def testSingleBucket(self):
def _map_fn(v):
return (v, array_ops.fill([v], v),
array_ops.fill([3], string_ops.as_string(v)))
input_dataset = (dataset_ops.Dataset.from_tensor_slices(
math_ops.range(32)).map(_map_fn))
bucketed_dataset = input_dataset.apply(
grouping.group_by_window(
lambda x, y, z: 0,
lambda k, bucket: self._dynamicPad(k, bucket, 32), 32))
iterator = bucketed_dataset.make_initializable_iterator()
init_op = iterator.initializer
get_next = iterator.get_next()
with self.cached_session() as sess:
sess.run(init_op)
which_bucket, bucketed_values = sess.run(get_next)
self.assertEqual(0, which_bucket)
expected_scalar_int = np.arange(32, dtype=np.int64)
expected_unk_int64 = np.zeros((32, 31)).astype(np.int64)
for i in range(32):
expected_unk_int64[i, :i] = i
expected_vec3_str = np.vstack(3 * [np.arange(32).astype(bytes)]).T
self.assertAllEqual(expected_scalar_int, bucketed_values[0])
self.assertAllEqual(expected_unk_int64, bucketed_values[1])
self.assertAllEqual(expected_vec3_str, bucketed_values[2])
def testDynamicWindowSize(self):
components = np.arange(100).astype(np.int64)
# Key fn: even/odd
# Reduce fn: batches of 5
# Window size fn: even=5, odd=10
def window_size_func(key):
window_sizes = constant_op.constant([5, 10], dtype=dtypes.int64)
return window_sizes[key]
dataset = dataset_ops.Dataset.from_tensor_slices(components).apply(
grouping.group_by_window(lambda x: x % 2, lambda _, xs: xs.batch(20),
None, window_size_func))
get_next = self.getNext(dataset)
with self.assertRaises(errors.OutOfRangeError):
batches = 0
while True:
result = self.evaluate(get_next())
is_even = all(x % 2 == 0 for x in result)
is_odd = all(x % 2 == 1 for x in result)
self.assertTrue(is_even or is_odd)
expected_batch_size = 5 if is_even else 10
self.assertEqual(expected_batch_size, result.shape[0])
batches += 1
self.assertEqual(batches, 15)
def testGroupByWindowDynamicBatchWithPartialBatchWithDropRemainder(self):
# This test exercises nested batch functionality, dynamic batch size
# and drop_remainder=True together.
dataset = dataset_ops.Dataset.range(40).map(lambda x: x % 2)
def reduce_fn(key, ds):
# key == 0 -> .batch(5)
# key == 1 -> .batch(10)
return ds.batch(batch_size=(key + 1) * 5, drop_remainder=True)
dataset = dataset.apply(
grouping.group_by_window(key_func=lambda x: x,
reduce_func=reduce_fn,
window_size=11))
dataset = distribute._RebatchDataset(dataset, num_workers=2)
self.assertEqual([[None]],
[ts.as_list() for ts in _flat_shapes(dataset)])
# The batches of 5 (value == 0) will be split into minibatches of (3, 2) and
# the batches of 10 (value == 1) split into minibatches of (5, 5)
# [(batch_size, value), ...]
pairs = [(3, 0), (2, 0), (3, 0), (2, 0), (5, 1), (5, 1), (3, 0),
(2, 0)]
expected_output = [[value] * batch_size for batch_size, value in pairs]
self.assertDatasetProduces(dataset, expected_output)
def testGroupByWindowDynamicBatch(self):
# {0, 1, 0, 1, ...}
dataset = dataset_ops.Dataset.range(40).map(lambda x: x % 2)
def reduce_fn(key, ds):
# key == 0 -> .batch(5)
# key == 1 -> .batch(10)
return ds.batch(batch_size=(key + 1) * 5)
dataset = dataset.apply(
grouping.group_by_window(key_func=lambda x: x,
reduce_func=reduce_fn,
window_size=10))
dataset = distribute._RebatchDataset(dataset, num_workers=2)
self.assertEqual([[None]],
[ts.as_list() for ts in _flat_shapes(dataset)])
