Skip to content

P
PaddleDetection

PaddlePaddle / PaddleDetection
大约 1 年 前同步成功

通知 695
Star 11112
Fork 2696
P
PaddleDetection
提交 0ea122f0 编写于 作者: B Bai Yifan 提交者： whs
浏览文件
操作

Copy slim from release/1.6 to develop (#3758)

上级 3126a437
隐藏空白更改
内联 并排
Showing with 790 addition and 141 deletion
slim/distillation/README.md 0 → 100755
浏览文件 @ 0ea122f0
>运行该示例前请安装Paddle1.6或更高版本
# 检测模型蒸馏示例
## 概述
该示例使用PaddleSlim提供的[蒸馏策略](https://github.com/PaddlePaddle/models/blob/develop/PaddleSlim/docs/tutorial.md#3-蒸馏)对检测库中的模型进行蒸馏训练。
在阅读该示例前，建议您先了解以下内容：
- [检测库的常规训练方法](https://github.com/PaddlePaddle/models/tree/develop/PaddleCV/PaddleDetection)
- [PaddleSlim使用文档](https://github.com/PaddlePaddle/models/blob/develop/PaddleSlim/docs/usage.md)
## 配置文件说明
关于配置文件如何编写您可以参考：
- [PaddleSlim配置文件编写说明](https://github.com/PaddlePaddle/models/blob/develop/PaddleSlim/docs/usage.md#122-%E9%85%8D%E7%BD%AE%E6%96%87%E4%BB%B6%E7%9A%84%E4%BD%BF%E7%94%A8)
- [蒸馏策略配置文件编写说明](https://github.com/PaddlePaddle/models/blob/develop/PaddleSlim/docs/usage.md#23-蒸馏)
这里以ResNet34-YoloV3蒸馏MobileNetV1-YoloV3模型为例，首先，为了对`student model`和`teacher model`有个总体的认识，从而进一步确认蒸馏的对象，我们通过以下命令分别观察两个网络变量（Variable）的名称和形状：
```python
# 观察student model的Variable
for v in fluid.default_main_program().list_vars():
if "py_reader" not in v.name and "double_buffer" not in v.name and "generated_var" not in v.name:
print(v.name, v.shape)
# 观察teacher model的Variable
for v in teacher_program.list_vars():
print(v.name, v.shape)
```
经过对比可以发现，`student model`和`teacher model`的部分中间结果分别为：
```bash
# student model
conv2d_15.tmp_0
# teacher model
teacher_teacher_conv2d_1.tmp_0
```
所以，我们用`l2_distiller`对这两个特征图做蒸馏。在配置文件中进行如下配置：
```yaml
distillers:
l2_distiller:
class: 'L2Distiller'
teacher_feature_map: 'teacher_teacher_conv2d_1.tmp_0'
student_feature_map: 'conv2d_15.tmp_0'
distillation_loss_weight: 1
strategies:
distillation_strategy:
class: 'DistillationStrategy'
distillers: ['l2_distiller']
start_epoch: 0
end_epoch: 270
```
我们也可以根据上述操作为蒸馏策略选择其他loss，PaddleSlim支持的有`FSP_loss`, `L2_loss`和`softmax_with_cross_entropy_loss` 。
## 训练
根据[PaddleDetection/tools/train.py](https://github.com/PaddlePaddle/models/blob/develop/PaddleCV/PaddleDetection/tools/train.py)编写压缩脚本compress.py。
在该脚本中定义了Compressor对象，用于执行压缩任务。
您可以通过运行脚本`run.sh`运行该示例。
### 保存断点（checkpoint）
如果在配置文件中设置了`checkpoint_path`, 则在蒸馏任务执行过程中会自动保存断点，当任务异常中断时，
重启任务会自动从`checkpoint_path`路径下按数字顺序加载最新的checkpoint文件。如果不想让重启的任务从断点恢复，
需要修改配置文件中的`checkpoint_path`，或者将`checkpoint_path`路径下文件清空。
>注意：配置文件中的信息不会保存在断点中，重启前对配置文件的修改将会生效。
## 评估
如果在配置文件中设置了`checkpoint_path`，则每个epoch会保存一个压缩后的用于评估的模型，
该模型会保存在`${checkpoint_path}/${epoch_id}/eval_model/`路径下，包含`__model__`和`__params__`两个文件。
其中，`__model__`用于保存模型结构信息，`__params__`用于保存参数（parameters）信息。
如果不需要保存评估模型，可以在定义Compressor对象时，将`save_eval_model`选项设置为False（默认为True）。
运行命令为：
```
python ../eval.py \
--model_path ${checkpoint_path}/${epoch_id}/eval_model/ \
--model_name __model__ \
--params_name __params__ \
-c ../../configs/yolov3_mobilenet_v1_voc.yml \
-d "../../dataset/voc"
```
## 预测
如果在配置文件中设置了`checkpoint_path`，并且在定义Compressor对象时指定了`prune_infer_model`选项，则每个epoch都会
保存一个`inference model`。该模型是通过删除eval_program中多余的operators而得到的。
该模型会保存在`${checkpoint_path}/${epoch_id}/eval_model/`路径下，包含`__model__.infer`和`__params__`两个文件。
其中，`__model__.infer`用于保存模型结构信息，`__params__`用于保存参数（parameters）信息。
更多关于`prune_infer_model`选项的介绍，请参考：[Compressor介绍](https://github.com/PaddlePaddle/models/blob/develop/PaddleSlim/docs/usage.md#121-%E5%A6%82%E4%BD%95%E6%94%B9%E5%86%99%E6%99%AE%E9%80%9A%E8%AE%AD%E7%BB%83%E8%84%9A%E6%9C%AC)
### python预测
在脚本<a href="../infer.py">slim/infer.py</a>中展示了如何使用fluid python API加载使用预测模型进行预测。
运行命令为：
```
python ../infer.py \
--model_path ${checkpoint_path}/${epoch_id}/eval_model/ \
--model_name __model__.infer \
--params_name __params__ \
-c ../../configs/yolov3_mobilenet_v1_voc.yml \
--infer_dir ../../demo
```
### PaddleLite
该示例中产出的预测（inference）模型可以直接用PaddleLite进行加载使用。
关于PaddleLite如何使用，请参考：[PaddleLite使用文档](https://github.com/PaddlePaddle/Paddle-Lite/wiki#%E4%BD%BF%E7%94%A8)
## 示例结果
>当前release的结果并非超参调优后的最好结果，仅做示例参考，后续我们会优化当前结果。
### MobileNetV1-YOLO-V3
| FLOPS |Box AP|
|---|---|
|baseline|76.2 |
|蒸馏后|- |
## FAQ
slim/distillation/compress.py 0 → 100644
浏览文件 @ 0ea122f0
# Copyright (c) 2019 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 time
import multiprocessing
import numpy as np
from collections import deque, OrderedDict
from paddle.fluid.contrib.slim.core import Compressor
from paddle.fluid.framework import IrGraph
def set_paddle_flags(**kwargs):
for key, value in kwargs.items():
if os.environ.get(key, None) is None:
os.environ[key] = str(value)
# NOTE(paddle-dev): All of these flags should be set before
# `import paddle`. Otherwise, it would not take any effect.
set_paddle_flags(
FLAGS_eager_delete_tensor_gb=0, # enable GC to save memory
)
from paddle import fluid
import sys
sys.path.append("../../")
from ppdet.core.workspace import load_config, merge_config, create
from ppdet.data.data_feed import create_reader
from ppdet.utils.eval_utils import parse_fetches, eval_results
from ppdet.utils.stats import TrainingStats
from ppdet.utils.cli import ArgsParser
from ppdet.utils.check import check_gpu
import ppdet.utils.checkpoint as checkpoint
from ppdet.modeling.model_input import create_feed
import logging
FORMAT = '%(asctime)s-%(levelname)s: %(message)s'
logging.basicConfig(level=logging.INFO, format=FORMAT)
logger = logging.getLogger(__name__)
def eval_run(exe, compile_program, reader, keys, values, cls, test_feed):
"""
Run evaluation program, return program outputs.
