Skip to content

P
PaddleClas

PaddlePaddle / PaddleClas
1 年多 前同步成功

通知 116
Star 4999
Fork 1114
P
PaddleClas
未验证 提交 19d03a58 编写于 作者: D dyning 提交者： GitHub
浏览文件
操作

Merge pull request #56 from cuicheng01/master 

Update DataAugment and Distillation doc
上级 590995e0 8d0c458a
隐藏空白更改
内联 并排
Showing with 1019 addition and 7 deletion
configs/DataAugment/ResNet50_AutoAugment.yaml 0 → 100644
浏览文件 @ 19d03a58
mode: 'train'
ARCHITECTURE:
name: 'ResNet50'
pretrained_model: ""
model_save_dir: "./output/"
classes_num: 1000
total_images: 1281167
save_interval: 1
validate: True
valid_interval: 1
epochs: 300
topk: 5
image_shape: [3, 224, 224]
use_mix: False
ls_epsilon: -1
LEARNING_RATE:
function: 'Cosine'
params:
lr: 0.1
OPTIMIZER:
function: 'Momentum'
params:
momentum: 0.9
regularizer:
function: 'L2'
factor: 0.000100
TRAIN:
batch_size: 256
num_workers: 4
file_list: "./dataset/ILSVRC2012/train_list.txt"
data_dir: "./dataset/ILSVRC2012/"
shuffle_seed: 0
transforms:
- DecodeImage:
to_rgb: True
to_np: False
channel_first: False
- RandCropImage:
size: 224
- RandFlipImage:
flip_code: 1
- AutoAugment:
- NormalizeImage:
scale: 1./255.
mean: [0.485, 0.456, 0.406]
std: [0.229, 0.224, 0.225]
order: ''
- ToCHWImage:
VALID:
batch_size: 64
num_workers: 4
file_list: "./dataset/ILSVRC2012/val_list.txt"
data_dir: "./dataset/ILSVRC2012/"
shuffle_seed: 0
transforms:
- DecodeImage:
to_rgb: True
to_np: False
channel_first: False
- ResizeImage:
resize_short: 256
- CropImage:
size: 224
- NormalizeImage:
scale: 1.0/255.0
mean: [0.485, 0.456, 0.406]
std: [0.229, 0.224, 0.225]
order: ''
- ToCHWImage:
configs/DataAugment/ResNet50_Baseline.yaml 0 → 100644
浏览文件 @ 19d03a58
mode: 'train'
ARCHITECTURE:
name: 'ResNet50'
pretrained_model: ""
model_save_dir: "./output/"
classes_num: 1000
total_images: 1281167
save_interval: 1
validate: True
valid_interval: 1
epochs: 300
topk: 5
image_shape: [3, 224, 224]
use_mix: False
ls_epsilon: -1
LEARNING_RATE:
function: 'Cosine'
params:
lr: 0.1
OPTIMIZER:
function: 'Momentum'
params:
momentum: 0.9
regularizer:
function: 'L2'
factor: 0.000100
TRAIN:
batch_size: 256
num_workers: 4
file_list: "./dataset/ILSVRC2012/train_list.txt"
data_dir: "./dataset/ILSVRC2012/"
shuffle_seed: 0
transforms:
- DecodeImage:
to_rgb: True
to_np: False
channel_first: False
- RandCropImage:
size: 224
- RandFlipImage:
flip_code: 1
- NormalizeImage:
scale: 1./255.
mean: [0.485, 0.456, 0.406]
std: [0.229, 0.224, 0.225]
order: ''
- ToCHWImage:
VALID:
batch_size: 64
num_workers: 4
file_list: "./dataset/ILSVRC2012/val_list.txt"
data_dir: "./dataset/ILSVRC2012/"
shuffle_seed: 0
transforms:
- DecodeImage:
to_rgb: True
to_np: False
channel_first: False
- ResizeImage:
resize_short: 256
- CropImage:
size: 224
- NormalizeImage:
scale: 1.0/255.0
mean: [0.485, 0.456, 0.406]
std: [0.229, 0.224, 0.225]
order: ''
- ToCHWImage:
configs/DataAugment/ResNet50_Cutmix.yaml 0 → 100644
浏览文件 @ 19d03a58
mode: 'train'
ARCHITECTURE:
name: 'ResNet50'
pretrained_model: ""
model_save_dir: "./output/"
classes_num: 1000
total_images: 1281167
save_interval: 1
validate: True
valid_interval: 1
epochs: 300
topk: 5
image_shape: [3, 224, 224]
use_mix: True
ls_epsilon: -1
LEARNING_RATE:
function: 'Cosine'
params:
lr: 0.1
OPTIMIZER:
function: 'Momentum'
params:
momentum: 0.9
regularizer:
function: 'L2'
factor: 0.000100
TRAIN:
batch_size: 256
num_workers: 4
file_list: "./dataset/ILSVRC2012/train_list.txt"
data_dir: "./dataset/ILSVRC2012/"
shuffle_seed: 0
transforms:
- DecodeImage:
to_rgb: True
to_np: False
channel_first: False