# The batches of 5 (value == 0) will be split into minibatches of (3, 2) and
# the batches of 10 (value == 1) split into minibatches of (5, 5)
# [(batch_size, value), ...]
pairs = [(3, 0), (2, 0), (3, 0), (2, 0), (5, 1), (5, 1)]
pairs = pairs * 2
expected_output = [[value] * batch_size for batch_size, value in pairs]
self.assertDatasetProduces(dataset, expected_output)
def testSingleBucket(self):
def _map_fn(v):
return (v, array_ops.fill([v], v),
array_ops.fill([3], string_ops.as_string(v)))
input_dataset = dataset_ops.Dataset.from_tensor_slices(
math_ops.range(32)).map(_map_fn)
bucketed_dataset = input_dataset.apply(
grouping.group_by_window(
lambda x, y, z: 0,
lambda k, bucket: self._dynamicPad(k, bucket, 32), 32))
get_next = self.getNext(bucketed_dataset)
which_bucket, bucketed_values = self.evaluate(get_next())
self.assertEqual(0, which_bucket)
expected_scalar_int = np.arange(32, dtype=np.int64)
expected_unk_int64 = np.zeros((32, 31)).astype(np.int64)
for i in range(32):
expected_unk_int64[i, :i] = i
expected_vec3_str = np.vstack(3 * [np.arange(32).astype(bytes)]).T
self.assertAllEqual(expected_scalar_int, bucketed_values[0])
self.assertAllEqual(expected_unk_int64, bucketed_values[1])
self.assertAllEqual(expected_vec3_str, bucketed_values[2])
def testSingleBucket(self):
def _map_fn(v):
return (v, array_ops.fill([v], v),
array_ops.fill([3], string_ops.as_string(v)))
input_dataset = (
dataset_ops.Dataset.from_tensor_slices(math_ops.range(32)).map(_map_fn))
bucketed_dataset = input_dataset.apply(
grouping.group_by_window(
lambda x, y, z: 0,
lambda k, bucket: self._dynamicPad(k, bucket, 32), 32))
iterator = bucketed_dataset.make_initializable_iterator()
init_op = iterator.initializer
get_next = iterator.get_next()
with self.cached_session() as sess:
sess.run(init_op)
which_bucket, bucketed_values = sess.run(get_next)
self.assertEqual(0, which_bucket)
expected_scalar_int = np.arange(32, dtype=np.int64)
expected_unk_int64 = np.zeros((32, 31)).astype(np.int64)
for i in range(32):
expected_unk_int64[i, :i] = i
expected_vec3_str = np.vstack(3 * [np.arange(32).astype(bytes)]).T
self.assertAllEqual(expected_scalar_int, bucketed_values[0])
self.assertAllEqual(expected_unk_int64, bucketed_values[1])
self.assertAllEqual(expected_vec3_str, bucketed_values[2])
def testSimple(self):
components = np.random.randint(100, size=(200,)).astype(np.int64)
iterator = (
dataset_ops.Dataset.from_tensor_slices(components).map(lambda x: x * x)
.apply(
grouping.group_by_window(lambda x: x % 2, lambda _, xs: xs.batch(4),
4)).make_initializable_iterator())
init_op = iterator.initializer
get_next = iterator.get_next()
with self.cached_session() as sess:
sess.run(init_op)
counts = []
with self.assertRaises(errors.OutOfRangeError):
while True:
result = sess.run(get_next)
self.assertTrue(
all(x % 2 == 0
for x in result) or all(x % 2 == 1)
for x in result)
counts.append(result.shape[0])
self.assertEqual(len(components), sum(counts))
num_full_batches = len([c for c in counts if c == 4])
self.assertGreaterEqual(num_full_batches, 24)
self.assertTrue(all(c == 4 for c in counts[:num_full_batches]))
def testTwoLevelDistribute(self):
cluster_1_size = 3
dispatcher_1, workers_1 = self.start_cluster( # to avoid gcing workers, pylint: disable=unused-variable
cluster_1_size,
name="cluster_1")
dispatcher_2, workers_2 = self.start_cluster(1, name="cluster_2") # to avoid gcing workers, pylint: disable=unused-variable
num_sizes = 10
size_repeats = 5
strings = ["a" * i for i in range(num_sizes)] * size_repeats
ds = dataset_ops.Dataset.from_tensor_slices(strings)
ds = ds.shuffle(len(strings))
ds = _make_distributed_dataset(ds, dispatcher_1)
# Large enough so that all strings of the same size are windowed together.
window_size = cluster_1_size * size_repeats
batch_size = size_repeats
def key_func(x):
return math_ops.cast(string_ops.string_length_v2(x), dtypes.int64)
ds = ds.apply(
grouping.group_by_window(
key_func=key_func,
reduce_func=lambda _, x: x.batch(batch_size),
window_size=window_size))
ds = _make_distributed_dataset(ds, dispatcher_2)
it = iter(ds)
for _ in range(num_sizes):
element = next(it).numpy()
for _ in range(1, cluster_1_size):
self.assertAllEqual(next(it).numpy(), element)
self.assertEmpty(list(it))
def testGroupByWindowCardinality(self):
dataset = dataset_ops.Dataset.range(1).repeat().apply(
grouping.group_by_window(
lambda x: x % 2,
lambda key, window: dataset_ops.Dataset.from_tensors(key), 4))
self.assertEqual(self.evaluate(dataset.cardinality()),
dataset_ops.INFINITE)
def testConsumeWindowDatasetMoreThanOnce(self):
components = np.random.randint(50, size=(200,)).astype(np.int64)
def reduce_func(key, window):