"""
iter_id = 0
results = []
if len(cls) != 0:
values = []
for i in range(len(cls)):
_, accum_map = cls[i].get_map_var()
cls[i].reset(exe)
values.append(accum_map)
images_num = 0
start_time = time.time()
has_bbox = 'bbox' in keys
for data in reader():
data = test_feed.feed(data)
feed_data = {'image': data['image'], 'im_size': data['im_size']}
outs = exe.run(compile_program,
feed=feed_data,
fetch_list=[values[0]],
return_numpy=False)
outs.append(data['gt_box'])
outs.append(data['gt_label'])
outs.append(data['is_difficult'])
res = {
k: (np.array(v), v.recursive_sequence_lengths())
for k, v in zip(keys, outs)
}
results.append(res)
if iter_id % 100 == 0:
logger.info('Test iter {}'.format(iter_id))
iter_id += 1
images_num += len(res['bbox'][1][0]) if has_bbox else 1
logger.info('Test finish iter {}'.format(iter_id))
end_time = time.time()
fps = images_num / (end_time - start_time)
if has_bbox:
logger.info('Total number of images: {}, inference time: {} fps.'.
format(images_num, fps))
else:
logger.info('Total iteration: {}, inference time: {} batch/s.'.format(
images_num, fps))
return results
def main():
cfg = load_config(FLAGS.config)
if 'architecture' in cfg:
main_arch = cfg.architecture
else:
raise ValueError("'architecture' not specified in config file.")
merge_config(FLAGS.opt)
if 'log_iter' not in cfg:
cfg.log_iter = 20
# check if set use_gpu=True in paddlepaddle cpu version
check_gpu(cfg.use_gpu)
if cfg.use_gpu:
devices_num = fluid.core.get_cuda_device_count()
else:
devices_num = int(
os.environ.get('CPU_NUM', multiprocessing.cpu_count()))
if 'train_feed' not in cfg:
train_feed = create(main_arch + 'TrainFeed')
else:
train_feed = create(cfg.train_feed)
if 'eval_feed' not in cfg:
eval_feed = create(main_arch + 'EvalFeed')
else:
eval_feed = create(cfg.eval_feed)
place = fluid.CUDAPlace(0) if cfg.use_gpu else fluid.CPUPlace()
exe = fluid.Executor(place)
lr_builder = create('LearningRate')
optim_builder = create('OptimizerBuilder')
# build program
model = create(main_arch)
train_loader, train_feed_vars = create_feed(train_feed, iterable=True)
train_fetches = model.train(train_feed_vars)
loss = train_fetches['loss']
lr = lr_builder()
opt = optim_builder(lr)
opt.minimize(loss)
#for v in fluid.default_main_program().list_vars():
# if "py_reader" not in v.name and "double_buffer" not in v.name and "generated_var" not in v.name:
# print(v.name, v.shape)
cfg.max_iters = 258
train_reader = create_reader(train_feed, cfg.max_iters, FLAGS.dataset_dir)
train_loader.set_sample_list_generator(train_reader, place)
exe.run(fluid.default_startup_program())
# parse train fetches
train_keys, train_values, _ = parse_fetches(train_fetches)
train_keys.append('lr')
train_values.append(lr.name)
train_fetch_list = []
for k, v in zip(train_keys, train_values):
train_fetch_list.append((k, v))
print("train_fetch_list: {}".format(train_fetch_list))
eval_prog = fluid.Program()
startup_prog = fluid.Program()
with fluid.program_guard(eval_prog, startup_prog):
with fluid.unique_name.guard():
model = create(main_arch)
_, test_feed_vars = create_feed(eval_feed, iterable=True)
fetches = model.eval(test_feed_vars)
eval_prog = eval_prog.clone(True)
eval_reader = create_reader(eval_feed, args_path=FLAGS.dataset_dir)
test_data_feed = fluid.DataFeeder(test_feed_vars.values(), place)
# parse eval fetches
extra_keys = []
if cfg.metric == 'COCO':
extra_keys = ['im_info', 'im_id', 'im_shape']
if cfg.metric == 'VOC':
extra_keys = ['gt_box', 'gt_label', 'is_difficult']
eval_keys, eval_values, eval_cls = parse_fetches(fetches, eval_prog,
extra_keys)
eval_fetch_list = []
for k, v in zip(eval_keys, eval_values):
eval_fetch_list.append((k, v))
print("eval_fetch_list: {}".format(eval_fetch_list))
exe.run(startup_prog)
checkpoint.load_params(exe,
fluid.default_main_program(), cfg.pretrain_weights)
best_box_ap_list = []
def eval_func(program, scope):
results = eval_run(exe, program, eval_reader, eval_keys, eval_values,
eval_cls, test_data_feed)
resolution = None
is_bbox_normalized = False
if 'mask' in results[0]:
resolution = model.mask_head.resolution
box_ap_stats = eval_results(results, eval_feed, cfg.metric,
cfg.num_classes, resolution,
is_bbox_normalized, FLAGS.output_eval)
if len(best_box_ap_list) == 0:
best_box_ap_list.append(box_ap_stats[0])
elif box_ap_stats[0] > best_box_ap_list[0]:
best_box_ap_list[0] = box_ap_stats[0]
logger.info("Best test box ap: {}".format(best_box_ap_list[0]))
return best_box_ap_list[0]
test_feed = [('image', test_feed_vars['image'].name),
('im_size', test_feed_vars['im_size'].name)]
teacher_cfg = load_config(FLAGS.teacher_config)
teacher_arch = teacher_cfg.architecture
teacher_programs = []
teacher_program = fluid.Program()
teacher_startup_program = fluid.Program()
with fluid.program_guard(teacher_program, teacher_startup_program):
with fluid.unique_name.guard('teacher_'):
teacher_feed_vars = OrderedDict()
for name, var in train_feed_vars.items():
teacher_feed_vars[name] = teacher_program.global_block(
)._clone_variable(
var, force_persistable=False)
model = create(teacher_arch)
train_fetches = model.train(teacher_feed_vars)
#print("="*50+"teacher_model_params"+"="*50)
#for v in teacher_program.list_vars():
# print(v.name, v.shape)
#return
exe.run(teacher_startup_program)
assert FLAGS.teacher_pretrained and os.path.exists(
FLAGS.teacher_pretrained
), "teacher_pretrained should be set when teacher_model is not None."
