未验证 提交 90797a22 编写于 作者: W whs 提交者: GitHub

Add tutorial of PTQ for classification (#1705)

上级 5000cf3d
# 动态图离线量化
本示例介绍如何对动态图模型进行离线量化,示例以常用的MobileNetV1和MobileNetV3模型为例,介绍如何对其进行离线量化。
## 分类模型的离线量化流程
#### 准备数据
在当前目录下创建``data``文件夹,将``ImageNet``的验证集解压在``data``文件夹下,解压后``data/ILSVRC2012``文件夹下应包含以下文件:
- ``'val'``文件夹,验证图片
- ``'val_list.txt'``文件
#### 准备需要离线量化的模型
本示例直接使用[paddle vision](https://github.com/PaddlePaddle/Paddle/tree/develop/python/paddle/vision/models)内置的模型结构和预训练权重。通过以下命令查看支持的所有模型:
```
python ptq.py --help
```
## 启动命令
以MobileNetV1为例,通过以下脚本启动PTQ任务:
```bash
python ptq.py \
--data=dataset/ILSVRC2012/ \
--model=mobilenet_v1 \
--activation_observer='mse' \
--weight_observer='mse_channel_wise' \
--quant_batch_num=10 \
--quant_batch_size=10 \
--output_dir="output_ptq"
```
其中,通过 `activation_observer` 配置用于激活的量化算法,通过 `weight_observer` 配置用于权重的量化算法。
更多支持的量化算法,请执行以下命令查看:
```
python ptq.py --help
```
## 评估精度
执行以下命令,使用 PaddleInference 推理库测试推理精度:
```bash
python eval.py --model_path=output_ptq/mobilenet_v1/int8_infer/ --data_dir=dataset/ILSVRC2012/ --use_gpu=True
```
- 评估时支持CPU,并且不依赖TensorRT,MKLDNN。
## 量化结果
| 模型 | FP32模型准确率(Top1/Top5) | 量化方法(activation/weight) | 量化模型准确率(Top1/Top5) |
| ----------- | --------------------------- | ------------ | --------------------------- |
| MobileNetV1 | 70.10%/90.10% | mse / mes_channel_wise | 69.10%/89.80% |
| MobileNetV2 | 71.10%/90.90% | mse / mes_channel_wise | 70.70%/90.10% |
# Copyright (c) 2021 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.
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import os
import numpy as np
import time
import sys
import argparse
import math
import paddle
import paddle.inference as paddle_infer
from ptq import ImageNetValDataset
def eval():
# create predictor
model_file = os.path.join(FLAGS.model_path, FLAGS.model_filename)
params_file = os.path.join(FLAGS.model_path, FLAGS.params_filename)
config = paddle_infer.Config(model_file, params_file)
if FLAGS.use_gpu:
config.enable_use_gpu(1000, 0)
if not FLAGS.ir_optim:
config.switch_ir_optim(False)
predictor = paddle_infer.create_predictor(config)
input_names = predictor.get_input_names()
input_handle = predictor.get_input_handle(input_names[0])
output_names = predictor.get_output_names()
output_handle = predictor.get_output_handle(output_names[0])
# prepare data
val_dataset = ImageNetValDataset(FLAGS.data_dir)
eval_loader = paddle.io.DataLoader(
val_dataset, batch_size=FLAGS.batch_size, num_workers=5)
cost_time = 0.
total_num = 0.
correct_1_num = 0
correct_5_num = 0
for batch_id, data in enumerate(eval_loader()):
# set input
img_np = np.array([tensor.numpy() for tensor in data[0]])
label_np = np.array([tensor.numpy() for tensor in data[1]])
input_handle.reshape(img_np.shape)
input_handle.copy_from_cpu(img_np)
# run
t1 = time.time()
predictor.run()
t2 = time.time()
cost_time += (t2 - t1)
output_data = output_handle.copy_to_cpu()
# calculate accuracy
for i in range(len(label_np)):
label = label_np[i][0]
result = output_data[i, :]
index = result.argsort()
total_num += 1
if index[-1] == label:
correct_1_num += 1
if label in index[-5:]:
correct_5_num += 1
if batch_id % 10 == 0:
acc1 = correct_1_num / total_num
acc5 = correct_5_num / total_num
avg_time = cost_time / total_num
print(
"batch_id {}, acc1 {:.3f}, acc5 {:.3f}, avg time {:.5f} sec/img".
