def test_composite_reader(self):
ws = workspace.C.Workspace()
session = LocalSession(ws)
num_srcs = 3
names = ["src_{}".format(i) for i in range(num_srcs)]
size = 100
offsets = [i * size for i in range(num_srcs)]
src_dses = [make_source_dataset(ws, offset=offset, size=size, name=name)
for (name, offset) in zip(names, offsets)]
data = [ws.fetch_blob(str(src.field_blobs[0])) for src in src_dses]
# Sanity check we didn't overwrite anything
for d, offset in zip(data, offsets):
npt.assert_array_equal(d, range(offset, offset + size))
# Make an identically-sized empty destnation dataset
dst_ds_schema = schema.Struct(
*[
(name, src_ds.content().clone_schema())
for name, src_ds in zip(names, src_dses)
]
)
dst_ds = make_destination_dataset(ws, dst_ds_schema)
with TaskGroup() as tg:
reader = CompositeReader(names,
[src_ds.reader() for src_ds in src_dses])
pipe(reader, dst_ds.writer(), num_runtime_threads=3)
session.run(tg)
for i in range(num_srcs):
written_data = sorted(
ws.fetch_blob(str(dst_ds.content()[names[i]].label())))
npt.assert_array_equal(data[i], written_data, "i: {}".format(i))
def build_cache_step(self, overwrite=False):
"""Build a step for generating cache DB file.
If self.db_path exists and not overwritting, build an empty step.
Overwise, build a step as follows.
Pipe original reader to the _DatasetWriter,
so that dataset field blobs are populated.
Then save these blobs into a file.
Args:
overwrite: bool. If true, ignore the existing file
and build a new one overwritting the existing one anyway.
Returns:
build_cache_step: ExcutionStep.
The step to be run for building a cache DB file.
"""
if os.path.exists(self.db_path) and not overwrite:
# cache already exists, no need to rebuild it
return core.execution_step('build_step', [])
init_net = core.Net('init')
self._init_field_blobs_as_empty(init_net)
with Cluster(), core.NameScope(self.name), TaskGroup() as copy_tg:
pipe(self.original_reader, self.ds.writer(), num_threads=16)
copy_step = copy_tg.to_task().get_step()
save_net = core.Net('save')
self._save_field_blobs_to_db_file(save_net)
return core.execution_step('build_cache', [init_net, copy_step, save_net])
def test_composite_reader_builder(self):
ws = workspace.C.Workspace()
session = LocalSession(ws)
num_srcs = 3
names = ["src_{}".format(i) for i in range(num_srcs)]
size = 100
offsets = [i * size for i in range(num_srcs)]
src_ds_builders = [
TestReaderBuilder(offset=offset, size=size, name=name)
for (name, offset) in zip(names, offsets)
]
# Make an identically-sized empty destnation dataset
dst_ds_schema = schema.Struct(
*[
(name, src_ds_builder.schema())
for name, src_ds_builder in zip(names, src_ds_builders)
]
)
dst_ds = make_destination_dataset(ws, dst_ds_schema)
with TaskGroup() as tg:
reader_builder = CompositeReaderBuilder(
names, src_ds_builders)
reader_builder.setup(ws=ws)
pipe(reader_builder.new_reader(), dst_ds.writer(),
num_runtime_threads=3)
session.run(tg)
for name, offset in zip(names, offsets):
written_data = sorted(
ws.fetch_blob(str(dst_ds.content()[name].label())))
npt.assert_array_equal(range(offset, offset + size), written_data,
"name: {}".format(name))
def test_composite_reader_builder(self):
ws = workspace.C.Workspace()
session = LocalSession(ws)
num_srcs = 3
names = ["src_{}".format(i) for i in range(num_srcs)]
size = 100
offsets = [i * size for i in range(num_srcs)]
src_ds_builders = [
TestReaderBuilder(offset=offset, size=size, name=name)
for (name, offset) in zip(names, offsets)
]
# Make an identically-sized empty destnation dataset
dst_ds_schema = schema.Struct(
*[(name, src_ds_builder.schema())
for name, src_ds_builder in zip(names, src_ds_builders)])
dst_ds = make_destination_dataset(ws, dst_ds_schema)
with TaskGroup() as tg:
reader_builder = CompositeReaderBuilder(names, src_ds_builders)
reader_builder.setup(ws=ws)
pipe(reader_builder.new_reader(),
dst_ds.writer(),
num_runtime_threads=3)
session.run(tg)
for name, offset in zip(names, offsets):
written_data = sorted(
ws.fetch_blob(str(dst_ds.content()[name].label())))
npt.assert_array_equal(range(offset, offset + size), written_data,
"name: {}".format(name))
def build_cache_step(self, overwrite=False):
"""Build a step for generating cache DB file.