def if_exist(var):
return os.path.exists(os.path.join(FLAGS.teacher_pretrained, var.name))
fluid.io.load_vars(
exe,
FLAGS.teacher_pretrained,
main_program=teacher_program,
predicate=if_exist)
teacher_programs.append(teacher_program.clone(for_test=True))
com = Compressor(
place,
fluid.global_scope(),
fluid.default_main_program(),
train_reader=train_reader,
train_feed_list=[(key, value.name)
for key, value in train_feed_vars.items()],
train_fetch_list=train_fetch_list,
eval_program=eval_prog,
eval_reader=eval_reader,
eval_feed_list=test_feed,
eval_func={'map': eval_func},
eval_fetch_list=eval_fetch_list[0:1],
save_eval_model=True,
prune_infer_model=[["image", "im_size"], ["multiclass_nms_0.tmp_0"]],
teacher_programs=teacher_programs,
train_optimizer=None,
distiller_optimizer=opt,
log_period=20)
com.config(FLAGS.slim_file)
com.run()
if __name__ == '__main__':
parser = ArgsParser()
parser.add_argument(
"-t",
"--teacher_config",
default=None,
type=str,
help="Config file of teacher architecture.")
parser.add_argument(
"-s",
"--slim_file",
default=None,
type=str,
help="Config file of PaddleSlim.")
parser.add_argument(
"-r",
"--resume_checkpoint",
default=None,
type=str,
help="Checkpoint path for resuming training.")
parser.add_argument(
"--eval",
action='store_true',
default=False,
help="Whether to perform evaluation in train")
parser.add_argument(
"--teacher_pretrained",
default=None,
type=str,
help="Whether to use pretrained model.")
parser.add_argument(
"--output_eval",
default=None,
type=str,
help="Evaluation directory, default is current directory.")
parser.add_argument(
"-d",
"--dataset_dir",
default=None,
type=str,
help="Dataset path, same as DataFeed.dataset.dataset_dir")
FLAGS = parser.parse_args()
main()
slim/distillation/run.sh 0 → 100644
浏览文件 @ 0ea122f0
#!/usr/bin/env bash
# download pretrain model
root_url="https://paddlemodels.bj.bcebos.com/object_detection"
yolov3_r34_voc="yolov3_r34_voc.tar"
pretrain_dir='./pretrain'
if [ ! -d ${pretrain_dir} ]; then
mkdir ${pretrain_dir}
fi
cd ${pretrain_dir}
if [ ! -f ${yolov3_r34_voc} ]; then
wget ${root_url}/${yolov3_r34_voc}
tar xf ${yolov3_r34_voc}
fi
cd -
# enable GC strategy
export FLAGS_fast_eager_deletion_mode=1
export FLAGS_eager_delete_tensor_gb=0.0
# for distillation
#-----------------
export CUDA_VISIBLE_DEVICES=0,1,2,3
# Fixing name conflicts in distillation
cd ${pretrain_dir}/yolov3_r34_voc
for files in $(ls teacher_*)
do mv $files ${files#*_}
done
for files in $(ls *)
do mv $files "teacher_"$files
done
cd -
python -u compress.py \
-c ../../configs/yolov3_mobilenet_v1_voc.yml \
-t yolov3_resnet34.yml \
-s yolov3_mobilenet_v1_yolov3_resnet34_distillation.yml \
-o YoloTrainFeed.batch_size=64 \
-d ../../dataset/voc \
--teacher_pretrained ./pretrain/yolov3_r34_voc \
> yolov3_distallation.log 2>&1 &
tailf yolov3_distallation.log
slim/distillation/yolov3_mobilenet_v1_yolov3_resnet34_distillation.yml 0 → 100644
浏览文件 @ 0ea122f0
version: 1.0
distillers:
l2_distiller:
class: 'L2Distiller'
teacher_feature_map: 'teacher_teacher_conv2d_1.tmp_0'
student_feature_map: 'conv2d_15.tmp_0'
distillation_loss_weight: 1
strategies:
distillation_strategy:
class: 'DistillationStrategy'
distillers: ['l2_distiller']
start_epoch: 0
end_epoch: 270
compressor:
epoch: 271
checkpoint_path: './checkpoints/'
strategies:
- distillation_strategy
slim/quantization/yolov3_mobilenet_v1_voc.yml → slim/distillation/yolov3_resnet34.yml
浏览文件 @ 0ea122f0
architecture: YOLOv3
train_feed: YoloTrainFeed
eval_feed: YoloEvalFeed
test_feed: YoloTestFeed
use_gpu: true
max_iters: 1000
log_smooth_window: 20
save_dir: output
snapshot_iter: 2000
metric: VOC
map_type: 11point
pretrain_weights: https://paddlemodels.bj.bcebos.com/object_detection/yolov3_mobilenet_v1_voc.tar
weights: output/yolov3_mobilenet_v1_voc/model_final
num_classes: 20
weight_prefix_name: teacher_
YOLOv3:
backbone: MobileNet
backbone: ResNet
yolo_head: YOLOv3Head
MobileNet:
ResNet:
norm_type: sync_bn
freeze_at: 0
freeze_norm: false
norm_decay: 0.
conv_group_scale: 1
with_extra_blocks: false
depth: 34
feature_maps: [3, 4, 5]
YOLOv3Head:
anchor_masks: [[6, 7, 8], [3, 4, 5], [0, 1, 2]]
......@@ -38,50 +32,3 @@ YOLOv3Head:
nms_top_k: 1000
normalized: false
score_threshold: 0.01
LearningRate:
base_lr: 0.0001
schedulers:
- !PiecewiseDecay
gamma: 0.1
milestones:
- 1000
- 2000
#- !LinearWarmup
#start_factor: 0.
#steps: 1000
OptimizerBuilder:
optimizer:
momentum: 0.9
type: Momentum
regularizer:
factor: 0.0005
type: L2
YoloTrainFeed:
batch_size: 8
dataset:
dataset_dir: ../../dataset/voc
annotation: VOCdevkit/VOC_all/ImageSets/Main/train.txt
image_dir: VOCdevkit/VOC_all/JPEGImages
use_default_label: true
num_workers: 8
bufsize: 128
use_process: true
mixup_epoch: 250
YoloEvalFeed:
batch_size: 8
image_shape: [3, 608, 608]
dataset:
dataset_dir: ../../dataset/voc
annotation: VOCdevkit/VOC_all/ImageSets/Main/val.txt
image_dir: VOCdevkit/VOC_all/JPEGImages
use_default_label: true
YoloTestFeed:
batch_size: 1
image_shape: [3, 608, 608]
dataset:
use_default_label: true
slim/infer.py
浏览文件 @ 0ea122f0
......@@ -19,11 +19,13 @@ from __future__ import print_function
import os
import sys
import glob
import time
import numpy as np
from PIL import Image
sys.path.append("../../")
def set_paddle_flags(**kwargs):
for key, value in kwargs.items():
if os.environ.get(key, None) is None:
......@@ -117,20 +119,19 @@ def main():
test_images = get_test_images(FLAGS.infer_dir, FLAGS.infer_img)
test_feed.dataset.add_images(test_images)
place = fluid.CUDAPlace(0) if cfg.use_gpu else fluid.CPUPlace()
exe = fluid.Executor(place)
infer_prog, feed_var_names, fetch_list = fluid.io.load_inference_model(
dirname=FLAGS.model_path, model_filename=FLAGS.model_name,
params_filename=FLAGS.params_name,
executor=exe)
dirname=FLAGS.model_path,
model_filename=FLAGS.model_name,
params_filename=FLAGS.params_name,
executor=exe)
reader = create_reader(test_feed)
feeder = fluid.DataFeeder(place=place, feed_list=feed_var_names,
program=infer_prog)
feeder = fluid.DataFeeder(
place=place, feed_list=feed_var_names, program=infer_prog)
# parse infer fetches
assert cfg.metric in ['COCO', 'VOC'], \
......@@ -140,7 +141,9 @@ def main():
extra_keys = ['im_info', 'im_id', 'im_shape']
if cfg['metric'] == 'VOC':
extra_keys = ['im_id', 'im_shape']
keys, values, _ = parse_fetches({'bbox':fetch_list}, infer_prog, extra_keys)
keys, values, _ = parse_fetches({
'bbox': fetch_list
}, infer_prog, extra_keys)
# parse dataset category
if cfg.metric == 'COCO':
......@@ -166,9 +169,33 @@ def main():
imid2path = reader.imid2path
keys = ['bbox']
infer_time = True
compile_prog = fluid.compiler.CompiledProgram(infer_prog)
for iter_id, data in enumerate(reader()):
feed_data = [[d[0], d[1]] for d in data]
outs = exe.run(infer_prog,
# for infer time
if infer_time:
warmup_times = 10
repeats_time = 100
feed_data_dict = feeder.feed(feed_data)
for i in range(warmup_times):
exe.run(compile_prog,
feed=feed_data_dict,
fetch_list=fetch_list,
return_numpy=False)
start_time = time.time()
for i in range(repeats_time):
exe.run(compile_prog,
feed=feed_data_dict,
fetch_list=fetch_list,
return_numpy=False)
print("infer time: {} ms/sample".format((time.time() - start_time) *
1000 / repeats_time))
infer_time = False
outs = exe.run(compile_prog,
feed=feeder.feed(feed_data),
fetch_list=fetch_list,
return_numpy=False)
......@@ -258,10 +285,7 @@ if __name__ == '__main__':
default="tb_log_dir/image",
help='Tensorboard logging directory for image.')
