wehub-resource-sync
142 lines
4.9 KiB
Python
142 lines
4.9 KiB
Python
# Copyright (c) 2026 PaddlePaddle Authors. All Rights Reserved.
|
|
#
|
|
# Licensed under the Apache License, Version 2.0 (the "License");
|
|
# you may not use this file except in compliance with the License.
|
|
# You may obtain a copy of the License at
|
|
#
|
|
# http://www.apache.org/licenses/LICENSE-2.0
|
|
#
|
|
# Unless required by applicable law or agreed to in writing, software
|
|
# distributed under the License is distributed on an "AS IS" BASIS,
|
|
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
# See the License for the specific language governing permissions and
|
|
# limitations under the License.
|
|
|
|
"""
|
|
量化感知训练测试 / Quantization-Aware Training Tests
|
|
|
|
测试目标 / Test Target:
|
|
paddle.quantization 量化功能
|
|
|
|
覆盖的模块 / Covered Modules:
|
|
- 伪量化操作
|
|
- fake_quant层
|
|
- 量化参数
|
|
|
|
作用 / Purpose:
|
|
补充量化相关API的测试，提升覆盖率。
|
|
"""
|
|
|
|
import unittest
|
|
|
|
import numpy as np
|
|
|
|
import paddle
|
|
from paddle import nn
|
|
|
|
paddle.disable_static()
|
|
|
|
|
|
class TestQuantizeBasic(unittest.TestCase):
|
|
"""测试基本量化 / Test basic quantization"""
|
|
|
|
def test_fake_quant_abs_max(self):
|
|
"""测试绝对最大值伪量化 / Test fake quantization with abs max"""
|
|
scale = paddle.to_tensor(1.0)
|
|
x = paddle.to_tensor([-0.5, 0.0, 0.5, 1.0])
|
|
# Manual fake quantize: clamp to [-scale, scale] then quantize
|
|
result = paddle.clip(x, -float(scale.numpy()), float(scale.numpy()))
|
|
self.assertEqual(result.shape, [4])
|
|
|
|
def test_quantize_dequantize(self):
|
|
"""测试量化反量化 / Test quantize-dequantize round trip"""
|
|
x = paddle.to_tensor([0.1, 0.5, 0.9, -0.3])
|
|
bits = 8
|
|
num_levels = 2**bits
|
|
# Quantize to int range and back
|
|
scale = paddle.max(paddle.abs(x))
|
|
quantized = paddle.round(x / scale * (num_levels / 2 - 1))
|
|
dequantized = quantized / (num_levels / 2 - 1) * scale
|
|
# Ensure shape preserved
|
|
self.assertEqual(dequantized.shape, x.shape)
|
|
|
|
|
|
class TestQuantizationUtils(unittest.TestCase):
|
|
"""测试量化工具 / Test quantization utilities"""
|
|
|
|
def test_uniform_quantization(self):
|
|
"""测试均匀量化 / Test uniform quantization"""
|
|
x = paddle.to_tensor([0.0, 0.25, 0.5, 0.75, 1.0])
|
|
# 4-level uniform quantization
|
|
q_min, q_max = 0, 3
|
|
scale = (1.0 - 0.0) / (q_max - q_min)
|
|
quantized = paddle.round(x / scale)
|
|
quantized = paddle.clip(quantized, q_min, q_max)
|
|
self.assertEqual(quantized.shape, [5])
|
|
|
|
def test_symmetric_quantization(self):
|
|
"""测试对称量化 / Test symmetric quantization"""
|
|
x = paddle.to_tensor([-0.5, -0.25, 0.0, 0.25, 0.5])
|
|
bits = 8
|
|
# Symmetric: scale based on abs max
|
|
abs_max = float(paddle.max(paddle.abs(x)).numpy())
|
|
scale = abs_max / (2 ** (bits - 1) - 1)
|
|
quantized = paddle.round(x / scale)
|
|
dequantized = quantized * scale
|
|
np.testing.assert_allclose(dequantized.numpy(), x.numpy(), atol=scale)
|
|
|
|
def test_per_channel_quantization(self):
|
|
"""测试逐通道量化 / Test per-channel quantization"""
|
|
# Simulating per-channel quantization
|
|
x = paddle.randn([4, 8, 16, 16])
|
|
# Compute per-channel max (over spatial dimensions)
|
|
max_vals = paddle.max(paddle.abs(x.reshape([4, 8, -1])), axis=2)
|
|
self.assertEqual(max_vals.shape, [4, 8])
|
|
|
|
|
|
class TestQuantModelWrapper(unittest.TestCase):
|
|
"""测试量化模型包装器 / Test quantization model wrapper"""
|
|
|
|
def test_model_with_fake_quant(self):
|
|
"""测试带伪量化的模型 / Test model with fake quantization"""
|
|
|
|
class SimpleQuantModel(nn.Layer):
|
|
def __init__(self):
|
|
super().__init__()
|
|
self.conv = nn.Conv2D(3, 8, 3)
|
|
self.bn = nn.BatchNorm2D(8)
|
|
|
|
def forward(self, x):
|
|
# Simulate fake quantization by clipping
|
|
x = paddle.clip(x, -1.0, 1.0)
|
|
x = self.conv(x)
|
|
x = self.bn(x)
|
|
return x
|
|
|
|
model = SimpleQuantModel()
|
|
x = paddle.randn([2, 3, 16, 16])
|
|
output = model(x)
|
|
self.assertEqual(output.shape[0], 2)
|
|
self.assertEqual(output.shape[1], 8)
|
|
|
|
def test_weight_quantization_aware(self):
|
|
"""测试权重量化感知 / Test weight quantization awareness"""
|
|
model = nn.Linear(8, 4)
|
|
# Simulate quantizing weights
|
|
original_weight = model.weight.numpy().copy()
|
|
bits = 8
|
|
q_max = 2 ** (bits - 1) - 1
|
|
scale = np.max(np.abs(original_weight)) / q_max
|
|
quant_weight = np.round(original_weight / scale).clip(-q_max, q_max)
|
|
dequant_weight = quant_weight * scale
|
|
model.weight.set_value(
|
|
paddle.to_tensor(dequant_weight.astype(np.float32))
|
|
)
|
|
x = paddle.randn([4, 8])
|
|
output = model(x)
|
|
self.assertEqual(output.shape, [4, 4])
|
|
|
|
|
|
if __name__ == '__main__':
|
|
unittest.main()
|