If self.db_path exists and not overwritting, build an empty step.
Overwise, build a step as follows.
Pipe original reader to the _DatasetWriter,
so that dataset field blobs are populated.
Then save these blobs into a file.
Args:
overwrite: bool. If true, ignore the existing file
and build a new one overwritting the existing one anyway.
Returns:
build_cache_step: ExecutionStep.
The step to be run for building a cache DB file.
"""
if os.path.exists(self.db_path) and not overwrite:
# cache already exists, no need to rebuild it
return core.execution_step('build_step', [])
init_net = core.Net('init')
self._init_field_blobs_as_empty(init_net)
with Cluster(), core.NameScope(self.name), TaskGroup() as copy_tg:
pipe(self.original_reader, self.ds.writer(), num_threads=16)
copy_step = copy_tg.to_task().get_step()
save_net = core.Net('save')
self._save_field_blobs_to_db_file(save_net)
return core.execution_step('build_cache',
[init_net, copy_step, save_net])
def build_job(node_id):
all_outputs = []
with Job() as job:
with Node('reader' + str(node_id)):
with job.init_group:
init_net = core.Net('init_net' + str(node_id))
data_arr = Struct(('val', np.array(range(10))))
data = ConstRecord(init_net, data_arr)
ds = Dataset(data, name='dataset' + str(node_id))
full_reader = ds.reader(init_net)
total = init_net.Const([100])
Task(step=init_net)
def inc_total(rec):
net = core.Net('inc_total' + str(node_id))
net.Add([total, rec.val()], [total])
return [net]
epoch_reader = ReaderWithLimit(full_reader, num_iter=3)
pipe(epoch_reader, processor=inc_total)
job.add_stop_signal(epoch_reader.data_finished())
all_outputs.append(total)
total_fetcher = Task(step=core.Net('empty'), outputs=all_outputs)
return job, total_fetcher
def test_composite_reader(self):
ws = workspace.C.Workspace()
session = LocalSession(ws)
num_srcs = 3
names = ["src_{}".format(i) for i in range(num_srcs)]
size = 100
offsets = [i * size for i in range(num_srcs)]
src_dses = [
make_source_dataset(ws, offset=offset, size=size, name=name)
for (name, offset) in zip(names, offsets)
]
data = [ws.fetch_blob(str(src.field_blobs[0])) for src in src_dses]
# Sanity check we didn't overwrite anything
for d, offset in zip(data, offsets):
npt.assert_array_equal(d, range(offset, offset + size))
# Make an identically-sized empty destnation dataset
dst_ds_schema = schema.Struct(
*[(name, src_ds.content().clone_schema())
for name, src_ds in zip(names, src_dses)])
dst_ds = make_destination_dataset(ws, dst_ds_schema)
with TaskGroup() as tg:
reader = CompositeReader(names,
[src_ds.reader() for src_ds in src_dses])
pipe(reader, dst_ds.writer(), num_runtime_threads=3)
session.run(tg)
for i in range(num_srcs):
written_data = sorted(
ws.fetch_blob(str(dst_ds.content()[names[i]].label())))
npt.assert_array_equal(data[i], written_data, "i: {}".format(i))
def test_reader_with_limit(self):
ws = workspace.C.Workspace()
session = LocalSession(ws)
""" 1. feed full dataset """
src_init = core.Net('src_init')
src_values = Struct(('label', np.array(range(100))))
src_blobs = NewRecord(src_init, src_values)
src_ds = Dataset(src_blobs)
FeedRecord(src_blobs, src_values, ws)
ws.run(src_init)
""" 2. Read with limit smaller than size of dataset """
dst_init = core.Net('dst_init')
dst_ds = Dataset(src_values.clone_schema())
dst_ds.init_empty(dst_init)
ws.run(dst_init)
with TaskGroup() as tg:
reader = ReaderWithLimit(src_ds.reader(), num_iter=10)
pipe(reader, dst_ds.writer(), num_threads=8)
session.run(tg)
self.assertFalse(ws.blobs[str(reader.data_finished())].fetch())
self.assertEquals(
sorted(ws.blobs[str(dst_ds.content().label())].fetch()), range(10))
""" 3. Read with limit larger than size of dataset """
ws.run(dst_init)
with TaskGroup() as tg:
reader = ReaderWithLimit(src_ds.reader(), num_iter=110)
pipe(reader, dst_ds.writer(), num_threads=8)
session.run(tg)
self.assertEquals(
sorted(ws.blobs[str(dst_ds.content().label())].fetch()),
range(100))
self.assertTrue(ws.blobs[str(reader.data_finished())].fetch())
def _test_limit_reader_shared(self, reader_class, size, expected_read_len,
expected_finish, num_threads, read_delay,
**limiter_args):
ws, session, src_ds, dst_init, dst_ds = \
self._test_limit_reader_init_shared(size)