parser.add_argument(
'--model_path',
type=str,
default=None,
help="inference model path")
'--model_path', type=str, default=None, help="inference model path")
parser.add_argument(
'--model_name',
type=str,
......
slim/prune/README.md
浏览文件 @ 0ea122f0
......@@ -7,7 +7,8 @@
该示例使用PaddleSlim提供的[卷积通道剪裁压缩策略](https://github.com/PaddlePaddle/models/blob/develop/PaddleSlim/docs/tutorial.md#2-%E5%8D%B7%E7%A7%AF%E6%A0%B8%E5%89%AA%E8%A3%81%E5%8E%9F%E7%90%86)对检测库中的模型进行压缩。
在阅读该示例前，建议您先了解以下内容：
- <a href="../..README_cn.md">检测库的常规训练方法</a>
- <a href="../../README_cn.md">检测库的常规训练方法</a>
- [检测模型数据准备](https://github.com/PaddlePaddle/models/blob/develop/PaddleCV/PaddleDetection/docs/INSTALL_cn.md#%E6%95%B0%E6%8D%AE%E9%9B%86)
- [PaddleSlim使用文档](https://github.com/PaddlePaddle/models/blob/develop/PaddleSlim/docs/usage.md)
......@@ -29,7 +30,7 @@ from paddle.fluid.framework import IrGraph
from paddle.fluid import core
graph = IrGraph(core.Graph(train_prog.desc), for_test=True)
marked_nodes = set()
marked_nodes = set()
for op in graph.all_op_nodes():
print(op.name())
if op.name().find('conv') > -1:
......@@ -39,12 +40,12 @@ graph.draw('.', 'forward', marked_nodes)
该示例中MobileNetV1-YoloV3模型结构的可视化结果：<a href="./images/MobileNetV1-YoloV3.pdf">MobileNetV1-YoloV3.pdf</a>
同时通过以下命令观察目标卷积层的参数（parameters）的名称和shape:
同时通过以下命令观察目标卷积层的参数（parameters）的名称和shape:
```
for param in fluid.default_main_program().global_block().all_parameters():
if 'weights' in param.name:
print param.name, param.shape
print(param.name, param.shape)
```
......@@ -109,6 +110,7 @@ python compress.py \
-s yolov3_mobilenet_v1_slim.yaml \
-c ../../configs/yolov3_mobilenet_v1_voc.yml \
-o max_iters=258 \
YoloTrainFeed.batch_size=64 \
-d "../../dataset/voc"
```
......@@ -117,9 +119,9 @@ python compress.py \
如果要调整训练卡数，需要调整配置文件`yolov3_mobilenet_v1_voc.yml`中的以下参数：
- **max_iters:** 一个`epoch`中batch的数量，需要设置为`total_num / batch_size`, 其中`total_num`为训练样本总数量，`batch_size`为多卡上总的batch size.
- **YoloTrainFeed.batch_size:** 单张卡上的batch size, 受限于显存大小。
- **YoloTrainFeed.batch_size:** 当使用DataLoader时，表示单张卡上的batch size; 当使用普通reader时，则表示多卡上的总的`batch_size`。`batch_size`受限于显存大小。
- **LeaningRate.base_lr:** 根据多卡的总`batch_size`调整`base_lr`，两者大小正相关，可以简单的按比例进行调整。
- **LearningRate.schedulers.PiecewiseDecay.milestones：**请根据batch size的变化对其调整。
- **LearningRate.schedulers.PiecewiseDecay.milestones：** 请根据batch size的变化对其调整。
- **LearningRate.schedulers.PiecewiseDecay.LinearWarmup.steps：** 请根据batch size的变化对其进行调整。
......@@ -130,7 +132,7 @@ python compress.py \
-s yolov3_mobilenet_v1_slim.yaml \
-c ../../configs/yolov3_mobilenet_v1_voc.yml \
-o max_iters=258 \
-o YoloTrainFeed.batch_size = 16 \
YoloTrainFeed.batch_size=64 \
-d "../../dataset/voc"
```
......@@ -140,9 +142,9 @@ python compress.py \
-s yolov3_mobilenet_v1_slim.yaml \
-c ../../configs/yolov3_mobilenet_v1_voc.yml \
-o max_iters=516 \
-o LeaningRate.base_lr=0.005 \ # 0.001 /2
-o YoloTrainFeed.batch_size = 16 \
-o LearningRate.schedulers='[!PiecewiseDecay {gamma: 0.1, milestones: [110000, 124000]}, !LinearWarmup {start_factor: 0., steps: 2000}]' \
LeaningRate.base_lr=0.005 \
YoloTrainFeed.batch_size=32 \
LearningRate.schedulers='[!PiecewiseDecay {gamma: 0.1, milestones: [110000, 124000]}, !LinearWarmup {start_factor: 0., steps: 2000}]' \
-d "../../dataset/voc"
```
......@@ -166,6 +168,16 @@ python compress.py \
如果不需要保存评估模型，可以在定义Compressor对象时，将`save_eval_model`选项设置为False（默认为True）。
运行命令为：
```
python ../eval.py \
--model_path ${checkpoint_path}/${epoch_id}/eval_model/ \
--model_name __model__ \
--params_name __params__ \
-c ../../configs/yolov3_mobilenet_v1_voc.yml \
-d "../../dataset/voc"
```
## 预测
如果在配置文件中设置了`checkpoint_path`，并且在定义Compressor对象时指定了`prune_infer_model`选项，则每个epoch都会
......@@ -180,6 +192,16 @@ python compress.py \
在脚本<a href="../infer.py">PaddleDetection/tools/infer.py</a>中展示了如何使用fluid python API加载使用预测模型进行预测。
运行命令为：
```
python ../infer.py \
--model_path ${checkpoint_path}/${epoch_id}/eval_model/ \
--model_name __model__.infer \
--params_name __params__ \
-c ../../configs/yolov3_mobilenet_v1_voc.yml \
--infer_dir ../../demo
```
### PaddleLite
该示例中产出的预测（inference）模型可以直接用PaddleLite进行加载使用。
......@@ -187,13 +209,13 @@ python compress.py \
## 示例结果
> 当前release的结果并非超参调优后的最好结果，仅做示例参考，后续我们会优化当前结果。
### MobileNetV1-YOLO-V3
| FLOPS |top1_acc/top5_acc| model_size |Paddle Fluid inference time(ms)| Paddle Lite inference time(ms)|
| FLOPS |Box AP| model_size |Paddle Fluid inference time(ms)| Paddle Lite inference time(ms)|
|---|---|---|---|---|
|baseline|- |- |- |-|
|-10%|- |- |- |-|
|-30%|- |- |- |-|
|-50%|- |- |- |-|
|baseline|76.2 |93M |- |-|
|-50%|69.48 |51M |- |-|
## FAQ
slim/prune/compress.py
浏览文件 @ 0ea122f0
......@@ -24,11 +24,13 @@ import sys
sys.path.append("../../")
from paddle.fluid.contrib.slim import Compressor
def set_paddle_flags(**kwargs):
for key, value in kwargs.items():
if os.environ.get(key, None) is None:
os.environ[key] = str(value)