# Apply two different kinds of padding to the input: tight
# padding, and quantized (to a multiple of 10) padding.
return dataset_ops.Dataset.zip((
window.padded_batch(
4, padded_shapes=tensor_shape.TensorShape([None])),
window.padded_batch(
4, padded_shapes=ops.convert_to_tensor([(key + 1) * 10])),
))
dataset = dataset_ops.Dataset.from_tensor_slices(
components
).map(lambda x: array_ops.fill([math_ops.cast(x, dtypes.int32)], x)).apply(
grouping.group_by_window(
lambda x: math_ops.cast(array_ops.shape(x)[0] // 10, dtypes.int64),
reduce_func, 4))
get_next = self.getNext(dataset)
counts = []
with self.assertRaises(errors.OutOfRangeError):
while True:
tight_result, multiple_of_10_result = self.evaluate(get_next())
self.assertEqual(0, multiple_of_10_result.shape[1] % 10)
self.assertAllEqual(tight_result,
multiple_of_10_result[:, :tight_result.shape[1]])
counts.append(tight_result.shape[0])
self.assertEqual(len(components), sum(counts))
def testTwoLevelDistribute(self):
cluster_1_size = 3
cluster_1 = data_service_test_base.TestCluster(
num_workers=cluster_1_size)
cluster_2 = data_service_test_base.TestCluster(num_workers=1)
num_sizes = 10
size_repeats = 5
strings = ["a" * i for i in range(num_sizes)] * size_repeats
ds = dataset_ops.Dataset.from_tensor_slices(strings)
ds = ds.shuffle(len(strings))
ds = self.make_distributed_dataset(ds, cluster_1)
# Large enough so that all strings of the same size are windowed together.
window_size = cluster_1_size * size_repeats
batch_size = size_repeats
def key_func(x):
return math_ops.cast(string_ops.string_length_v2(x), dtypes.int64)
ds = ds.apply(
grouping.group_by_window(
key_func=key_func,
reduce_func=lambda _, x: x.batch(batch_size),
window_size=window_size))
ds = self.make_distributed_dataset(ds, cluster_2)
get_next = self.getNext(ds)
for _ in range(num_sizes):
element = self.evaluate(get_next())
for _ in range(1, cluster_1_size):
self.assertAllEqual(self.evaluate(get_next()), element)
self.assertEmpty(self.getIteratorOutput(get_next))
def testSimple(self):
components = np.random.randint(100, size=(200, )).astype(np.int64)
iterator = (dataset_ops.Dataset.from_tensor_slices(components).map(
lambda x: x * x).apply(
grouping.group_by_window(lambda x: x % 2,
lambda _, xs: xs.batch(4),
4)).make_initializable_iterator())
init_op = iterator.initializer
get_next = iterator.get_next()
with self.cached_session() as sess:
sess.run(init_op)
counts = []
with self.assertRaises(errors.OutOfRangeError):
while True:
result = sess.run(get_next)
self.assertTrue(
all(x % 2 == 0 for x in result) or all(x % 2 == 1)
for x in result)
counts.append(result.shape[0])
self.assertEqual(len(components), sum(counts))
num_full_batches = len([c for c in counts if c == 4])
self.assertGreaterEqual(num_full_batches, 24)
self.assertTrue(all(c == 4 for c in counts[:num_full_batches]))
def make_group_by_window_dataset(var):
def reduce_fn(key, bucket):
del key, bucket
return dataset_ops.Dataset.from_tensors(var)
return dataset_ops.Dataset.from_tensors(0).repeat(10).apply(
grouping.group_by_window(lambda _: 0, reduce_fn, 10))
def testRoundRobinBucketizing(self):
# Tests a common use case for round robin reads. At each step, all
# consumers should get batches with the same bucket size.
cluster = self.create_cluster(num_workers=4)
num_elements = 100
ds = dataset_ops.Dataset.range(num_elements, output_type=dtypes.int32)
ds = ds.shuffle(num_elements)
low_bucket_max = 30
mid_bucket_max = 60
bucket_boundaries = [low_bucket_max, mid_bucket_max]
batch_size = 10
num_consumers = 3
bucket_batch_sizes = [batch_size] * (len(bucket_boundaries) + 1)
ds = ds.apply(
grouping.bucket_by_sequence_length(lambda x: x,
bucket_boundaries,
bucket_batch_sizes,
drop_remainder=True))
ds = ds.apply(
grouping.group_by_window(
lambda x: math_ops.cast(x[1], dtypes.int64),
lambda _, x: dataset_ops.Dataset.from_tensors(x),
window_size=num_consumers))
ds = ds.flat_map(lambda x: x)
ds = ds.repeat()
consumers = []
for consumer_index in range(num_consumers):
consumers.append(
self.make_distributed_dataset(ds,
cluster,
job_name="test",
consumer_index=consumer_index,
num_consumers=num_consumers))