# Read without limiter
# WorkspaceType.GLOBAL is required because we are fetching
# reader.data_finished() after the TaskGroup finishes.
with TaskGroup(workspace_type=WorkspaceType.GLOBAL) as tg:
if read_delay > 0:
reader = reader_class(ReaderWithDelay(src_ds.reader(),
read_delay),
**limiter_args)
else:
reader = reader_class(src_ds.reader(), **limiter_args)
pipe(reader, dst_ds.writer(), num_runtime_threads=num_threads)
session.run(tg)
read_len = len(sorted(ws.blobs[str(dst_ds.content().label())].fetch()))
self.assertEqual(read_len, expected_read_len)
self.assertEqual(
sorted(ws.blobs[str(dst_ds.content().label())].fetch()),
list(range(expected_read_len))
)
self.assertEqual(ws.blobs[str(reader.data_finished())].fetch(),
expected_finish)
def test_reader_with_limit(self):
ws = workspace.C.Workspace()
session = LocalSession(ws)
""" 1. feed full dataset """
src_init = core.Net('src_init')
src_values = Struct(('label', np.array(range(100))))
src_blobs = NewRecord(src_init, src_values)
src_ds = Dataset(src_blobs)
FeedRecord(src_blobs, src_values, ws)
ws.run(src_init)
""" 2. Read with limit smaller than size of dataset """
dst_init = core.Net('dst_init')
dst_ds = Dataset(src_values.clone_schema())
dst_ds.init_empty(dst_init)
ws.run(dst_init)
with TaskGroup() as tg:
reader = ReaderWithLimit(src_ds.reader(), num_iter=10)
pipe(reader, dst_ds.writer(), num_threads=8)
session.run(tg)
self.assertFalse(ws.blobs[str(reader.data_finished())].fetch())
self.assertEquals(
sorted(ws.blobs[str(dst_ds.content().label())].fetch()), range(10))
""" 3. Read with limit larger than size of dataset """
ws.run(dst_init)
with TaskGroup() as tg:
reader = ReaderWithLimit(src_ds.reader(), num_iter=110)
pipe(reader, dst_ds.writer(), num_threads=8)
session.run(tg)
self.assertEquals(
sorted(ws.blobs[str(dst_ds.content().label())].fetch()), range(100))
self.assertTrue(ws.blobs[str(reader.data_finished())].fetch())
def _test_limit_reader_shared(self, reader_class, size, expected_read_len,
expected_finish, num_threads, read_delay,
**limiter_args):
ws, session, src_ds, dst_ds = \
self._test_limit_reader_init_shared(size)