format(batch_id, acc1, acc5, avg_time))
if FLAGS.test_samples > 0 and \
(batch_id + 1)* FLAGS.batch_size >= FLAGS.test_samples:
break
acc1 = correct_1_num / total_num
acc5 = correct_5_num / total_num
avg_time = cost_time / total_num
print("End test: test image {}".format(total_num))
print("test_acc1: {:.4f}; test_acc5: {:.4f}; avg time: {:.5f} sec/img".
format(acc1, acc5, avg_time))
print("\n")
if __name__ == '__main__':
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument(
'--model_path', type=str, default="", help="The inference model path.")
parser.add_argument(
'--model_filename',
type=str,
default="model.pdmodel",
help="model filename")
parser.add_argument(
'--params_filename',
type=str,
default="model.pdiparams",
help="params filename")
parser.add_argument(
'--data_dir',
type=str,
default="dataset/ILSVRC2012/",
help="The ImageNet dataset root dir.")
parser.add_argument(
'--test_samples',
type=int,
default=-1,
help="Test samples. If set -1, use all test samples")
parser.add_argument(
'--batch_size', type=int, default=10, help="Batch size.")
parser.add_argument(
'--use_gpu', type=bool, default=False, help=" Whether use gpu or not.")
parser.add_argument(
'--ir_optim', type=bool, default=False, help="Enable ir optim.")
FLAGS = parser.parse_args()
eval()
# Copyright (c) 2021 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.
from __future__ import division
from __future__ import print_function
import argparse
import six
from inspect import isfunction
import os
import time
import random
from types import FunctionType
from typing import Dict
import numpy as np
from PIL import Image
import paddle
from paddle.io import Dataset
from paddle.vision.transforms import transforms
import paddle.vision.models as models
from paddle.quantization import QuantConfig
from paddle.quantization import PTQ
from tqdm import tqdm
from paddleslim.quant.observers import HistObserver, KLObserver, EMDObserver, MSEObserver, AVGObserver
from paddleslim.quant.observers import MSEChannelWiseWeightObserver, AbsMaxChannelWiseWeightObserver
import sys
sys.path.append(os.path.dirname("__file__"))
sys.path.append(
os.path.join(os.path.dirname("__file__"), os.path.pardir, os.path.pardir))
SUPPORT_MODELS: Dict[str, FunctionType] = {}
for _name, _module in models.__dict__.items():
if isfunction(_module) and 'pretrained' in _module.__code__.co_varnames:
SUPPORT_MODELS[_name] = _module
ACTIVATION_OBSERVERS: Dict[str, type] = {
'hist': HistObserver,
'kl': KLObserver,
'emd': EMDObserver,
'mse': MSEObserver,
'avg': AVGObserver,
}
WEIGHT_OBSERVERS: Dict[str, type] = {
'mse_channel_wise': MSEChannelWiseWeightObserver,
'abs_max_channel_wise': AbsMaxChannelWiseWeightObserver,
}
class ImageNetValDataset(Dataset):
def __init__(self, data_dir, image_size=224, resize_short_size=256):
super(ImageNetValDataset, self).__init__()
val_file_list = os.path.join(data_dir, 'val_list.txt')
test_file_list = os.path.join(data_dir, 'test_list.txt')
self.data_dir = data_dir
normalize = transforms.Normalize(
mean=[123.675, 116.28, 103.53], std=[58.395, 57.120, 57.375])
self.transform = transforms.Compose([
transforms.Resize(resize_short_size),
transforms.CenterCrop(image_size),
transforms.Transpose(), normalize
])
with open(val_file_list) as flist:
lines = [line.strip() for line in flist]
self.data = [line.split() for line in lines]
def __getitem__(self, index):
img_path, label = self.data[index]
img_path = os.path.join(self.data_dir, img_path)
img = Image.open(img_path).convert('RGB')
label = np.array([label]).astype(np.int64)
return self.transform(img), label
def __len__(self):
return len(self.data)
def test(net, dataset):
valid_loader = paddle.io.DataLoader(dataset, batch_size=1)
net.eval()
batch_id = 0
acc_top1_ns = []
acc_top5_ns = []
eval_reader_cost = 0.0
eval_run_cost = 0.0
total_samples = 0
reader_start = time.time()
for data in tqdm(valid_loader()):