# Use parallel interleave to read from consumers in parallel.
ds = dataset_ops.Dataset.from_tensor_slices(consumers)
ds = ds.interleave(lambda x: x.prefetch(num_elements),
cycle_length=num_consumers,
num_parallel_calls=num_consumers)
num_rounds = 10
get_next = self.getNext(ds, requires_initialization=True)
results = []
for _ in range(num_rounds):
results.append(self.evaluate(get_next()))
def get_bucket(elem):
bucket_ind = 0
while bucket_ind < len(bucket_boundaries
) and elem >= bucket_boundaries[bucket_ind]:
bucket_ind += 1
return bucket_ind
for i in range(0, len(results), num_consumers):
batches = results[num_consumers * i:num_consumers * i +
num_consumers]
bucket_inds = [get_bucket(batch[0]) for batch in batches]
for bucket_ind in bucket_inds[1:]:
self.assertEqual(bucket_inds[0], bucket_ind)
def testGroupByWindowWithAutotune(self):
dataset = dataset_ops.Dataset.range(1000).apply(
grouping.group_by_window(
lambda x: x // 10,
lambda key, window: dataset_ops.Dataset.from_tensors(key), 4))
dataset = dataset.map(lambda x: x + 1, num_parallel_calls=-1)
get_next = self.getNext(dataset)
self.evaluate(get_next())
def testShortCircuit(self):
dataset = dataset_ops.Dataset.range(10)
dataset = dataset.apply(
grouping.group_by_window(lambda x: x, lambda _, window: window.batch(1),
1))
self.assertDatasetProduces(
dataset, expected_output=[[i] for i in range(10)])
def make_group_by_window_dataset(var):
def reduce_fn(key, bucket):
del key, bucket
return dataset_ops.Dataset.from_tensors(var)
return dataset_ops.Dataset.from_tensors(0).repeat(10).apply(
grouping.group_by_window(lambda _: 0, reduce_fn, 10))
def testEmpty(self):
dataset = dataset_ops.Dataset.range(4).apply(
grouping.group_by_window(lambda _: 0, lambda _, xs: xs, 0))
get_next = self.getNext(dataset)
with self.assertRaisesRegexp(
errors.InvalidArgumentError,
"Window size must be greater than zero, but got 0."):
print(self.evaluate(get_next()))
def testEvenOddBuckets(self):
def _map_fn(v):
return (v, array_ops.fill([v], v),
array_ops.fill([3], string_ops.as_string(v)))
input_dataset = (
dataset_ops.Dataset.from_tensor_slices(math_ops.range(64)).map(_map_fn))
bucketed_dataset = input_dataset.apply(
grouping.group_by_window(
lambda x, y, z: math_ops.cast(x % 2, dtypes.int64),
lambda k, bucket: self._dynamicPad(k, bucket, 32), 32))
iterator = bucketed_dataset.make_initializable_iterator()
init_op = iterator.initializer
get_next = iterator.get_next()
with self.cached_session() as sess:
sess.run(init_op)
# Get two minibatches (one containing even values, one containing odds)
which_bucket_even, bucketed_values_even = sess.run(get_next)
which_bucket_odd, bucketed_values_odd = sess.run(get_next)
# Count number of bucket_tensors.
self.assertEqual(3, len(bucketed_values_even))
self.assertEqual(3, len(bucketed_values_odd))
# Ensure bucket 0 was used for all minibatch entries.
self.assertAllEqual(0, which_bucket_even)
self.assertAllEqual(1, which_bucket_odd)
# Test the first bucket outputted, the events starting at 0
expected_scalar_int = np.arange(0, 32 * 2, 2, dtype=np.int64)
expected_unk_int64 = np.zeros((32, 31 * 2)).astype(np.int64)
for i in range(0, 32):
expected_unk_int64[i, :2 * i] = 2 * i
expected_vec3_str = np.vstack(
3 * [np.arange(0, 32 * 2, 2).astype(bytes)]).T
self.assertAllEqual(expected_scalar_int, bucketed_values_even[0])
self.assertAllEqual(expected_unk_int64, bucketed_values_even[1])
self.assertAllEqual(expected_vec3_str, bucketed_values_even[2])
# Test the second bucket outputted, the odds starting at 1
expected_scalar_int = np.arange(1, 32 * 2 + 1, 2, dtype=np.int64)
expected_unk_int64 = np.zeros((32, 31 * 2 + 1)).astype(np.int64)
for i in range(0, 32):
expected_unk_int64[i, :2 * i + 1] = 2 * i + 1
expected_vec3_str = np.vstack(
3 * [np.arange(1, 32 * 2 + 1, 2).astype(bytes)]).T
self.assertAllEqual(expected_scalar_int, bucketed_values_odd[0])
self.assertAllEqual(expected_unk_int64, bucketed_values_odd[1])
self.assertAllEqual(expected_vec3_str, bucketed_values_odd[2])
def testEvenOddBuckets(self):
def _map_fn(v):
return (v, array_ops.fill([v], v),
array_ops.fill([3], string_ops.as_string(v)))
input_dataset = (dataset_ops.Dataset.from_tensor_slices(
math_ops.range(64)).map(_map_fn))
bucketed_dataset = input_dataset.apply(
grouping.group_by_window(
lambda x, y, z: math_ops.cast(x % 2, dtypes.int64),
lambda k, bucket: self._dynamicPad(k, bucket, 32), 32))
iterator = bucketed_dataset.make_initializable_iterator()
init_op = iterator.initializer
get_next = iterator.get_next()
with self.cached_session() as sess:
sess.run(init_op)
# Get two minibatches (one containing even values, one containing odds)
which_bucket_even, bucketed_values_even = sess.run(get_next)
which_bucket_odd, bucketed_values_odd = sess.run(get_next)
# Count number of bucket_tensors.
self.assertEqual(3, len(bucketed_values_even))
self.assertEqual(3, len(bucketed_values_odd))