# Read without limiter
# WorkspaceType.GLOBAL is required because we are fetching
# reader.data_finished() after the TaskGroup finishes.
with TaskGroup(workspace_type=WorkspaceType.GLOBAL) as tg:
if read_delay > 0:
reader = reader_class(
ReaderWithDelay(src_ds.reader(), read_delay),
**limiter_args)
else:
reader = reader_class(src_ds.reader(), **limiter_args)
pipe(reader, dst_ds.writer(), num_runtime_threads=num_threads)
session.run(tg)
read_len = len(sorted(ws.blobs[str(dst_ds.content().label())].fetch()))
self.assertEqual(read_len, expected_read_len)
self.assertEqual(
sorted(ws.blobs[str(dst_ds.content().label())].fetch()),
list(range(expected_read_len)))
self.assertEqual(ws.blobs[str(reader.data_finished())].fetch(),
expected_finish)
def _read_all_data(ws, reader, session):
dst_ds = make_destination_dataset(ws, reader.schema().clone_schema())
with TaskGroup() as tg:
pipe(reader, dst_ds.writer(), num_runtime_threads=8)
session.run(tg)
return ws.blobs[str(dst_ds.content().label())].fetch()
def _read_all_data(ws, reader, session):
dst_ds = make_destination_dataset(ws, reader.schema().clone_schema())
with TaskGroup() as tg:
pipe(reader, dst_ds.writer(), num_runtime_threads=8)
session.run(tg)
return ws.blobs[str(dst_ds.content().label())].fetch()
def test_dequeue_many(self):
init_net = core.Net('init')
N = 17
NUM_DEQUEUE_RECORDS = 3
src_values = Struct(
('uid', np.array(range(N))),
('value', 0.1 * np.array(range(N))))
expected_dst = Struct(
('uid', 2 * np.array(range(N))),
('value', np.array(N * [0.0])))
with core.NameScope('init'):
src_blobs = NewRecord(init_net, src_values)
dst_blobs = InitEmptyRecord(init_net, src_values.clone_schema())
counter = init_net.Const(0)
ONE = init_net.Const(1)
def proc1(rec):
with core.NameScope('proc1'):
out = NewRecord(ops, rec)
ops.Add([rec.uid(), rec.uid()], [out.uid()])
out.value.set(blob=rec.value(), unsafe=True)
return out
def proc2(rec):
with core.NameScope('proc2'):
out = NewRecord(ops, rec)
out.uid.set(blob=rec.uid(), unsafe=True)
ops.Sub([rec.value(), rec.value()], [out.value()])
ops.Add([counter, ONE], [counter])
return out
src_ds = Dataset(src_blobs)
dst_ds = Dataset(dst_blobs)
with TaskGroup() as tg:
out1 = pipe(
src_ds.reader(),
output=Queue(
capacity=11, num_dequeue_records=NUM_DEQUEUE_RECORDS),
processor=proc1)
out2 = pipe(out1, processor=proc2)
pipe(out2, dst_ds.writer())
ws = workspace.C.Workspace()
FeedRecord(src_blobs, src_values, ws)
session = LocalSession(ws)
session.run(init_net)
session.run(tg)
output = FetchRecord(dst_blobs, ws=ws)
num_dequeues = ws.blobs[str(counter)].fetch()
self.assertEquals(
num_dequeues, int(math.ceil(float(N) / NUM_DEQUEUE_RECORDS)))
for a, b in zip(output.field_blobs(), expected_dst.field_blobs()):
np.testing.assert_array_equal(a, b)
def test_dequeue_many(self):
init_net = core.Net('init')
N = 17
NUM_DEQUEUE_RECORDS = 3
src_values = Struct(
('uid', np.array(range(N))),
('value', 0.1 * np.array(range(N))))
expected_dst = Struct(
('uid', 2 * np.array(range(N))),
('value', np.array(N * [0.0])))
with core.NameScope('init'):
src_blobs = NewRecord(init_net, src_values)
dst_blobs = InitEmptyRecord(init_net, src_values.clone_schema())
counter = init_net.Const(0)
ONE = init_net.Const(1)
def proc1(rec):
with core.NameScope('proc1'):
out = NewRecord(ops, rec)
ops.Add([rec.uid(), rec.uid()], [out.uid()])
out.value.set(blob=rec.value(), unsafe=True)
return out