- RandCropImage:
size: 224
- RandFlipImage:
flip_code: 1
- NormalizeImage:
scale: 1./255.
mean: [0.485, 0.456, 0.406]
std: [0.229, 0.224, 0.225]
order: ''
- ToCHWImage:
mix:
- CutmixOperator:
alpha: 0.2
VALID:
batch_size: 64
num_workers: 4
file_list: "./dataset/ILSVRC2012/val_list.txt"
data_dir: "./dataset/ILSVRC2012/"
shuffle_seed: 0
transforms:
- DecodeImage:
to_rgb: True
to_np: False
channel_first: False
- ResizeImage:
resize_short: 256
- CropImage:
size: 224
- NormalizeImage:
scale: 1.0/255.0
mean: [0.485, 0.456, 0.406]
std: [0.229, 0.224, 0.225]
order: ''
- ToCHWImage:
configs/DataAugment/ResNet50_Cutout.yaml 0 → 100644
浏览文件 @ 19d03a58
mode: 'train'
ARCHITECTURE:
name: 'ResNet50'
pretrained_model: ""
model_save_dir: "./output/"
classes_num: 1000
total_images: 1281167
save_interval: 1
validate: True
valid_interval: 1
epochs: 300
topk: 5
image_shape: [3, 224, 224]
use_mix: False
ls_epsilon: -1
LEARNING_RATE:
function: 'Cosine'
params:
lr: 0.1
OPTIMIZER:
function: 'Momentum'
params:
momentum: 0.9
regularizer:
function: 'L2'
factor: 0.000100
TRAIN:
batch_size: 256
num_workers: 4
file_list: "./dataset/ILSVRC2012/train_list.txt"
data_dir: "./dataset/ILSVRC2012/"
shuffle_seed: 0
transforms:
- DecodeImage:
to_rgb: True
to_np: False
channel_first: False
- RandCropImage:
size: 224
- RandFlipImage:
flip_code: 1
- NormalizeImage:
scale: 1./255.
mean: [0.485, 0.456, 0.406]
std: [0.229, 0.224, 0.225]
order: ''
- Cutout:
n_holes: 1
length: 112
- ToCHWImage:
VALID:
batch_size: 64
num_workers: 4
file_list: "./dataset/ILSVRC2012/val_list.txt"
data_dir: "./dataset/ILSVRC2012/"
shuffle_seed: 0
transforms:
- DecodeImage:
to_rgb: True
to_np: False
channel_first: False
- ResizeImage:
resize_short: 256
- CropImage:
size: 224
- NormalizeImage:
scale: 1.0/255.0
mean: [0.485, 0.456, 0.406]
std: [0.229, 0.224, 0.225]
order: ''
- ToCHWImage:
configs/DataAugment/ResNet50_GridMask.yaml 0 → 100644
浏览文件 @ 19d03a58
mode: 'train'
ARCHITECTURE:
name: 'ResNet50'
pretrained_model: ""
model_save_dir: "./output/"
classes_num: 1000
total_images: 1281167
save_interval: 1
validate: True
valid_interval: 1
epochs: 300
topk: 5
image_shape: [3, 224, 224]
use_mix: False
ls_epsilon: -1
LEARNING_RATE:
function: 'Cosine'
params:
lr: 0.1
OPTIMIZER:
function: 'Momentum'
params:
momentum: 0.9
regularizer:
function: 'L2'
factor: 0.000100
TRAIN:
batch_size: 256
num_workers: 4
file_list: "./dataset/ILSVRC2012/train_list.txt"
data_dir: "./dataset/ILSVRC2012/"
shuffle_seed: 0
transforms:
- DecodeImage:
to_rgb: True
to_np: False
channel_first: False
- RandCropImage:
size: 224
- RandFlipImage:
flip_code: 1
- NormalizeImage:
scale: 1./255.
mean: [0.485, 0.456, 0.406]
std: [0.229, 0.224, 0.225]
order: ''
- GridMask:
d1: 96
d2: 224
rotate: 1
ratio: 0.5
mode: 0
- ToCHWImage:
VALID:
batch_size: 64
num_workers: 4
file_list: "./dataset/ILSVRC2012/val_list.txt"
data_dir: "./dataset/ILSVRC2012/"
shuffle_seed: 0
transforms:
- DecodeImage:
to_rgb: True
to_np: False
channel_first: False
- ResizeImage:
resize_short: 256
- CropImage:
size: 224
- NormalizeImage:
scale: 1.0/255.0
mean: [0.485, 0.456, 0.406]
std: [0.229, 0.224, 0.225]
order: ''
- ToCHWImage:
configs/DataAugment/ResNet50_HideAndSeek.yaml 0 → 100644
浏览文件 @ 19d03a58
mode: 'train'
ARCHITECTURE:
name: 'ResNet50'
pretrained_model: ""
model_save_dir: "./output/"
classes_num: 1000
total_images: 1281167
save_interval: 1
validate: True
valid_interval: 1
epochs: 300
topk: 5
image_shape: [3, 224, 224]
use_mix: False
ls_epsilon: -1
LEARNING_RATE:
function: 'Cosine'
params:
