Piotr KrzemińskiandMaxim Vafin
6d8dcb059d
* Support GetAttr with packed params * Apply suggestions from code review * [PT FE] Add quantized types as normal types to decoder * [PT FE] Add decoder dequantize, add dtypes to quantize * [PT FE] Add dequantize example * [PT FE] Implement replacer for quantized nodes * [PT FE] Register replacer for quantize/dequantize * [PT FE] Remove unwanted junk from previous version * [PT FE] Fix building mistakes for frontend * [PT FE] Clang fix * [PT FE] Ease of use upgrade to quantize funcs * [PT FE] Clang format * [PT FE] Introduce new version of quantize/dequantize * [PT FE] Remove unwanted files from new version * [PT FE] Fix style * [PT FE] Add QuantizedPtNode replacer, fix accuracy error * [PT FE] Add improved version of quantize/dequantize with shared_ptrs * [PT FE] Fix utils shared ptr reference error * [PT FE] Quantize now takes correct input for operations * [PT FE] Upgrade quantize method * [PT FE] Add BFS for dequantize, add quantize_per_channel * [PT FE] Add missing replacer to frontend, improve tests * [PT FE] Rename replacer -> remover, remove unwanted header files * [PT FE] Change function declarations to return ov::Output instead of shared ptr * [PT FE] Add missing context mark node * [PT FE] Remove unknown modifications to ie_c_api * [PT FE] Remove fp16 support, turn off int32 tests * [PT FE] Clang format * [PT FE] Fix quantize_per_tensor * [PT FE] Minor fixes from review * [PT FE] Remove dequantize, remove helpers, replacer now removes nodes instead * [PT FE] Rename Replacer to Remover for dequantize nodes * [PT FE] Clang format * [PT FE] Move comments to header files, minor import fixes * [PT FE] Fix clang format * [PT FE] Fix dtype issue * [PT FE] Fix quantize_per_channel tests * Apply suggestions from code review Removing sporadic tests from precommit * Apply suggestions from code review * [PT FE] Fix conversion errors for aten::quantize_per_channel * [PT FE] Mark axis 2 as xfail --------- Co-authored-by: Maxim Vafin <maxim.vafin@intel.com>
96 lines
3.9 KiB
Python
96 lines
3.9 KiB
Python
# Copyright (C) 2018-2023 Intel Corporation
|
|
# SPDX-License-Identifier: Apache-2.0
|
|
|
|
import numpy as np
|
|
import pytest
|
|
import torch
|
|
|
|
from pytorch_layer_test_class import PytorchLayerTest
|
|
|
|
class aten_quantize_per_tensor_aten_dequantize(torch.nn.Module):
|
|
def __init__(self, scale, zero_point, dtype) -> None:
|
|
torch.nn.Module.__init__(self)
|
|
self.scale = scale
|
|
self.zero_point = zero_point
|
|
self.dtype = dtype
|
|
|
|
def forward(self, input_tensor):
|
|
quantized_tensor = torch.quantize_per_tensor(input_tensor, self.scale, self.zero_point, self.dtype)
|
|
dequantized_tensor = torch.dequantize(quantized_tensor)
|
|
return dequantized_tensor
|
|
|
|
class aten_quantize_per_channel_aten_dequantize(torch.nn.Module):
|
|
def __init__(self, scales, zero_points, dtype, axis) -> None:
|
|
torch.nn.Module.__init__(self)
|
|
self.scales = torch.Tensor(scales)
|
|
self.zero_points = torch.Tensor(zero_points)
|
|
self.dtype = dtype
|
|
self.axis = axis
|
|
def forward(self, input_tensor):
|
|
quantized_tensor = torch.quantize_per_channel(input_tensor, self.scales, self.zero_points, self.axis, self.dtype)
|
|
dequantized_tensor = torch.dequantize(quantized_tensor)
|
|
return dequantized_tensor
|
|
|
|
class TestQuantizePerTensorDequantize(PytorchLayerTest):
|
|
def _prepare_input(self):
|
|
return (np.array(5.00 * np.random.rand(100, 100) + 5.00, dtype=np.float32),)
|
|
|
|
@pytest.mark.parametrize("scale", [
|
|
1.0, 0.21, 0.62
|
|
])
|
|
@pytest.mark.parametrize("zero_point", [
|
|
0, 4, -7
|
|
])
|
|
@pytest.mark.parametrize("dtype", [
|
|
torch.quint8,
|
|
torch.qint8,
|
|
pytest.param(torch.qint32, marks=pytest.mark.skip(
|
|
reason="Not supported with FakeQuantize."))
|
|
])
|
|
@pytest.mark.nightly
|
|
# @pytest.mark.precommit - sporadic issue
|
|
def test_quantize_per_tensor_dequantize(self, scale, zero_point, dtype, ie_device, precision, ir_version):
|
|
if dtype == torch.quint8: zero_point = abs(zero_point)
|
|
self._test(aten_quantize_per_tensor_aten_dequantize(scale, zero_point, dtype), None, ["aten::quantize_per_tensor", "aten::dequantize"],
|
|
ie_device, precision, ir_version, )
|
|
|
|
class TestQuantizePerChannelDequantize(PytorchLayerTest):
|
|
def _prepare_input(self):
|
|
return (np.array(5.00 * np.random.rand(5, 6, 7, 8) + 5.00, dtype=np.float32),)
|
|
|
|
@pytest.mark.parametrize("scale, zero_point, axis", [
|
|
[
|
|
np.array([1.0, 0.21, 0.62, 0.5, 0.74], dtype=np.float32),
|
|
np.array([0, -1, 2, -3, 4], dtype=np.int32),
|
|
0
|
|
],
|
|
[
|
|
np.array([1.0, 0.62, 0.74, 0.11, 0.89, 0.32], dtype=np.float32),
|
|
np.array([0, 2, 4, -5, 6, -7], dtype=np.int32),
|
|
1
|
|
],
|
|
pytest.param(
|
|
np.array([1.0, 0.21, 0.62, 0.5, 0.11, 0.89, 0.32], dtype=np.float32),
|
|
np.array([0, -1, 2, -3, 4, -5, -7], dtype=np.int32),
|
|
2,
|
|
marks=pytest.mark.skip(reason="Axis = 2 not supported in FakeQuantize.")),
|
|
[
|
|
np.array([1.0, 0.21, 0.62, 0.5, 0.74, 0.11, 0.89, 0.32], dtype=np.float32),
|
|
np.array([0, -1, 2, -3, 4, -5, 6, -7], dtype=np.int32),
|
|
3
|
|
],
|
|
])
|
|
@pytest.mark.parametrize("dtype", [
|
|
torch.quint8,
|
|
torch.qint8,
|
|
pytest.param(torch.qint32, marks=pytest.mark.skip(
|
|
reason="Not supported with FakeQuantize."))
|
|
])
|
|
@pytest.mark.nightly
|
|
# @pytest.mark.precommit - sporadic issue
|
|
def test_quantize_per_channel_dequantize(self, scale, zero_point, axis, dtype, ie_device, precision, ir_version):
|
|
np.random.shuffle(scale), np.random.shuffle(zero_point)
|
|
if dtype == torch.quint8: zero_point = abs(zero_point)
|
|
self._test(aten_quantize_per_channel_aten_dequantize(scale, zero_point, axis, dtype), None, ["aten::quantize_per_channel", "aten::dequantize"],
|
|
ie_device, precision, ir_version, )
|