def proc2(rec):
with core.NameScope('proc2'):
out = NewRecord(ops, rec)
out.uid.set(blob=rec.uid(), unsafe=True)
ops.Sub([rec.value(), rec.value()], [out.value()])
ops.Add([counter, ONE], [counter])
return out
src_ds = Dataset(src_blobs)
dst_ds = Dataset(dst_blobs)
with TaskGroup() as tg:
out1 = pipe(
src_ds.reader(),
output=Queue(
capacity=11, num_dequeue_records=NUM_DEQUEUE_RECORDS),
processor=proc1)
out2 = pipe(out1, processor=proc2)
pipe(out2, dst_ds.writer())
ws = workspace.C.Workspace()
FeedRecord(src_blobs, src_values, ws)
session = LocalSession(ws)
session.run(init_net)
session.run(tg)
output = FetchRecord(dst_blobs, ws=ws)
num_dequeues = ws.blobs[str(counter)].fetch()
self.assertEquals(
num_dequeues, int(math.ceil(float(N) / NUM_DEQUEUE_RECORDS)))
for a, b in zip(output.field_blobs(), expected_dst.field_blobs()):
np.testing.assert_array_equal(a, b)
def read_all_data(ws, reader, session):
dst_init = core.Net('dst_init')
with core.NameScope('dst'):
dst_ds = Dataset(reader.schema().clone_schema())
dst_ds.init_empty(dst_init)
session.run(dst_init)
with TaskGroup(workspace_type=WorkspaceType.GLOBAL) as tg:
pipe(reader, dst_ds.writer(), num_runtime_threads=8)
session.run(tg)
return ws.blobs[str(dst_ds.content().label())].fetch()
def read_all_data(ws, reader, session):
dst_init = core.Net('dst_init')
with core.NameScope('dst'):
dst_ds = Dataset(reader.schema().clone_schema())
dst_ds.init_empty(dst_init)
session.run(dst_init)
with TaskGroup(workspace_type=WorkspaceType.GLOBAL) as tg:
pipe(reader, dst_ds.writer(), num_runtime_threads=8)
session.run(tg)
return ws.blobs[str(dst_ds.content().label())].fetch()
def build_pipeline(node_id):
with Node('trainer_%d' % node_id):
with Job.current().init_group, Task():
data_arr = Struct(('val', np.array(list(range(10)))))
data = ConstRecord(ops, data_arr)
ds = Dataset(data, name='dataset:%d' % node_id)
full_reader = ds.reader(ops)
total = ops.Const([100])
def inc_total(rec):
ops.Add([total, rec.val()], [total])
epoch_reader = ReaderWithLimit(full_reader, num_iter=3)
pipe(epoch_reader, processor=inc_total)
Job.current().add_stop_signal(epoch_reader.data_finished())
return [total]
def build_pipeline(node_id):
with Node('trainer_%d' % node_id):
with Job.current().init_group, Task():
data_arr = Struct(('val', np.array(list(range(10)))))
data = ConstRecord(ops, data_arr)
ds = Dataset(data, name='dataset:%d' % node_id)
full_reader = ds.reader(ops)
total = ops.Const([100])
def inc_total(rec):
ops.Add([total, rec.val()], [total])
epoch_reader = ReaderWithLimit(full_reader, num_iter=3)
pipe(epoch_reader, processor=inc_total)
Job.current().add_stop_signal(epoch_reader.data_finished())
return [total]
def build_cache(self, cache_path, overwrite=False):
if not self.has_cache() or overwrite:
self.cache_path = cache_path
if self.has_cache() and not overwrite:
# cache already exists, no need to rebuild it
return core.execution_step('build_step', [])
init_net = core.Net('init')
self._init_dataset(init_net)
with Cluster(), core.NameScope(self.name), TaskGroup() as copy_tg:
pipe(self.original_reader, self.ds.writer(), num_threads=16)
copy_step = copy_tg.to_task().get_step()
save_net = core.Net('save')
self._save_to_file(save_net)
return core.execution_step('build_cache', [init_net, copy_step, save_net])
def test_local_session(self):
init_net = core.Net('init')
src_values = Struct(
('uid', np.array([1, 2, 6])),
