mirror of
https://github.com/zebrajr/pytorch.git
synced 2025-12-06 12:20:52 +01:00
13ad4739a6
### Summary: This PR implements PTQ for APoT FakeQuant. It runs models (Resnet-18 pre-trained model, ImageNet dataset) to compare accuracy metrics for different qconfig settings of uniform vs. APoT quantized activation and weight. According to the collected accuracy stats, model #2 (uniform activation and APoT weight) appears to have a slight improvement in accuracy compared to model #1 (uniform activation and uniform weight) for 8-bit and significant improvement for 4-bit (see "Accuracy Stats" section below). ### Test Plan: Run models with: `python test/quantization/core/experimental/fx_graph_mode_apot.py` ### Accuracy Stats: 8-bit (Uniform int8, APoT b = 8 k = 2) **Model #1:** Uniform activation, uniform weight (FX Graph Mode quantized) Evaluation accuracy on test dataset: 64.43% (Top-1), 85.62% (Top-5) **Model #2:** Uniform activation, APoT weight (FX Graph Mode quantized) Evaluation accuracy on test dataset: 64.51% (Top-1), 85.78% (Top-5) **Model #3:** APoT activation, APoT weight (FX Graph Mode quantized) Evaluation accuracy on test dataset: 64.32% (Top-1), 85.78% (Top-5) 4-bit (Uniform int4, APoT b = 4 k = 2) **Model #1:** Uniform activation, uniform weight (FX Graph Mode quantized) Evaluation accuracy on test dataset: 45.63% (Top-1), 71.96% (Top-5) **Model #2:** Uniform activation, APoT weight (FX Graph Mode quantized) Evaluation accuracy on test dataset: 64.24% (Top-1), 85.56% (Top-5) **Model #3:** APoT activation, APoT weight (FX Graph Mode quantized) Evaluation accuracy on test dataset: 45.40% (Top-1), 76.21% (Top-5) **Full Precision model (FX Graph Mode quantized)** Evaluation accuracy on test dataset: 69.76% (Top-1), 89.08% (Top-5) **Eager mode quantized model** Evaluation accuracy on test dataset: 69.49% (Top-1), 88.90% (Top-5) Pull Request resolved: https://github.com/pytorch/pytorch/pull/81040 Approved by: https://github.com/jerryzh168
47 lines
2.5 KiB
Python
47 lines
2.5 KiB
Python
import torch
|
|
from torch.ao.quantization.qconfig import QConfig
|
|
from torch.ao.quantization import MinMaxObserver
|
|
from torch.ao.quantization.fake_quantize import FakeQuantize
|
|
from torch.ao.quantization.experimental.fake_quantize import APoTFakeQuantize
|
|
|
|
"""
|
|
Default symmetric fake_quant for activations.
|
|
"""
|
|
default_symmetric_fake_quant = FakeQuantize.with_args(observer=MinMaxObserver,
|
|
qscheme=torch.per_tensor_symmetric,
|
|
dtype=torch.quint8)
|
|
|
|
"""
|
|
Default symmetric fake_quant for weights.
|
|
"""
|
|
default_weight_symmetric_fake_quant = FakeQuantize.with_args(observer=MinMaxObserver,
|
|
qscheme=torch.per_tensor_symmetric,
|
|
dtype=torch.qint8)
|
|
|
|
# uniform activation and weight, b=8 k=2
|
|
uniform_qconfig_8bit = QConfig(activation=default_symmetric_fake_quant,
|
|
weight=default_weight_symmetric_fake_quant.with_args)
|
|
|
|
# uniform activation, APoT weight, b=8 k=2
|
|
apot_weight_qconfig_8bit = QConfig(activation=default_symmetric_fake_quant.with_args,
|
|
weight=APoTFakeQuantize.with_args(b=8, k=2, dtype=torch.qint8))
|
|
|
|
# APoT activation and uniform weight, b=8 k=2
|
|
apot_qconfig_8bit = QConfig(activation=APoTFakeQuantize.with_args(b=8, k=2, dtype=torch.quint8),
|
|
weight=APoTFakeQuantize.with_args(b=8, k=2, dtype=torch.qint8))
|
|
|
|
# uniform activation and weight, b=4 k=2
|
|
uniform_qconfig_4bit = QConfig(activation=default_symmetric_fake_quant.with_args(quant_min=0,
|
|
quant_max=15),
|
|
weight=default_weight_symmetric_fake_quant.with_args(quant_min=0,
|
|
quant_max=15))
|
|
|
|
# uniform activation, APoT weight, b=4 k=2
|
|
apot_weight_qconfig_4bit = QConfig(activation=default_symmetric_fake_quant.with_args(quant_min=0,
|
|
quant_max=15),
|
|
weight=APoTFakeQuantize.with_args(b=4, k=2, dtype=torch.qint8))
|
|
|
|
# APoT activation and uniform weight, b=4 k=2
|
|
apot_qconfig_4bit = QConfig(activation=APoTFakeQuantize.with_args(b=4, k=2, dtype=torch.quint8),
|
|
weight=APoTFakeQuantize.with_args(b=4, k=2, dtype=torch.qint8))
|