def test_invalid_num_lines_leads_to_disabling_progress_bar(num_lines, _nncf_caplog):
for _ in ProgressBar(TEST_RANGE, num_lines=num_lines):
pass
assert len(_nncf_caplog.records) == 1
record = next(iter(_nncf_caplog.records))
assert record.levelno == logging.WARNING
def test_invalid_total_leads_to_disabling_progress_bar(total, _nncf_caplog):
for _ in ProgressBar(TEST_RANGE, total=total):
pass
assert len(_nncf_caplog.records) == 1
record = next(iter(_nncf_caplog.records))
assert record.levelno == logging.WARNING
def test_can_iterate_with_warning_for_iterable_without_len(_nncf_caplog):
for _ in ProgressBar(enumerate(TEST_RANGE)):
pass
assert len(_nncf_caplog.records) == 1
record = next(iter(_nncf_caplog.records))
assert record.levelno == logging.WARNING
def test_can_limit_number_of_iterations(_nncf_caplog):
for _ in ProgressBar(TEST_RANGE, total=2):
pass
assert _nncf_caplog.record_tuples == [
('nncf', 20, ' ████████ | 1 / 2'),
('nncf', 20, ' ████████████████ | 2 / 2')
]
def test_can_print_by_default__with_enumerate_and_total(_nncf_caplog):
for _ in ProgressBar(enumerate(TEST_RANGE), total=3.0):
pass
assert _nncf_caplog.record_tuples == [
('nncf', 20, ' █████ | 1 / 3'),
('nncf', 20, ' ██████████ | 2 / 3'),
('nncf', 20, ' ████████████████ | 3 / 3')
]
def test_can_print_by_default(_nncf_caplog):
for _ in ProgressBar(TEST_RANGE):
pass
assert _nncf_caplog.record_tuples == [
('nncf', 20, ' █████ | 1 / 3'),
('nncf', 20, ' ██████████ | 2 / 3'),
('nncf', 20, ' ████████████████ | 3 / 3')
]
def test_can_print_collections_bigger_than_num_lines(_nncf_caplog):
for _ in ProgressBar(range(11), num_lines=3):
pass
assert _nncf_caplog.record_tuples == [
('nncf', 20, ' ███████ | 5 / 11'),
('nncf', 20, ' ██████████████ | 10 / 11'),
('nncf', 20, ' ████████████████ | 11 / 11')
]
def test_can_print_collections_less_than_num_lines(_nncf_caplog):
desc = 'desc'
for _ in ProgressBar(TEST_RANGE, desc=desc, num_lines=4):
pass
assert _nncf_caplog.record_tuples == [
('nncf', 20, 'desc █████ | 1 / 3'),
('nncf', 20, 'desc ██████████ | 2 / 3'),
('nncf', 20, 'desc ████████████████ | 3 / 3')
]
def _run_model_inference(self, data_loader, num_init_steps, device):
for i, loaded_item in ProgressBar(
enumerate(data_loader),
total=num_init_steps,
desc=self.progressbar_description,
):
if num_init_steps is not None and i >= num_init_steps:
break
args_kwargs_tuple = data_loader.get_inputs(loaded_item)
self._infer_batch(args_kwargs_tuple, device)
def test_can_print_with_another_logger(_nncf_caplog):
name = "test"
logger = logging.getLogger(name)
logger.setLevel(logging.INFO)
for _ in ProgressBar(enumerate(TEST_RANGE), logger=logger, total=2):
pass
assert _nncf_caplog.record_tuples == [
('test', 20, ' ████████ | 1 / 2'),
('test', 20, ' ████████████████ | 2 / 2')
]
def run(self, model: tf.keras.Model) -> None:
"""
Runs the batch-norm statistics adaptation algorithm.
:param model: A model for which the algorithm will be applied.
"""
if self._device is not None:
raise ValueError(
'TF implementation of batchnorm adaptation algorithm '
'does not support switch of devices. Model initial device '
'is used by default for batchnorm adaptation.')
with BNTrainingStateSwitcher(model):
for (x, _) in ProgressBar(islice(self._data_loader,
self._num_bn_adaptation_steps),
total=self._num_bn_adaptation_steps,
desc='BatchNorm statistics adaptation'):
model(x, training=True)
def run(self, model: tf.keras.Model) -> None:
layer_statistics = []
op_statistics = []
handles = []
for layer in model.layers:
if isinstance(layer, FakeQuantize):
self._register_layer_statistics(layer, layer_statistics,
handles)
elif isinstance(layer, NNCFWrapper):
self._register_op_statistics(layer, op_statistics, handles)
for (x, _) in ProgressBar(islice(self.dataset, self.num_steps),
total=self.num_steps,
desc='Collecting tensor statistics/data'):
model(x, training=False)
for layer, collector in layer_statistics:
target_stat = collector.get_statistics()
minmax_stats = tf_convert_stat_to_min_max_tensor_stat(target_stat)
layer.apply_range_initialization(
tf.squeeze(minmax_stats.min_values),
tf.squeeze(minmax_stats.max_values))
layer.enabled = True
for layer, op_name, op, collector in op_statistics:
weights = layer.get_operation_weights(op_name)
target_stat = collector.get_statistics()
minmax_stats = tf_convert_stat_to_min_max_tensor_stat(target_stat)
min_values = minmax_stats.min_values
if len(min_values.shape) != 1:
min_values = tf.squeeze(min_values)
max_values = minmax_stats.max_values
if len(max_values.shape) != 1:
max_values = tf.squeeze(max_values)
op.apply_range_initialization(weights, min_values, max_values)
op.enabled = True
for handle in handles:
handle.remove()
for x, _ in self.dataset:
model(x, training=False)
break
def test_type_error_happens_for_iteration_none(_nncf_caplog):
with pytest.raises(TypeError):
for _ in ProgressBar(None):
pass
def test_can_iterate_over_empty_iterable(caplog):
for _ in ProgressBar([]):
pass
assert caplog.record_tuples == []