# Ensure bucket 0 was used for all minibatch entries.
self.assertAllEqual(0, which_bucket_even)
self.assertAllEqual(1, which_bucket_odd)
# Test the first bucket outputted, the events starting at 0
expected_scalar_int = np.arange(0, 32 * 2, 2, dtype=np.int64)
expected_unk_int64 = np.zeros((32, 31 * 2)).astype(np.int64)
for i in range(0, 32):
expected_unk_int64[i, :2 * i] = 2 * i
expected_vec3_str = np.vstack(
3 * [np.arange(0, 32 * 2, 2).astype(bytes)]).T
self.assertAllEqual(expected_scalar_int, bucketed_values_even[0])
self.assertAllEqual(expected_unk_int64, bucketed_values_even[1])
self.assertAllEqual(expected_vec3_str, bucketed_values_even[2])
# Test the second bucket outputted, the odds starting at 1
expected_scalar_int = np.arange(1, 32 * 2 + 1, 2, dtype=np.int64)
expected_unk_int64 = np.zeros((32, 31 * 2 + 1)).astype(np.int64)
for i in range(0, 32):
expected_unk_int64[i, :2 * i + 1] = 2 * i + 1
expected_vec3_str = np.vstack(
3 * [np.arange(1, 32 * 2 + 1, 2).astype(bytes)]).T
self.assertAllEqual(expected_scalar_int, bucketed_values_odd[0])
self.assertAllEqual(expected_unk_int64, bucketed_values_odd[1])
self.assertAllEqual(expected_vec3_str, bucketed_values_odd[2])
def testSmallGroups(self):
components = np.array([0, 0, 0, 0, 1, 1, 1, 1, 1, 0, 0, 0], dtype=np.int64)
dataset = dataset_ops.Dataset.from_tensor_slices(components).apply(
grouping.group_by_window(lambda x: x % 2, lambda _, xs: xs.batch(4), 4))
get_next = self.getNext(dataset)
self.assertAllEqual([0, 0, 0, 0], self.evaluate(get_next()))
self.assertAllEqual([1, 1, 1, 1], self.evaluate(get_next()))
# The small outputs at the end are deterministically produced in key
# order.
self.assertAllEqual([0, 0, 0], self.evaluate(get_next()))
self.assertAllEqual([1], self.evaluate(get_next()))
def testEmpty(self):
iterator = dataset_ops.make_initializable_iterator(
dataset_ops.Dataset.range(4).apply(
grouping.group_by_window(lambda _: 0, lambda _, xs: xs, 0)))
init_op = iterator.initializer
get_next = iterator.get_next()
with self.cached_session() as sess:
self.evaluate(init_op)
with self.assertRaisesRegexp(
errors.InvalidArgumentError,
"Window size must be greater than zero, but got 0."):
print(self.evaluate(get_next))
def testEmpty(self):
iterator = (dataset_ops.Dataset.range(4).apply(
grouping.group_by_window(lambda _: 0, lambda _, xs: xs,
0)).make_initializable_iterator())
init_op = iterator.initializer
get_next = iterator.get_next()
with self.cached_session() as sess:
sess.run(init_op)
with self.assertRaisesRegexp(
errors.InvalidArgumentError,
"Window size must be greater than zero, but got 0."):
print(sess.run(get_next))
def testSmallGroups(self):
components = np.array([0, 0, 0, 0, 1, 1, 1, 1, 1, 0, 0, 0], dtype=np.int64)
iterator = (
dataset_ops.Dataset.from_tensor_slices(components).apply(
grouping.group_by_window(lambda x: x % 2, lambda _, xs: xs.batch(4),
4)).make_initializable_iterator())
init_op = iterator.initializer
get_next = iterator.get_next()
with self.cached_session() as sess:
sess.run(init_op)
self.assertAllEqual([0, 0, 0, 0], sess.run(get_next))
self.assertAllEqual([1, 1, 1, 1], sess.run(get_next))
# The small outputs at the end are deterministically produced in key
# order.
self.assertAllEqual([0, 0, 0], sess.run(get_next))
self.assertAllEqual([1], sess.run(get_next))
def testEvenOddBucketsFilterOutAllOdd(self):
def _map_fn(v):
return {
"x": v,
"y": array_ops.fill([v], v),
"z": array_ops.fill([3], string_ops.as_string(v))
}
def _dynamic_pad_fn(bucket, window, _):
return dataset_ops.Dataset.zip(
(dataset_ops.Dataset.from_tensors(bucket),
window.padded_batch(
32, {
"x": tensor_shape.TensorShape([]),
"y": tensor_shape.TensorShape([None]),
"z": tensor_shape.TensorShape([3])
})))
input_dataset = (
dataset_ops.Dataset.from_tensor_slices(math_ops.range(128)).map(_map_fn)
.filter(lambda d: math_ops.equal(d["x"] % 2, 0)))
bucketed_dataset = input_dataset.apply(
grouping.group_by_window(
lambda d: math_ops.cast(d["x"] % 2, dtypes.int64),
lambda k, bucket: _dynamic_pad_fn(k, bucket, 32), 32))
iterator = bucketed_dataset.make_initializable_iterator()
init_op = iterator.initializer
get_next = iterator.get_next()
with self.cached_session() as sess:
sess.run(init_op)