('value', np.array([1.4, 1.6, 1.7])))
expected_dst = Struct(
('uid', np.array([2, 4, 12])),
('value', np.array([0.0, 0.0, 0.0])))
with core.NameScope('init'):
src_blobs = NewRecord(init_net, src_values)
dst_blobs = InitEmptyRecord(init_net, src_values.clone_schema())
def proc1(rec):
net = core.Net('proc1')
with core.NameScope('proc1'):
out = NewRecord(net, rec)
net.Add([rec.uid(), rec.uid()], [out.uid()])
out.value.set(blob=rec.value(), unsafe=True)
return [net], out
def proc2(rec):
net = core.Net('proc2')
with core.NameScope('proc2'):
out = NewRecord(net, rec)
out.uid.set(blob=rec.uid(), unsafe=True)
net.Sub([rec.value(), rec.value()], [out.value()])
return [net], out
src_ds = Dataset(src_blobs)
dst_ds = Dataset(dst_blobs)
with TaskGroup() as tg:
out1 = pipe(src_ds.reader(), processor=proc1)
out2 = pipe(out1, processor=proc2)
pipe(out2, dst_ds.writer())
ws = workspace.C.Workspace()
FeedRecord(src_blobs, src_values, ws)
session = LocalSession(ws)
session.run(init_net)
session.run(tg)
output = FetchRecord(dst_blobs, ws=ws)
for a, b in zip(output.field_blobs(), expected_dst.field_blobs()):
np.testing.assert_array_equal(a, b)
def test_local_session(self):
init_net = core.Net('init')
src_values = Struct(
('uid', np.array([1, 2, 6])),
('value', np.array([1.4, 1.6, 1.7])))
expected_dst = Struct(
('uid', np.array([2, 4, 12])),
('value', np.array([0.0, 0.0, 0.0])))
with core.NameScope('init'):
src_blobs = NewRecord(init_net, src_values)
dst_blobs = InitEmptyRecord(init_net, src_values.clone_schema())
def proc1(rec):
net = core.Net('proc1')
with core.NameScope('proc1'):
out = NewRecord(net, rec)
net.Add([rec.uid(), rec.uid()], [out.uid()])
out.value.set(blob=rec.value(), unsafe=True)
return [net], out
def proc2(rec):
net = core.Net('proc2')
with core.NameScope('proc2'):
out = NewRecord(net, rec)
out.uid.set(blob=rec.uid(), unsafe=True)
net.Sub([rec.value(), rec.value()], [out.value()])
return [net], out
src_ds = Dataset(src_blobs)
dst_ds = Dataset(dst_blobs)
with TaskGroup() as tg:
out1 = pipe(src_ds.reader(), processor=proc1)
out2 = pipe(out1, processor=proc2)
pipe(out2, dst_ds.writer())
ws = workspace.C.Workspace()
FeedRecord(src_blobs, src_values, ws)
session = LocalSession(ws)
session.run(init_net)
session.run(tg)
output = FetchRecord(dst_blobs, ws=ws)
for a, b in zip(output.field_blobs(), expected_dst.field_blobs()):
np.testing.assert_array_equal(a, b)
def build_job():
with Node('reader'):
with Job() as job:
with job.init_group:
init_net = core.Net('init_net')
data_arr = Struct(('val', np.array(range(10))))
data = ConstRecord(init_net, data_arr)
ds = Dataset(data)
full_reader = ds.reader(init_net)
total = init_net.Const([100])
Task(step=init_net)
def inc_total(rec):
net = core.Net('inc_total')
net.Add([total, rec.val()], [total])
return [net]
epoch_reader = ReaderWithLimit(full_reader, num_iter=3)
pipe(epoch_reader, processor=inc_total)
job.add_stop_signal(epoch_reader.data_finished())
total_fetcher = Task(step=core.Net('empty'), outputs=[total])
return job, total_fetcher
def test_reader_with_limit(self):
ws = workspace.C.Workspace()
session = LocalSession(ws)
""" 1. feed full dataset """
src_ds = init_dataset(ws)
""" 2. Read with limit smaller than size of dataset """
dst_init = core.Net('dst_init')
with core.NameScope('dst'):
dst_ds = Dataset(src_ds.content().clone_schema())
dst_ds.init_empty(dst_init)
ws.run(dst_init)
