
print(masked_tensor_model(a))

Epoch 1/3
1/1 [==============================] - 1s 1s/step - loss: 0.7110
Epoch 2/3
1/1 [==============================] - 0s 6ms/step - loss: 0.6215
Epoch 3/3
1/1 [==============================] - 0s 5ms/step - loss: 0.5670
tf.Tensor(
[[ 0.20307031]
 [-0.32614586]
 [ 1.0157624 ]], shape=(3, 1), dtype=float32)

masked_tensor_model_path = tempfile.mkdtemp()
tf.saved_model.save(masked_tensor_model, masked_tensor_model_path)
imported_model = tf.saved_model.load(masked_tensor_model_path)
imported_model(a)

WARNING:absl:Function `_wrapped_model` contains input name(s) args_0 with unsupported characters which will be renamed to args_0_1 in the SavedModel.
INFO:tensorflow:Assets written to: /tmpfs/tmp/tmp3_ax5e_0/assets
INFO:tensorflow:Assets written to: /tmpfs/tmp/tmp3_ax5e_0/assets
<tf.Tensor: shape=(3, 1), dtype=float32, numpy=
array([[ 0.20307031],
       [-0.32614586],
       [ 1.0157624 ]], dtype=float32)>

class CustomModule(tf.Module):
  def __init__(self, variable_value):
    super().__init__()
    self.v = tf.Variable(variable_value)

  @tf.function
  def grow(self, x: MaskedTensor):
    """Increase values in `x` by multiplying them by `self.v`."""
    return MaskedTensor(x.values * self.v, x.mask)

module = CustomModule(100.0)

module.grow.get_concrete_function(MaskedTensor.Spec(shape=None,
                                                    dtype=tf.float32))
custom_module_path = tempfile.mkdtemp()
tf.saved_model.save(module, custom_module_path)
imported_model = tf.saved_model.load(custom_module_path)
imported_model.grow(MaskedTensor([1., 2, 3], [False, True, False]))

INFO:tensorflow:Assets written to: /tmpfs/tmp/tmp9apon9h2/assets
INFO:tensorflow:Assets written to: /tmpfs/tmp/tmp9apon9h2/assets
<MaskedTensor [_, 200.0, _]>

class CustomModuleWrapper(tf.Module):
  def __init__(self, variable_value):
    super().__init__()
    self.v = tf.Variable(variable_value)

  @tf.function
  def var_weighted_mean(self, x: MaskedTensor):
    """Mean value of unmasked values in x, weighted by self.v."""
    x = MaskedTensor(x.values * self.v, x.mask)
    return (tf.reduce_sum(x.with_default(0)) /
            tf.reduce_sum(tf.cast(x.mask, x.dtype)))

  @tf.function()
  def var_weighted_mean_wrapper(self, x_values, x_mask):
    """Raw tensor wrapper for var_weighted_mean."""
    return self.var_weighted_mean(MaskedTensor(x_values, x_mask))

module = CustomModuleWrapper([3., 2., 8., 5.])

module.var_weighted_mean_wrapper.get_concrete_function(
    tf.TensorSpec(None, tf.float32), tf.TensorSpec(None, tf.bool))
custom_module_path = tempfile.mkdtemp()
tf.saved_model.save(module, custom_module_path)
imported_model = tf.saved_model.load(custom_module_path)
x = MaskedTensor([1., 2., 3., 4.], [False, True, False, True])
imported_model.var_weighted_mean_wrapper(x.values, x.mask)

INFO:tensorflow:Assets written to: /tmpfs/tmp/tmpts_yhd88/assets
INFO:tensorflow:Assets written to: /tmpfs/tmp/tmpts_yhd88/assets
<tf.Tensor: shape=(), dtype=float32, numpy=12.0>

ds = tf.data.Dataset.from_tensors(Pastry(5, 5))
iter(ds).next()

Pastry(sweetness=<tf.Tensor: shape=(), dtype=int32, numpy=5>, chewiness=<tf.Tensor: shape=(), dtype=int32, numpy=5>)

mt = MaskedTensor(tf.reshape(range(20), [5, 4]), tf.ones([5, 4]))
ds = tf.data.Dataset.from_tensor_slices(mt)
for value in ds:
  print(value)

<MaskedTensor [0, 1, 2, 3]>
<MaskedTensor [4, 5, 6, 7]>
<MaskedTensor [8, 9, 10, 11]>
<MaskedTensor [12, 13, 14, 15]>
<MaskedTensor [16, 17, 18, 19]>

def value_gen():
  for i in range(2, 7):
    yield MaskedTensor(range(10), [j%i != 0 for j in range(10)])

ds = tf.data.Dataset.from_generator(
    value_gen, output_signature=MaskedTensor.Spec(shape=[10], dtype=tf.int32))
for value in ds:
  print(value)

<MaskedTensor [_, 1, _, 3, _, 5, _, 7, _, 9]>
<MaskedTensor [_, 1, 2, _, 4, 5, _, 7, 8, _]>
<MaskedTensor [_, 1, 2, 3, _, 5, 6, 7, _, 9]>
<MaskedTensor [_, 1, 2, 3, 4, _, 6, 7, 8, 9]>
<MaskedTensor [_, 1, 2, 3, 4, 5, _, 7, 8, 9]>

batched_ds = ds.batch(2)
for value in batched_ds:
  print(value)

<MaskedTensor [[_, 1, _, 3, _, 5, _, 7, _, 9], [_, 1, 2, _, 4, 5, _, 7, 8, _]]>
<MaskedTensor [[_, 1, 2, 3, _, 5, 6, 7, _, 9], [_, 1, 2, 3, 4, _, 6, 7, 8, 9]]>
<MaskedTensor [[_, 1, 2, 3, 4, 5, _, 7, 8, 9]]>

unbatched_ds = batched_ds.unbatch()
for value in unbatched_ds:
  print(value)

<MaskedTensor [_, 1, _, 3, _, 5, _, 7, _, 9]>
<MaskedTensor [_, 1, 2, _, 4, 5, _, 7, 8, _]>
<MaskedTensor [_, 1, 2, 3, _, 5, 6, 7, _, 9]>
<MaskedTensor [_, 1, 2, 3, 4, _, 6, 7, 8, 9]>
<MaskedTensor [_, 1, 2, 3, 4, 5, _, 7, 8, 9]>