# Get two minibatches ([0, 2, ...] and [64, 66, ...])
which_bucket0, bucketed_values_even0 = sess.run(get_next)
which_bucket1, bucketed_values_even1 = sess.run(get_next)
# Ensure that bucket 1 was completely filtered out
self.assertAllEqual(0, which_bucket0)
self.assertAllEqual(0, which_bucket1)
self.assertAllEqual(
np.arange(0, 64, 2, dtype=np.int64), bucketed_values_even0["x"])
self.assertAllEqual(
np.arange(64, 128, 2, dtype=np.int64), bucketed_values_even1["x"])
def testImmediateOutput(self):
components = np.array(
[0, 0, 0, 0, 1, 1, 1, 1, 2, 2, 0, 0, 2, 2, 0, 0], dtype=np.int64)
dataset = dataset_ops.Dataset.from_tensor_slices(components).repeat(
-1).apply(
grouping.group_by_window(lambda x: x % 3, lambda _, xs: xs.batch(4),
4))
get_next = self.getNext(dataset)
# The input is infinite, so this test demonstrates that:
# 1. We produce output without having to consume the entire input,
# 2. Different buckets can produce output at different rates, and
# 3. For deterministic input, the output is deterministic.
for _ in range(3):
self.assertAllEqual([0, 0, 0, 0], self.evaluate(get_next()))
self.assertAllEqual([1, 1, 1, 1], self.evaluate(get_next()))
self.assertAllEqual([2, 2, 2, 2], self.evaluate(get_next()))
self.assertAllEqual([0, 0, 0, 0], self.evaluate(get_next()))
def testReduceFuncError(self):
components = np.random.randint(100, size=(200,)).astype(np.int64)
def reduce_func(_, xs):
# Introduce an incorrect padded shape that cannot (currently) be
# detected at graph construction time.
return xs.padded_batch(
4,
padded_shapes=(tensor_shape.TensorShape([]),
constant_op.constant([5], dtype=dtypes.int64) * -1))
dataset = dataset_ops.Dataset.from_tensor_slices(
components).map(lambda x: (x, ops.convert_to_tensor([x * x]))).apply(
grouping.group_by_window(lambda x, _: x % 2, reduce_func, 32))
get_next = self.getNext(dataset)
with self.assertRaises(errors.InvalidArgumentError):
self.evaluate(get_next())
def testEvenOddBucketsFilterOutAllOdd(self):
def _map_fn(v):
return {
"x": v,
"y": array_ops.fill([v], v),
"z": array_ops.fill([3], string_ops.as_string(v))
}
def _dynamic_pad_fn(bucket, window, _):
return dataset_ops.Dataset.zip(
(dataset_ops.Dataset.from_tensors(bucket),
window.padded_batch(
32, {
"x": tensor_shape.TensorShape([]),
"y": tensor_shape.TensorShape([None]),
"z": tensor_shape.TensorShape([3])
})))
input_dataset = (dataset_ops.Dataset.from_tensor_slices(
math_ops.range(128)).map(_map_fn).filter(
lambda d: math_ops.equal(d["x"] % 2, 0)))
bucketed_dataset = input_dataset.apply(
grouping.group_by_window(
lambda d: math_ops.cast(d["x"] % 2, dtypes.int64),
lambda k, bucket: _dynamic_pad_fn(k, bucket, 32), 32))
iterator = bucketed_dataset.make_initializable_iterator()
init_op = iterator.initializer
get_next = iterator.get_next()
with self.cached_session() as sess:
sess.run(init_op)
