IntenseWebs/openvino
3
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

[POT] get_num_levels function (#8393)

Browse Source 
* added function for calculating the number of discret levels in the input tensors and tests for this function

* added function for calculating the number of discret levels in the input tensors and tests for this function

* fixed pylint issues

* changed the function for delta estimation from mean to min

* added empty delta array processing in get_num_levels func and tests for it
This commit is contained in:
Tagir Rakipov 2021-11-10 15:27:09 +03:00 committed by GitHub
parent dd5efdca85
commit ac0582c2d9
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
2 changed files with 37 additions and 11 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

21
tools/pot/openvino/tools/pot/algorithms/quantization/fake_quantize.py
View File

@ -450,3 +450,24 @@ def create_renamed_layers_mapping(model, stats_layout):
name_change_to = node['orig_node_name'] if port_id is None else (node['orig_node_name'], port_id)
changed_names_map[layer_name] = name_change_to
return changed_names_map
def get_num_levels(x: np.ndarray) -> int:
"""
Calculates the number of discret levels of the values
in the input NumPy tensor x
:param x: the input tensor
:return the number of discret value levels in the input tensor x
"""
NUM_BINS = 256
x = x.flatten()
hist, _ = np.histogram(x, NUM_BINS)
non_empty_bins = [i for i, v in enumerate(hist) if v > 0]
deltas = [non_empty_bins[i]-non_empty_bins[i-1] for i in range(1, len(non_empty_bins))]
if deltas == []:
return 0
d = min(deltas)
if d == 1:
return -1
return round(NUM_BINS / d)

27
tools/pot/tests/test_tensor_statistics.py
View File

@ -7,6 +7,8 @@ import pytest
from openvino.tools.pot.statistics.function_selector import AGGREGATION_FN, ACTIVATIONS_STATS_FN, WEIGHTS_STATS_FN, \
get_aggregation_function, get_stats_function_for_activations, get_stats_function_for_weights, PERCHANNEL, PERTENSOR
from openvino.tools.pot.algorithms.quantization.fake_quantize import get_num_levels
INPUT_SHAPES = [(2, 2, 1), (2, 2, 2)]
AGG_INPUTS = [np.reshape(np.array(range(np.prod(shape)), dtype=np.float32), shape) for shape in INPUT_SHAPES]
@ -137,14 +139,17 @@ WEIGHTS_CH_STATS_FUNCTIONS = [(name, True) for name in
WEIGHTS_STATS_FN[PERCHANNEL].registry_dict.keys()]
@pytest.mark.parametrize(
'name, transpose', WEIGHTS_CH_STATS_FUNCTIONS,
ids=['{}_{}'.format(fn[0], fn[1]) for fn in WEIGHTS_CH_STATS_FUNCTIONS])
def test_weights_transpose_function(name, transpose):
fn = get_stats_function_for_weights(name, PERCHANNEL)
if name in ['quantile', 'abs_quantile']:
result = fn(INPUT, q=1e-2, transpose=transpose)
else:
result = fn(INPUT, transpose=transpose)
expected = GOLD_VALUES_CH_TRANS_WEIGHT_FUNCTIONS[name]
np.testing.assert_almost_equal(result, expected)
NUM_LEVELS_PARAMS = [
(np.random.randint, (0, 2, (3, 100, 100)), 1, 1),
(np.random.randint, (-32, 32, (3, 100, 100)), 64, 1),
(np.random.randint, (-32, 32, (3, 100, 100)), 64, 1/512),
(np.random.rand, (3, 100, 100), -1, 1),
(np.random.randint, (0, 1, (3, 100, 100)), 0, 1)
]
@pytest.mark.parametrize('gen_func,params,expected,coef', NUM_LEVELS_PARAMS)
def test_get_num_levels_function(gen_func, params, expected, coef):
test_1 = gen_func(*params) * coef
result = get_num_levels(test_1)
assert result == expected