# NOTE(paddle-dev): All of these flags should be set before
# `import paddle`. Otherwise, it would not take any effect.
set_paddle_flags(
......@@ -48,6 +50,8 @@ import logging
FORMAT = '%(asctime)s-%(levelname)s: %(message)s'
logging.basicConfig(level=logging.INFO, format=FORMAT)
logger = logging.getLogger(__name__)
def eval_run(exe, compile_program, reader, keys, values, cls, test_feed):
"""
Run evaluation program, return program outputs.
......@@ -66,8 +70,7 @@ def eval_run(exe, compile_program, reader, keys, values, cls, test_feed):
has_bbox = 'bbox' in keys
for data in reader():
data = test_feed.feed(data)
feed_data = {'image': data['image'],
'im_size': data['im_size']}
feed_data = {'image': data['image'], 'im_size': data['im_size']}
outs = exe.run(compile_program,
feed=feed_data,
fetch_list=[values[0]],
......@@ -147,9 +150,7 @@ def main():
optimizer = optim_builder(lr)
optimizer.minimize(loss)
train_reader = create_reader(train_feed, cfg.max_iters * devices_num,
FLAGS.dataset_dir)
train_reader = create_reader(train_feed, cfg.max_iters, FLAGS.dataset_dir)
train_loader.set_sample_list_generator(train_reader, place)
# parse train fetches
......@@ -157,7 +158,7 @@ def main():
train_keys.append("lr")
train_values.append(lr.name)
train_fetch_list=[]
train_fetch_list = []
for k, v in zip(train_keys, train_values):
train_fetch_list.append((k, v))
......@@ -181,8 +182,8 @@ def main():
if cfg.metric == 'VOC':
extra_keys = ['gt_box', 'gt_label', 'is_difficult']
eval_keys, eval_values, eval_cls = parse_fetches(fetches, eval_prog,
extra_keys)
eval_fetch_list=[]
extra_keys)
eval_fetch_list = []
for k, v in zip(eval_keys, eval_values):
eval_fetch_list.append((k, v))
......@@ -195,21 +196,20 @@ def main():
#place = fluid.CPUPlace()
#exe = fluid.Executor(place)
results = eval_run(exe, program, eval_reader,
eval_keys, eval_values, eval_cls, test_data_feed)
results = eval_run(exe, program, eval_reader, eval_keys, eval_values,
eval_cls, test_data_feed)
resolution = None
if 'mask' in results[0]:
resolution = model.mask_head.resolution
box_ap_stats = eval_results(results, eval_feed, cfg.metric, cfg.num_classes,
resolution, False, FLAGS.output_eval)
box_ap_stats = eval_results(results, eval_feed, cfg.metric,
cfg.num_classes, resolution, False,
FLAGS.output_eval)
if len(best_box_ap_list) == 0:
best_box_ap_list.append(box_ap_stats[0])
elif box_ap_stats[0] > best_box_ap_list[0]:
best_box_ap_list[0] = box_ap_stats[0]
checkpoint.save(exe, train_prog, os.path.join(save_dir,"best_model"))
logger.info("Best test box ap: {}".format(
best_box_ap_list[0]))
logger.info("Best test box ap: {}".format(best_box_ap_list[0]))
return best_box_ap_list[0]
test_feed = [('image', test_feed_vars['image'].name),
......@@ -228,13 +228,12 @@ def main():
eval_func={'map': eval_func},
eval_fetch_list=[eval_fetch_list[0]],
save_eval_model=True,
prune_infer_model=[["image", "im_size"],["multiclass_nms_0.tmp_0"]],
prune_infer_model=[["image", "im_size"], ["multiclass_nms_0.tmp_0"]],
train_optimizer=None)
com.config(FLAGS.slim_file)
com.run()
if __name__ == '__main__':
parser = ArgsParser()
parser.add_argument(
......
slim/quantization/README.md
浏览文件 @ 0ea122f0
......@@ -37,28 +37,76 @@
- config: 检测库的配置，其中配置了训练超参数、数据集信息等。
- slim_file: PaddleSlim的配置文件，参见[配置文件说明](#配置文件说明)。
您可以通过运行以下命令运行该示例，请确保已正确下载[pretrained model](https://github.com/PaddlePaddle/models/tree/develop/PaddleCV/image_classification#%E5%B7%B2%E5%8F%91%E5%B8%83%E6%A8%A1%E5%9E%8B%E5%8F%8A%E5%85%B6%E6%80%A7%E8%83%BD)。
您可以通过运行以下命令运行该示例。
step1: 开启显存优化策略
step1: 设置gpu卡
```
export FLAGS_fast_eager_deletion_mode=1
export FLAGS_eager_delete_tensor_gb=0.0
export CUDA_VISIBLE_DEVICES=0
```
step2: 设置gpu卡,目前的超参设置适合2卡训练
step2: 开始训练
使用PaddleDetection提供的配置文件在用8卡进行训练：
```
python compress.py \
-s yolov3_mobilenet_v1_slim.yaml \
-c ../../configs/yolov3_mobilenet_v1_voc.yml \
-d "../../dataset/voc" \
-o max_iters=258 \
LearningRate.base_lr=0.0001 \
LearningRate.schedulers="[!PiecewiseDecay {gamma: 0.1, milestones: [258, 516]}]" \
pretrain_weights=https://paddlemodels.bj.bcebos.com/object_detection/yolov3_mobilenet_v1_voc.tar \
YoloTrainFeed.batch_size=64
```
>通过命令行覆盖设置max_iters选项，因为PaddleDetection中训练是以`batch`为单位迭代的，并没有涉及`epoch`的概念，但是PaddleSlim需要知道当前训练进行到第几个`epoch`, 所以需要将`max_iters`设置为一个`epoch`内的`batch`的数量。
如果要调整训练卡数，需要调整配置文件`yolov3_mobilenet_v1_voc.yml`中的以下参数：
- **max_iters:** 一个`epoch`中batch的数量，需要设置为`total_num / batch_size`, 其中`total_num`为训练样本总数量，`batch_size`为多卡上总的batch size.