# Get two minibatches ([0, 2, ...] and [64, 66, ...])
which_bucket0, bucketed_values_even0 = sess.run(get_next)
which_bucket1, bucketed_values_even1 = sess.run(get_next)
# Ensure that bucket 1 was completely filtered out
self.assertAllEqual(0, which_bucket0)
self.assertAllEqual(0, which_bucket1)
self.assertAllEqual(np.arange(0, 64, 2, dtype=np.int64),
bucketed_values_even0["x"])
self.assertAllEqual(np.arange(64, 128, 2, dtype=np.int64),
bucketed_values_even1["x"])
def testGroupByWindowBatching(self, drop_remainder):
dataset = dataset_ops.Dataset.from_tensor_slices(
[[array_ops.constant(i, dtype=dtypes.int64)] * 3 for i in range(40)])
reduce_fn = lambda bucket_id, ds: ds.batch(
batch_size=10, drop_remainder=drop_remainder)
dataset = dataset.apply(
grouping.group_by_window(
key_func=lambda x: x[0] % 4, reduce_func=reduce_fn, window_size=10))
rebatched_dataset = distribute._RebatchDataset(dataset, num_workers=2)
self.assertEqual([[5, 3] if drop_remainder else [None, 3]],
[ts.as_list() for ts in _flat_shapes(rebatched_dataset)])
# pylint: disable=g-complex-comprehension
expected_output = [[[j + i * 4 + k * 20] * 3
for i in range(5)]
for j in range(4)
for k in range(2)]
self.assertDatasetProduces(rebatched_dataset, expected_output)
def testGroupByWindowDynamicBatch(self, drop_remainder):
dataset = dataset_ops.Dataset.range(40).map(lambda x: x % 2)
reduce_fn = lambda bucket_id, ds: ds.batch( # pylint: disable=g-long-lambda
batch_size=(bucket_id + 1) * 5,
drop_remainder=drop_remainder)
dataset = dataset.apply(
grouping.group_by_window(key_func=lambda x: x,
reduce_func=reduce_fn,
window_size=10))
dataset = distribute._RebatchDataset(dataset, num_workers=2)
self.assertEqual([[None]],
[ts.as_list() for ts in _flat_shapes(dataset)])
# pylint: disable=g-complex-comprehension
x = [(2, 0), (2, 0), (2, 0), (2, 0), (2, 0), (5, 1), (5, 1), (2, 0),
(2, 0), (2, 0), (2, 0), (2, 0), (5, 1), (5, 1)]
expected_output = [[value] * batch_size for batch_size, value in x]
self.assertDatasetProduces(dataset, expected_output)
def testSmallGroups(self):
components = np.array([0, 0, 0, 0, 1, 1, 1, 1, 1, 0, 0, 0],
dtype=np.int64)
iterator = (dataset_ops.Dataset.from_tensor_slices(components).apply(
grouping.group_by_window(lambda x: x % 2,
lambda _, xs: xs.batch(4),
4)).make_initializable_iterator())
init_op = iterator.initializer
get_next = iterator.get_next()
with self.cached_session() as sess:
sess.run(init_op)
self.assertAllEqual([0, 0, 0, 0], sess.run(get_next))
self.assertAllEqual([1, 1, 1, 1], sess.run(get_next))
# The small outputs at the end are deterministically produced in key
# order.
self.assertAllEqual([0, 0, 0], sess.run(get_next))
self.assertAllEqual([1], sess.run(get_next))
def testGroupByWindowStaticBatch(self):
dataset = dataset_ops.Dataset.from_tensor_slices(
[[array_ops.constant(i, dtype=dtypes.int64)] * 3 for i in range(40)])
reduce_fn = lambda bucket_id, ds: ds.batch( # pylint: disable=g-long-lambda
batch_size=10)
dataset = dataset.apply(
grouping.group_by_window(
key_func=lambda x: x[0] % 4, reduce_func=reduce_fn, window_size=10))
rebatched_dataset = distribute._RebatchDataset(dataset, num_replicas=2)
self.assertEqual([[None, 3]],
[ts.as_list() for ts in _flat_shapes(rebatched_dataset)])
# pylint: disable=g-complex-comprehension
expected_output = [[[j + i * 4 + k * 20] * 3
for i in range(5)]
for j in range(4)
for k in range(2)]
self.assertDatasetProduces(rebatched_dataset, expected_output)
def testStatefulGroupByWindowNotCheckpointable(self):
stateful_key_func = self._statefulInt64Func
key_func = lambda _: math_ops.cast(0, dtypes.int64)
stateful_reduce_func = lambda _, x: self._statefulDatasetFunc(x)
reduce_func = lambda _, x: x
stateful_window_func = self._statefulInt64Func
window_func = lambda x: math_ops.cast(0, dtypes.int64)
test_cases = [
(stateful_key_func, reduce_func, window_func),
(key_func, stateful_reduce_func, window_func),
(key_func, reduce_func, stateful_window_func),
]
for key_func_fn, reduce_func_fn, window_func in test_cases:
dataset = dataset_ops.Dataset.range(10)
dataset = dataset.apply(
grouping.group_by_window(
key_func_fn, reduce_func_fn, window_size_func=window_func))
self._assertNotCheckpointable(dataset)
def testSimple(self):
components = np.random.randint(100, size=(200,)).astype(np.int64)
dataset = dataset_ops.Dataset.from_tensor_slices(
components).map(lambda x: x * x).apply(
grouping.group_by_window(lambda x: x % 2, lambda _, xs: xs.batch(4),
4))
get_next = self.getNext(dataset)
counts = []
with self.assertRaises(errors.OutOfRangeError):
while True:
result = self.evaluate(get_next())
self.assertTrue(
all(x % 2 == 0 for x in result) or all(x % 2 == 1) for x in result)
counts.append(result.shape[0])
self.assertEqual(len(components), sum(counts))
num_full_batches = len([c for c in counts if c == 4])
self.assertGreaterEqual(num_full_batches, 24)
self.assertTrue(all(c == 4 for c in counts[:num_full_batches]))
def testGroupByWindowDynamicBatchWithPartialBatch(self):
# {0, 1, 0, 1, ...}
dataset = dataset_ops.Dataset.range(40).map(lambda x: x % 2)
def reduce_fn(key, ds):
# key == 0 -> .batch(5)
# key == 1 -> .batch(10)
return ds.batch(batch_size=(key + 1) * 5)
dataset = dataset.apply(
grouping.group_by_window(
key_func=lambda x: x, reduce_func=reduce_fn, window_size=11))
dataset = distribute._RebatchDataset(dataset, num_replicas=2)
self.assertEqual([[None]], [ts.as_list() for ts in _flat_shapes(dataset)])
pairs = [(3, 0), (2, 0), (3, 0), (2, 0), (1, 0), (0, 0), (5, 1), (5, 1),
(1, 1), (0, 1), (3, 0), (2, 0), (2, 0), (2, 0), (5, 1), (4, 1)]
expected_output = [[value] * batch_size for batch_size, value in pairs]
self.assertDatasetProduces(dataset, expected_output)
def testImmediateOutput(self):
components = np.array(
[0, 0, 0, 0, 1, 1, 1, 1, 2, 2, 0, 0, 2, 2, 0, 0], dtype=np.int64)
iterator = (
dataset_ops.Dataset.from_tensor_slices(components).repeat(-1).apply(
grouping.group_by_window(lambda x: x % 3, lambda _, xs: xs.batch(4),
4)).make_initializable_iterator())
init_op = iterator.initializer
get_next = iterator.get_next()
with self.cached_session() as sess:
sess.run(init_op)