eval_reader_cost += time.time() - reader_start
image = data[0]
label = data[1]
eval_start = time.time()
out = net(image)
acc_top1 = paddle.metric.accuracy(input=out, label=label, k=1)
acc_top5 = paddle.metric.accuracy(input=out, label=label, k=5)
eval_run_cost += time.time() - eval_start
batch_size = image.shape[0]
total_samples += batch_size
acc_top1_ns.append(np.mean(acc_top1.numpy()))
acc_top5_ns.append(np.mean(acc_top5.numpy()))
batch_id += 1
reader_start = time.time()
return np.mean(np.array(acc_top1_ns)), np.mean(np.array(acc_top5_ns))
def calibrate(model, dataset, batch_num, batch_size, num_workers=1):
data_loader = paddle.io.DataLoader(
dataset, batch_size=batch_size, num_workers=num_workers)
pbar = tqdm(total=batch_num)
for idx, data in enumerate(data_loader()):
model(data[0])
pbar.update(1)
if (batch_num > 0) and (idx + 1 >= batch_num):
break
pbar.close()
def main():
num_workers = 5
if FLAGS.ce_test:
# set seed
seed = 111
paddle.seed(seed)
np.random.seed(seed)
random.seed(seed)
num_workers = 0
# 1 load model
fp32_model = SUPPORT_MODELS[FLAGS.model](pretrained=True)
if FLAGS.pretrain_weight:
info_dict = paddle.load(FLAGS.pretrain_weight)
fp32_model.load_dict(info_dict)
print('Finish loading model weights:{}'.format(FLAGS.pretrain_weight))
fp32_model.eval()
val_dataset = ImageNetValDataset(FLAGS.data)
# 2 quantizations
activation_observer = ACTIVATION_OBSERVERS[FLAGS.activation_observer]()
weight_observer = WEIGHT_OBSERVERS[FLAGS.weight_observer]()
config = QuantConfig(weight=None, activation=None)
config.add_type_config(
paddle.nn.Conv2D,
activation=activation_observer,
weight=weight_observer)
ptq = PTQ(config)
top1, top5 = test(fp32_model, val_dataset)
print(
f"\033[31mBaseline(FP32): top1/top5 = {top1*100:.2f}%/{top5*100:.2f}%\033[0m"
)
quant_model = ptq.quantize(fp32_model)
print("Start PTQ calibration for quantization")
calibrate(
quant_model,
val_dataset,
FLAGS.quant_batch_num,
FLAGS.quant_batch_size,
num_workers=num_workers)
infer_model = ptq.convert(quant_model, inplace=True)
top1, top5 = test(infer_model, val_dataset)
print(
f"\033[31mPTQ with {FLAGS.activation_observer}/{FLAGS.weight_observer}: top1/top5 = {top1*100:.2f}%/{top5*100:.2f}%\033[0m"
)
dummy_input = paddle.static.InputSpec(
shape=[None, 3, 224, 224], dtype='float32')
paddle.jit.save(infer_model, "./int8_infer", [dummy_input])
if __name__ == '__main__':
parser = argparse.ArgumentParser("Quantization on ImageNet")
# model
parser.add_argument(
"--model",
type=str,
choices=SUPPORT_MODELS.keys(),
default='mobilenet_v1',
help="model name", )
parser.add_argument(
"--pretrain_weight",
type=str,
default=None,
help="pretrain weight path")
parser.add_argument(
"--output_dir", type=str, default='output', help="save dir")
# data
parser.add_argument(
'--data',
default="/dataset/ILSVRC2012",
help=
'path to dataset (should have subdirectories named "train" and "val"',
required=True, )
parser.add_argument(
'--val_dir',
default="val",
help='the dir that saves val images for paddle.Model')
# quantization
parser.add_argument(
"--activation_observer",
default='mse',
type=str,
choices=ACTIVATION_OBSERVERS.keys(),
help="batch num for quant")
parser.add_argument(
"--weight_observer",
default='mse_channel_wise',
choices=WEIGHT_OBSERVERS.keys(),
type=str,
help="batch size for quant")
# train
parser.add_argument(
"--quant_batch_num", default=10, type=int, help="batch num for quant")
parser.add_argument(
"--quant_batch_size", default=10, type=int, help="batch size for quant")
parser.add_argument(
'--ce_test', default=False, type=bool, help="Whether to CE test.")
FLAGS = parser.parse_args()
print("----------- Configuration Arguments -----------")
for arg, value in sorted(six.iteritems(vars(FLAGS))):
print("%s: %s" % (arg, value))
print("------------------------------------------------")
main()
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册