# WorkspaceType.GLOBAL is required because we are fetching
# reader.data_finished() after the TaskGroup finishes.
with TaskGroup(workspace_type=WorkspaceType.GLOBAL) as tg:
reader = ReaderWithLimit(src_ds.reader(), num_iter=10)
pipe(reader, dst_ds.writer(), num_threads=8)
session.run(tg)
self.assertFalse(ws.blobs[str(reader.data_finished())].fetch())
self.assertEquals(
sorted(ws.blobs[str(dst_ds.content().label())].fetch()),
list(range(10))
)
""" 3. Read with limit larger than size of dataset """
ws.run(dst_init)
with TaskGroup(workspace_type=WorkspaceType.GLOBAL) as tg:
reader = ReaderWithLimit(src_ds.reader(), num_iter=110)
pipe(reader, dst_ds.writer(), num_runtime_threads=8)
session.run(tg)
self.assertEquals(
sorted(ws.blobs[str(dst_ds.content().label())].fetch()),
list(range(100))
)
self.assertTrue(ws.blobs[str(reader.data_finished())].fetch())
""" 4. Read without counter """
ws.run(dst_init)
with TaskGroup(workspace_type=WorkspaceType.GLOBAL) as tg:
reader = ReaderWithLimit(src_ds.reader(), num_iter=None)
pipe(reader, dst_ds.writer(), num_threads=8)
session.run(tg)
self.assertEquals(
sorted(ws.blobs[str(dst_ds.content().label())].fetch()),
list(range(100))
)
self.assertTrue(ws.blobs[str(reader.data_finished())].fetch())
""" 5. Read using the same reader without resetting workspace """
session.run(tg)
self.assertEquals(
sorted(ws.blobs[str(dst_ds.content().label())].fetch()),
sorted(list(range(100)) * 2)
)
def test_runtime_threads(self):
ws = workspace.C.Workspace()
session = LocalSession(ws)
src_ds = init_dataset(ws)
totals = [None] * 3
def proc(rec):
# executed once
with ops.task_init():
counter1 = ops.CreateCounter([], ['global_counter'])
counter2 = ops.CreateCounter([], ['global_counter2'])
counter3 = ops.CreateCounter([], ['global_counter3'])
# executed once per thread
with ops.task_instance_init():
task_counter = ops.CreateCounter([], ['task_counter'])
# executed on each iteration
ops.CountUp(counter1)
ops.CountUp(task_counter)
# executed once per thread
with ops.task_instance_exit():
with ops.loop(ops.RetrieveCount(task_counter)):
ops.CountUp(counter2)
ops.CountUp(counter3)
# executed once
with ops.task_exit():
totals[0] = final_output(ops.RetrieveCount(counter1))
totals[1] = final_output(ops.RetrieveCount(counter2))
totals[2] = final_output(ops.RetrieveCount(counter3))
return rec
# Read full data set from original reader
with TaskGroup() as tg:
pipe(src_ds.reader(), num_runtime_threads=8, processor=proc)
session.run(tg)
self.assertEqual(totals[0].fetch(), 100)
self.assertEqual(totals[1].fetch(), 100)
self.assertEqual(totals[2].fetch(), 8)
# Read with a count-limited reader
with TaskGroup() as tg:
q1 = pipe(src_ds.reader(), num_runtime_threads=2)
q2 = pipe(
ReaderWithLimit(q1.reader(), num_iter=25),
num_runtime_threads=3)
pipe(q2, processor=proc, num_runtime_threads=6)
session.run(tg)
self.assertEqual(totals[0].fetch(), 25)
self.assertEqual(totals[1].fetch(), 25)
self.assertEqual(totals[2].fetch(), 6)
def test_runtime_threads(self):
ws = workspace.C.Workspace()
session = LocalSession(ws)
src_ds = init_dataset(ws)
totals = [None] * 3
def proc(rec):
# executed once
with ops.task_init():
counter1 = ops.CreateCounter([], ['global_counter'])
counter2 = ops.CreateCounter([], ['global_counter2'])
counter3 = ops.CreateCounter([], ['global_counter3'])
# executed once per thread
with ops.task_instance_init():
task_counter = ops.CreateCounter([], ['task_counter'])
# executed on each iteration
ops.CountUp(counter1)
ops.CountUp(task_counter)
# executed once per thread
with ops.task_instance_exit():
with ops.loop(ops.RetrieveCount(task_counter)):
ops.CountUp(counter2)
ops.CountUp(counter3)
# executed once
with ops.task_exit():
totals[0] = final_output(ops.RetrieveCount(counter1))
totals[1] = final_output(ops.RetrieveCount(counter2))
totals[2] = final_output(ops.RetrieveCount(counter3))
return rec
""" 1. Feed full dataset """
with TaskGroup() as tg:
pipe(src_ds.reader(), num_runtime_threads=8, processor=proc)
session.run(tg)
self.assertEquals(totals[0].fetch(), 100)
self.assertEquals(totals[1].fetch(), 100)
self.assertEquals(totals[2].fetch(), 8)
""" 2. Add a few steps in between """
with TaskGroup() as tg:
q1 = pipe(src_ds.reader(), num_runtime_threads=2)
q2 = pipe(
ReaderWithLimit(q1.reader(), num_iter=25),
num_runtime_threads=3)
pipe(q2, processor=proc, num_runtime_threads=6)
session.run(tg)
self.assertEquals(totals[0].fetch(), 25)
self.assertEquals(totals[1].fetch(), 25)
self.assertEquals(totals[2].fetch(), 6)