# The input is infinite, so this test demonstrates that:
# 1. We produce output without having to consume the entire input,
# 2. Different buckets can produce output at different rates, and
# 3. For deterministic input, the output is deterministic.
for _ in range(3):
self.assertAllEqual([0, 0, 0, 0], sess.run(get_next))
self.assertAllEqual([1, 1, 1, 1], sess.run(get_next))
self.assertAllEqual([2, 2, 2, 2], sess.run(get_next))
self.assertAllEqual([0, 0, 0, 0], sess.run(get_next))
def group_by_window(key_func,
reduce_func,
window_size=None,
window_size_func=None):
"""A transformation that groups windows of elements by key and reduces them.
This transformation maps each consecutive element in a dataset to a key
using `key_func` and groups the elements by key. It then applies
`reduce_func` to at most `window_size_func(key)` elements matching the same
key. All except the final window for each key will contain
`window_size_func(key)` elements; the final window may be smaller.
You may provide either a constant `window_size` or a window size determined by
the key through `window_size_func`.
Args:
key_func: A function mapping a nested structure of tensors
(having shapes and types defined by `self.output_shapes` and
`self.output_types`) to a scalar `tf.int64` tensor.
reduce_func: A function mapping a key and a dataset of up to `window_size`
consecutive elements matching that key to another dataset.
window_size: A `tf.int64` scalar `tf.Tensor`, representing the number of
consecutive elements matching the same key to combine in a single
batch, which will be passed to `reduce_func`. Mutually exclusive with
`window_size_func`.
window_size_func: A function mapping a key to a `tf.int64` scalar
`tf.Tensor`, representing the number of consecutive elements matching
the same key to combine in a single batch, which will be passed to
`reduce_func`. Mutually exclusive with `window_size`.
Returns:
A `Dataset` transformation function, which can be passed to
`tf.data.Dataset.apply`.
Raises:
ValueError: if neither or both of {`window_size`, `window_size_func`} are
passed.
"""
return grouping.group_by_window(key_func, reduce_func, window_size,
window_size_func)
def testImmediateOutput(self):
components = np.array([0, 0, 0, 0, 1, 1, 1, 1, 2, 2, 0, 0, 2, 2, 0, 0],
dtype=np.int64)
iterator = (dataset_ops.Dataset.from_tensor_slices(components).repeat(
-1).apply(
grouping.group_by_window(lambda x: x % 3,
lambda _, xs: xs.batch(4),
4)).make_initializable_iterator())
init_op = iterator.initializer
get_next = iterator.get_next()
with self.cached_session() as sess:
sess.run(init_op)
# The input is infinite, so this test demonstrates that:
# 1. We produce output without having to consume the entire input,
# 2. Different buckets can produce output at different rates, and
# 3. For deterministic input, the output is deterministic.
for _ in range(3):
self.assertAllEqual([0, 0, 0, 0], sess.run(get_next))
self.assertAllEqual([1, 1, 1, 1], sess.run(get_next))
self.assertAllEqual([2, 2, 2, 2], sess.run(get_next))
self.assertAllEqual([0, 0, 0, 0], sess.run(get_next))
def testShard(self):
filename = self._createFile()
dataset = readers.TFRecordDataset([filename])
def reduce_func(key, dataset):
shard_filename = string_ops.string_join(
[filename, string_ops.as_string(key)])
writer = writers.TFRecordWriter(shard_filename)
writer.write(dataset.map(lambda _, x: x))
return dataset_ops.Dataset.from_tensors(shard_filename)
dataset = dataset.enumerate()
dataset = dataset.apply(
grouping.group_by_window(lambda i, _: i % 2, reduce_func,
dtypes.int64.max))
get_next = self.getNext(dataset)
for i in range(2):
shard_filename = (filename + str(i)).encode()
self.assertEqual(self.evaluate(get_next()), shard_filename)
for j, r in enumerate(
tf_record.tf_record_iterator(shard_filename)):
self.assertAllEqual(self._record(i + 2 * j), r)
def _build_dataset(self, components):
return dataset_ops.Dataset.from_tensor_slices(components).repeat(-1).apply(
grouping.group_by_window(lambda x: x % 3, lambda _, xs: xs.batch(4), 4))