- **YoloTrainFeed.batch_size:** 当使用DataLoader时，表示单张卡上的batch size; 当使用普通reader时，则表示多卡上的总的batch_size。batch_size受限于显存大小。
- **LeaningRate.base_lr:** 根据多卡的总`batch_size`调整`base_lr`，两者大小正相关，可以简单的按比例进行调整。
- **LearningRate.schedulers.PiecewiseDecay.milestones：** 请根据batch size的变化对其调整。
- **LearningRate.schedulers.PiecewiseDecay.LinearWarmup.steps：** 请根据batch size的变化对其进行调整。
以下为4卡训练示例，通过命令行覆盖`yolov3_mobilenet_v1_voc.yml`中的参数：
```
export CUDA_VISIBLE_DEVICES=0,1
python compress.py \
-s yolov3_mobilenet_v1_slim.yaml \
-c ../../configs/yolov3_mobilenet_v1_voc.yml \
-d "../../dataset/voc" \
-o max_iters=258 \
LearningRate.base_lr=0.0001 \
LearningRate.schedulers="[!PiecewiseDecay {gamma: 0.1, milestones: [258, 516]}]" \
pretrain_weights=https://paddlemodels.bj.bcebos.com/object_detection/yolov3_mobilenet_v1_voc.tar \
YoloTrainFeed.batch_size=64
```
step3: 开始训练
以下为2卡训练示例，受显存所制，单卡`batch_size`不变, 总`batch_size`减小，`base_lr`减小，一个epoch内batch数量增加，同时需要调整学习率相关参数，如下：
```
python compress.py \
-s yolov3_mobilenet_v1_slim.yaml \
-c yolov3_mobilenet_v1_voc.yml
-c ../../configs/yolov3_mobilenet_v1_voc.yml \
-d "../../dataset/voc" \
-o max_iters=516 \
LearningRate.base_lr=0.00005 \
LearningRate.schedulers="[!PiecewiseDecay {gamma: 0.1, milestones: [516, 1012]}]" \
pretrain_weights=https://paddlemodels.bj.bcebos.com/object_detection/yolov3_mobilenet_v1_voc.tar \
YoloTrainFeed.batch_size=32
```
通过`python compress.py --help`查看可配置参数。
通过`python ../../tools/configure.py ${option_name} help`查看如何通过命令行覆盖配置文件`yolov3_mobilenet_v1_voc.yml`中的参数。
### 训练时的模型结构
这部分介绍来源于[量化low-level API介绍](https://github.com/PaddlePaddle/models/tree/develop/PaddleSlim/quant_low_level_api#1-%E9%87%8F%E5%8C%96%E8%AE%AD%E7%BB%83low-level-apis%E4%BB%8B%E7%BB%8D)。
PaddlePaddle框架中有四个和量化相关的IrPass, 分别是QuantizationTransformPass、QuantizationFreezePass、ConvertToInt8Pass以及TransformForMobilePass。在训练时，对网络应用了QuantizationTransformPass，作用是在网络中的conv2d、depthwise_conv2d、mul等算子的各个输入前插入连续的量化op和反量化op，并改变相应反向算子的某些输入。示例图如下：
PaddlePaddle框架中和量化相关的IrPass, 分别有QuantizationTransformPass、QuantizationFreezePass、ConvertToInt8Pass。在训练时，对网络应用了QuantizationTransformPass，作用是在网络中的conv2d、depthwise_conv2d、mul等算子的各个输入前插入连续的量化op和反量化op，并改变相应反向算子的某些输入。示例图如下：
<p align="center">
<img src="./images/TransformPass.png" height=400 width=520 hspace='10'/> <br />
......@@ -76,10 +124,10 @@ PaddlePaddle框架中有四个和量化相关的IrPass, 分别是QuantizationTra
### 保存评估和预测模型
如果在配置文件的量化策略中设置了`float_model_save_path`, `int8_model_save_path`, `mobile_model_save_path`, 在训练结束后，会保存模型量化压缩之后用于预测的模型。接下来介绍这三种预测模型的区别。
如果在配置文件的量化策略中设置了`float_model_save_path`, `int8_model_save_path` 在训练结束后，会保存模型量化压缩之后用于预测的模型。接下来介绍这2种预测模型的区别。
#### FP32模型
在介绍量化训练时的模型结构时介绍了PaddlePaddle框架中有四个和量化相关的IrPass, 分别是QuantizationTransformPass、QuantizationFreezePass、ConvertToInt8Pass以及TransformForMobilePass。FP32模型是在应用QuantizationFreezePass并删除eval_program中多余的operators之后，保存的模型。
在介绍量化训练时的模型结构时介绍了PaddlePaddle框架中和量化相关的IrPass, 分别是QuantizationTransformPass、QuantizationFreezePass、ConvertToInt8Pass。FP32模型是在应用QuantizationFreezePass并删除eval_program中多余的operators之后，保存的模型。
QuantizationFreezePass主要用于改变IrGraph中量化op和反量化op的顺序，即将类似图1中的量化op和反量化op顺序改变为图2中的布局。除此之外，QuantizationFreezePass还会将`conv2d`、`depthwise_conv2d`、`mul`等算子的权重离线量化为int8_t范围内的值(但数据类型仍为float32)，以减少预测过程中对权重的量化操作，示例如图2：
......@@ -97,19 +145,13 @@ QuantizationFreezePass主要用于改变IrGraph中量化op和反量化op的顺
<strong>图3：应用ConvertToInt8Pass后的结果</strong>
</p>
#### mobile模型
经TransformForMobilePass转换后，用户可得到兼容[paddle-lite](https://github.com/PaddlePaddle/Paddle-Lite)移动端预测库的量化模型。paddle-mobile中的量化op和反量化op的名称分别为`quantize`和`dequantize`。`quantize`算子和PaddlePaddle框架中的`fake_quantize_abs_max`算子簇的功能类似，`dequantize` 算子和PaddlePaddle框架中的`fake_dequantize_max_abs`算子簇的功能相同。若选择paddle-mobile执行量化训练输出的模型，则需要将`fake_quantize_abs_max`等算子改为`quantize`算子以及将`fake_dequantize_max_abs`等算子改为`dequantize`算子，示例如图4：
<p align="center">
<img src="./images/TransformForMobilePass.png" height=400 width=400 hspace='10'/> <br />
<strong>图4：应用TransformForMobilePass后的结果</strong>
</p>
> 综上，可得在量化过程中有以下几种模型结构：
1. 原始模型
2. 经QuantizationTransformPass之后得到的适用于训练的量化模型结构，在${checkpoint_path}下保存的`eval_model`是这种结构，在训练过程中每个epoch结束时也使用这个网络结构进行评估，虽然这个模型结构不是最终想要的模型结构，但是每个epoch的评估结果可用来挑选模型。
3. 经QuantizationFreezePass之后得到的FP32模型结构，具体结构已在上面进行介绍。本文档中列出的数据集的评估结果是对FP32模型结构进行评估得到的结果。这种模型结构在训练过程中只会保存一次，也就是在量化配置文件中设置的`end_epoch`结束时进行保存，如果想将其他epoch的训练结果转化成FP32模型，可使用脚本 <a href='./freeze.py'>PaddleSlim/classification/quantization/freeze.py</a>进行转化，具体使用方法在[评估](#评估)中介绍。
4. 经ConvertToInt8Pass之后得到的8-bit模型结构，具体结构已在上面进行介绍。这种模型结构在训练过程中只会保存一次，也就是在量化配置文件中设置的`end_epoch`结束时进行保存，如果想将其他epoch的训练结果转化成8-bit模型，可使用脚本 <a href='./freeze.py'>slim/quantization/freeze.py</a>进行转化，具体使用方法在[评估](#评估)中介绍。
5. 经TransformForMobilePass之后得到的mobile模型结构，具体结构已在上面进行介绍。这种模型结构在训练过程中只会保存一次，也就是在量化配置文件中设置的`end_epoch`结束时进行保存，如果想将其他epoch的训练结果转化成mobile模型，可使用脚本 <a href='./freeze.py'>slim/quantization/freeze.py</a>进行转化，具体使用方法在[评估](#评估)中介绍。
## 评估
......@@ -128,21 +170,24 @@ python ../eval.py \
--model_path ${checkpoint_path}/${epoch_id}/eval_model/ \
--model_name __model__ \
--params_name __params__ \
-c yolov3_mobilenet_v1_voc.yml
-c ../../configs/yolov3_mobilenet_v1_voc.yml \
-d "../../dataset/voc"
```
在评估之后，选取效果最好的epoch的模型，可使用脚本 <a href='./freeze.py'>slim/quantization/freeze.py</a>将该模型转化为以上介绍的三种模型：FP32模型，int8模型，mobile模型，需要配置的参数为：
在评估之后，选取效果最好的epoch的模型，可使用脚本 <a href='./freeze.py'>slim/quantization/freeze.py</a>将该模型转化为以上介绍的2种模型：FP32模型，int8模型，需要配置的参数为：
- model_path, 加载的模型路径，`为${checkpoint_path}/${epoch_id}/eval_model/`
- weight_quant_type 模型参数的量化方式，和配置文件中的类型保持一致
- save_path `FP32`, `8-bit`, `mobile`模型的保存路径，分别为 `${save_path}/float/`, `${save_path}/int8/`, `${save_path}/mobile/`
