def test_multibranch_propagation_with_fq_moving():
TEST_CASES_PATH = TEST_ROOT / 'data' / 'test_cases_refs'
model_path = (TEST_CASES_PATH /
'test_ig_border_case_with_fq_moving.xml').as_posix()
weights_path = (TEST_CASES_PATH /
'test_ig_border_case_with_fq_moving.bin').as_posix()
ignored_params = {
"scope": [
'8/WithoutBiases', '9/WithoutBiases', '10/WithoutBiases',
'11/WithoutBiases'
]
}
config = Dict({'model': model_path, 'weights': weights_path})
model = load_model(config)
hardware_config = HardwareConfig.from_json(
(HARDWARE_CONFIG_PATH / 'cpu.json').as_posix())
quantized_model = GraphTransformer(hardware_config).insert_fake_quantize(
model, ignored_params)
node = get_node_by_name(quantized_model, '14')
for node_input in get_node_inputs(node)[:2]:
assert node_input.type == 'FakeQuantize'
assert get_node_inputs(node)[2].type == 'Concat'
node = get_node_by_name(quantized_model, '12')
for node_input in get_node_inputs(node)[:2]:
assert node_input.type == 'FakeQuantize'
assert len(get_nodes_by_type(quantized_model, ['FakeQuantize'])) == 6
def cut_fq_node(model, node_list, graph_transformer, tmp_path):
model_ = load_model(model.model_params)
quantized_model = graph_transformer.insert_fake_quantize(model_)
cropped_model = quantized_model
for node_name in node_list:
node = get_node_by_name(cropped_model, node_name)
for parent_node in nu.get_node_inputs(node):
if parent_node and parent_node and parent_node.type == 'FakeQuantize':
cropped_model, *_ = graph_transformer.remove_fq_nodes(quantized_model, [parent_node.name])
break
check_model(tmp_path, cropped_model, model.model_name + '_cut_fq', model.framework)
def test_multibranch_propagation_with_fq_moving(tmp_path, models, model_name, model_framework):
ignored_params = {
"scope": ['Convolution_104', 'Convolution_152', 'Convolution_8', 'Convolution_56']
}
model = models.get(model_name, model_framework, tmp_path)
model = load_model(model.model_params)
hardware_config = HardwareConfig.from_json((HARDWARE_CONFIG_PATH / 'cpu.json').as_posix())
quantized_model = GraphTransformer(hardware_config).insert_fake_quantize(model, ignored_params)
node = get_node_by_name(quantized_model, '14')
for node_input in get_node_inputs(node)[:2]:
assert node_input.type == 'FakeQuantize'
assert get_node_inputs(node)[2].type == 'Concat'
node = get_node_by_name(quantized_model, '12')
for node_input in get_node_inputs(node)[:2]:
assert node_input.type == 'FakeQuantize'
assert len(get_nodes_by_type(quantized_model, ['FakeQuantize'])) == 6
def _test_unify_scales(model_, to_unify_):
for _, fqs in to_unify_:
ranges = []
for fq in fqs:
fq = get_node_by_name(model_, fq)
fq_inputs = nu.get_node_inputs(fq)[1:]
ranges.append(
tuple(
fqut.get_node_value(fq_input)
for fq_input in fq_inputs))
assert all([
np.array_equal(r, ranges[0][i])
for i, r in enumerate(ranges[-1])
])