- save_path `FP32`, `8-bit` 模型的保存路径，分别为 `${save_path}/float/`, `${save_path}/int8/`
运行命令示例：
```
python freeze.py \
--model_path ${checkpoint_path}/${epoch_id}/eval_model/ \
--weight_quant_type ${weight_quant_type} \
--save_path ${any path you want}
--save_path ${any path you want} \
-c ../../configs/yolov3_mobilenet_v1_voc.yml \
-d "../../dataset/voc"
```
### 最终评估模型
......@@ -150,10 +195,11 @@ python freeze.py \
运行命令为：
```
python ../eval.py \
--model_path ${float_model_path}
--model_path ${float_model_path}
--model_name model \
--params_name weights \
-c yolov3_mobilenet_v1_voc.yml
-c ../../configs/yolov3_mobilenet_v1_voc.yml \
-d "../../dataset/voc"
```
## 预测
......@@ -169,7 +215,7 @@ python ../infer.py \
--model_path ${save_path}/float \
--model_name model \
--params_name weights \
-c yolov3_mobilenet_v1_voc.yml \
-c ../../configs/yolov3_mobilenet_v1_voc.yml \
--infer_dir ../../demo
```
......@@ -180,7 +226,9 @@ FP32模型可使用PaddleLite进行加载预测，可参见教程[Paddle-Lite如
## 示例结果
### MobileNetV1
>当前release的结果并非超参调优后的最好结果，仅做示例参考，后续我们会优化当前结果。
### MobileNetV1-YOLO-V3
| weight量化方式 | activation量化方式| Box ap |Paddle Fluid inference time(ms)| Paddle Lite inference time(ms)|
|---|---|---|---|---|
......@@ -189,9 +237,5 @@ FP32模型可使用PaddleLite进行加载预测，可参见教程[Paddle-Lite如
|abs_max|moving_average_abs_max|- |- |-|
|channel_wise_abs_max|abs_max|- |- |-|
>训练超参：
## FAQ
slim/quantization/compress.py
浏览文件 @ 0ea122f0
......@@ -28,11 +28,13 @@ from paddle.fluid.contrib.slim import Compressor
from paddle.fluid.framework import IrGraph
from paddle.fluid import core
def set_paddle_flags(**kwargs):
for key, value in kwargs.items():
if os.environ.get(key, None) is None:
os.environ[key] = str(value)
# NOTE(paddle-dev): All of these flags should be set before
# `import paddle`. Otherwise, it would not take any effect.
set_paddle_flags(
......@@ -46,7 +48,7 @@ from ppdet.data.data_feed import create_reader
from ppdet.utils.eval_utils import parse_fetches, eval_results
from ppdet.utils.stats import TrainingStats
from ppdet.utils.cli import ArgsParser
from ppdet.utils.cli import ArgsParser, print_total_cfg
from ppdet.utils.check import check_gpu, check_version
import ppdet.utils.checkpoint as checkpoint
from ppdet.modeling.model_input import create_feed
......@@ -55,6 +57,8 @@ import logging
FORMAT = '%(asctime)s-%(levelname)s: %(message)s'
logging.basicConfig(level=logging.INFO, format=FORMAT)
logger = logging.getLogger(__name__)
def eval_run(exe, compile_program, reader, keys, values, cls, test_feed):
"""
Run evaluation program, return program outputs.
......@@ -73,11 +77,10 @@ def eval_run(exe, compile_program, reader, keys, values, cls, test_feed):
has_bbox = 'bbox' in keys
for data in reader():
data = test_feed.feed(data)
feed_data = {'image': data['image'],
'im_size': data['im_size']}
feed_data = {'image': data['image'], 'im_size': data['im_size']}
outs = exe.run(compile_program,
feed=feed_data,
fetch_list=values[0],
fetch_list=[values[0]],
return_numpy=False)
outs.append(data['gt_box'])
outs.append(data['gt_label'])
......@@ -118,8 +121,8 @@ def main():
# check if set use_gpu=True in paddlepaddle cpu version
check_gpu(cfg.use_gpu)
# print_total_cfg(cfg)
#check_version()
if cfg.use_gpu:
devices_num = fluid.core.get_cuda_device_count()
else:
......@@ -155,16 +158,14 @@ def main():
optimizer = optim_builder(lr)
optimizer.minimize(loss)
train_reader = create_reader(train_feed, cfg.max_iters * devices_num,
FLAGS.dataset_dir)
train_reader = create_reader(train_feed, cfg.max_iters, FLAGS.dataset_dir)
train_loader.set_sample_list_generator(train_reader, place)
# parse train fetches
train_keys, train_values, _ = parse_fetches(train_fetches)
train_values.append(lr)
train_fetch_list=[]
train_fetch_list = []
for k, v in zip(train_keys, train_values):
train_fetch_list.append((k, v))
print("train_fetch_list: {}".format(train_fetch_list))
......@@ -188,14 +189,13 @@ def main():
if cfg.metric == 'VOC':
extra_keys = ['gt_box', 'gt_label', 'is_difficult']
eval_keys, eval_values, eval_cls = parse_fetches(fetches, eval_prog,
extra_keys)
extra_keys)
# print(eval_values)
eval_fetch_list=[]
eval_fetch_list = []
for k, v in zip(eval_keys, eval_values):
eval_fetch_list.append((k, v))
exe.run(startup_prog)
start_iter = 0
......@@ -208,21 +208,20 @@ def main():
#place = fluid.CPUPlace()
#exe = fluid.Executor(place)
results = eval_run(exe, program, eval_reader,
eval_keys, eval_values, eval_cls, test_data_feed)
results = eval_run(exe, program, eval_reader, eval_keys, eval_values,
eval_cls, test_data_feed)
resolution = None
if 'mask' in results[0]:
resolution = model.mask_head.resolution
box_ap_stats = eval_results(results, eval_feed, cfg.metric, cfg.num_classes,
resolution, False, FLAGS.output_eval)
box_ap_stats = eval_results(results, eval_feed, cfg.metric,
cfg.num_classes, resolution, False,
FLAGS.output_eval)
if len(best_box_ap_list) == 0:
best_box_ap_list.append(box_ap_stats[0])
elif box_ap_stats[0] > best_box_ap_list[0]:
best_box_ap_list[0] = box_ap_stats[0]
checkpoint.save(exe, train_prog, os.path.join(save_dir,"best_model"))
logger.info("Best test box ap: {}".format(
best_box_ap_list[0]))
logger.info("Best test box ap: {}".format(best_box_ap_list[0]))
return best_box_ap_list[0]
test_feed = [('image', test_feed_vars['image'].name),
......@@ -240,12 +239,12 @@ def main():
eval_feed_list=test_feed,
eval_func={'map': eval_func},
eval_fetch_list=[eval_fetch_list[0]],
prune_infer_model=[["image", "im_size"], ["multiclass_nms_0.tmp_0"]],
train_optimizer=None)
com.config(FLAGS.slim_file)
com.run()
if __name__ == '__main__':
parser = ArgsParser()
parser.add_argument(
......
slim/quantization/freeze.py
浏览文件 @ 0ea122f0
......@@ -32,11 +32,13 @@ from paddle.fluid.contrib.slim.quantization import QuantizationFreezePass
from paddle.fluid.contrib.slim.quantization import ConvertToInt8Pass
from paddle.fluid.contrib.slim.quantization import TransformForMobilePass
def set_paddle_flags(**kwargs):
for key, value in kwargs.items():
if os.environ.get(key, None) is None:
os.environ[key] = str(value)
# NOTE(paddle-dev): All of these flags should be set before
# `import paddle`. Otherwise, it would not take any effect.
set_paddle_flags(
......@@ -59,6 +61,8 @@ import logging
FORMAT = '%(asctime)s-%(levelname)s: %(message)s'
logging.basicConfig(level=logging.INFO, format=FORMAT)
logger = logging.getLogger(__name__)
def eval_run(exe, compile_program, reader, keys, values, cls, test_feed):
"""
Run evaluation program, return program outputs.
......@@ -71,8 +75,7 @@ def eval_run(exe, compile_program, reader, keys, values, cls, test_feed):
has_bbox = 'bbox' in keys
for data in reader():
data = test_feed.feed(data)
feed_data = {'image': data['image'],
'im_size': data['im_size']}
feed_data = {'image': data['image'], 'im_size': data['im_size']}
outs = exe.run(compile_program,
feed=feed_data,
fetch_list=values[0],
......@@ -123,7 +126,6 @@ def main():
devices_num = int(
os.environ.get('CPU_NUM', multiprocessing.cpu_count()))
if 'eval_feed' not in cfg:
eval_feed = create(main_arch + 'EvalFeed')
else:
......@@ -138,85 +140,79 @@ def main():
#eval_pyreader.decorate_sample_list_generator(eval_reader, place)
test_data_feed = fluid.DataFeeder(test_feed_vars.values(), place)
assert os.path.exists(FLAGS.model_path)
infer_prog, feed_names, fetch_targets = fluid.io.load_inference_model(
dirname=FLAGS.model_path, executor=exe,
model_filename='__model__',
params_filename='__params__')
dirname=FLAGS.model_path,
executor=exe,
model_filename='__model__.infer',
params_filename='__params__')
eval_keys = ['bbox', 'gt_box', 'gt_label', 'is_difficult']
eval_values = ['multiclass_nms_0.tmp_0', 'gt_box', 'gt_label', 'is_difficult']
eval_values = [
'multiclass_nms_0.tmp_0', 'gt_box', 'gt_label', 'is_difficult'
]
eval_cls = []
eval_values[0] = fetch_targets[0]
results = eval_run(exe, infer_prog, eval_reader,
eval_keys, eval_values, eval_cls, test_data_feed)
results = eval_run(exe, infer_prog, eval_reader, eval_keys, eval_values,
eval_cls, test_data_feed)
resolution = None
if 'mask' in results[0]:
resolution = model.mask_head.resolution
box_ap_stats = eval_results(results, eval_feed, cfg.metric, cfg.num_classes,
resolution, False, FLAGS.output_eval)
resolution, False, FLAGS.output_eval)
logger.info("freeze the graph for inference")
test_graph = IrGraph(core.Graph(infer_prog.desc), for_test=True)
freeze_pass = QuantizationFreezePass(
scope=fluid.global_scope(),
place=place,
weight_quantize_type=FLAGS.weight_quant_type)
scope=fluid.global_scope(),
place=place,
weight_quantize_type=FLAGS.weight_quant_type)
freeze_pass.apply(test_graph)
server_program = test_graph.to_program()
fluid.io.save_inference_model(
dirname=os.path.join(FLAGS.save_path, 'float'),
feeded_var_names=feed_names,
target_vars=fetch_targets,
executor=exe,
main_program=server_program,
model_filename='model',
params_filename='weights')
dirname=os.path.join(FLAGS.save_path, 'float'),
feeded_var_names=feed_names,
target_vars=fetch_targets,
executor=exe,
main_program=server_program,
model_filename='model',
params_filename='weights')
logger.info("convert the weights into int8 type")
convert_int8_pass = ConvertToInt8Pass(
scope=fluid.global_scope(),
place=place)
scope=fluid.global_scope(), place=place)
convert_int8_pass.apply(test_graph)
server_int8_program = test_graph.to_program()
fluid.io.save_inference_model(
dirname=os.path.join(FLAGS.save_path, 'int8'),
feeded_var_names=feed_names,
target_vars=fetch_targets,
executor=exe,
main_program=server_int8_program,
model_filename='model',
params_filename='weights')
dirname=os.path.join(FLAGS.save_path, 'int8'),
feeded_var_names=feed_names,
target_vars=fetch_targets,
executor=exe,
main_program=server_int8_program,
model_filename='model',
params_filename='weights')
logger.info("convert the freezed pass to paddle-lite execution")
mobile_pass = TransformForMobilePass()
mobile_pass.apply(test_graph)
mobile_program = test_graph.to_program()
fluid.io.save_inference_model(
dirname=os.path.join(FLAGS.save_path, 'mobile'),
feeded_var_names=feed_names,
target_vars=fetch_targets,
executor=exe,
main_program=mobile_program,
model_filename='model',
params_filename='weights')
dirname=os.path.join(FLAGS.save_path, 'mobile'),
feeded_var_names=feed_names,
target_vars=fetch_targets,
executor=exe,
main_program=mobile_program,
model_filename='model',
params_filename='weights')
if __name__ == '__main__':
parser = ArgsParser()
parser.add_argument(
"-m",
"--model_path",
default=None,
type=str,
help="path of checkpoint")
"-m", "--model_path", default=None, type=str, help="path of checkpoint")
parser.add_argument(
"--output_eval",
default=None,
......
Markdown is supported
0% 或 .
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 或