lr: 0.1
OPTIMIZER:
function: 'Momentum'
params:
momentum: 0.9
regularizer:
function: 'L2'
factor: 0.000100
TRAIN:
batch_size: 256
num_workers: 4
file_list: "./dataset/ILSVRC2012/train_list.txt"
data_dir: "./dataset/ILSVRC2012/"
shuffle_seed: 0
transforms:
- DecodeImage:
to_rgb: True
to_np: False
channel_first: False
- RandCropImage:
size: 224
- RandFlipImage:
flip_code: 1
- NormalizeImage:
scale: 1./255.
mean: [0.485, 0.456, 0.406]
std: [0.229, 0.224, 0.225]
order: ''
- HideAndSeek:
- ToCHWImage:
VALID:
batch_size: 64
num_workers: 4
file_list: "./dataset/ILSVRC2012/val_list.txt"
data_dir: "./dataset/ILSVRC2012/"
shuffle_seed: 0
transforms:
- DecodeImage:
to_rgb: True
to_np: False
channel_first: False
- ResizeImage:
resize_short: 256
- CropImage:
size: 224
- NormalizeImage:
scale: 1.0/255.0
mean: [0.485, 0.456, 0.406]
std: [0.229, 0.224, 0.225]
order: ''
- ToCHWImage:
configs/DataAugment/ResNet50_Mixup.yaml 0 → 100644
浏览文件 @ 19d03a58
mode: 'train'
ARCHITECTURE:
name: 'ResNet50'
pretrained_model: ""
model_save_dir: "./output/"
classes_num: 1000
total_images: 1281167
save_interval: 1
validate: True
valid_interval: 1
epochs: 300
topk: 5
image_shape: [3, 224, 224]
use_mix: True
ls_epsilon: -1
LEARNING_RATE:
function: 'Cosine'
params:
lr: 0.1
OPTIMIZER:
function: 'Momentum'
params:
momentum: 0.9
regularizer:
function: 'L2'
factor: 0.000100
TRAIN:
batch_size: 256
num_workers: 4
file_list: "./dataset/ILSVRC2012/train_list.txt"
data_dir: "./dataset/ILSVRC2012/"
shuffle_seed: 0
transforms:
- DecodeImage:
to_rgb: True
to_np: False
channel_first: False
- RandCropImage:
size: 224
- RandFlipImage:
flip_code: 1
- NormalizeImage:
scale: 1./255.
mean: [0.485, 0.456, 0.406]
std: [0.229, 0.224, 0.225]
order: ''
- ToCHWImage:
mix:
- MixupOperator:
alpha: 0.2
VALID:
batch_size: 64
num_workers: 4
file_list: "./dataset/ILSVRC2012/val_list.txt"
data_dir: "./dataset/ILSVRC2012/"
shuffle_seed: 0
transforms:
- DecodeImage:
to_rgb: True
to_np: False
channel_first: False
- ResizeImage:
resize_short: 256
- CropImage:
size: 224
- NormalizeImage:
scale: 1.0/255.0
mean: [0.485, 0.456, 0.406]
std: [0.229, 0.224, 0.225]
order: ''
- ToCHWImage:
configs/DataAugment/ResNet50_RandAugment.yaml 0 → 100644
浏览文件 @ 19d03a58
mode: 'train'
ARCHITECTURE:
name: 'ResNet50'
pretrained_model: ""
model_save_dir: "./output/"
classes_num: 1000
total_images: 1281167
save_interval: 1
validate: True
valid_interval: 1
epochs: 300
topk: 5
image_shape: [3, 224, 224]
use_mix: False
ls_epsilon: -1
LEARNING_RATE:
function: 'Cosine'
params:
lr: 0.1
OPTIMIZER:
function: 'Momentum'
params:
momentum: 0.9
regularizer:
function: 'L2'
factor: 0.000100
TRAIN:
batch_size: 256
num_workers: 4
file_list: "./dataset/ILSVRC2012/train_list.txt"
data_dir: "./dataset/ILSVRC2012/"
shuffle_seed: 0
transforms:
- DecodeImage:
to_rgb: True
to_np: False
channel_first: False
- RandCropImage:
size: 224
- RandFlipImage:
flip_code: 1
- RandAugment:
num_layers: 2
magnitude: 5
- NormalizeImage:
scale: 1./255.
mean: [0.485, 0.456, 0.406]
std: [0.229, 0.224, 0.225]
order: ''
- ToCHWImage:
VALID:
batch_size: 64
num_workers: 4
file_list: "./dataset/ILSVRC2012/val_list.txt"
data_dir: "./dataset/ILSVRC2012/"
shuffle_seed: 0
transforms:
- DecodeImage:
to_rgb: True
to_np: False
channel_first: False
- ResizeImage:
resize_short: 256
- CropImage:
size: 224
- NormalizeImage:
scale: 1.0/255.0
mean: [0.485, 0.456, 0.406]
std: [0.229, 0.224, 0.225]
order: ''
- ToCHWImage:
configs/DataAugment/ResNet50_RandomErasing.yaml 0 → 100644
浏览文件 @ 19d03a58
mode: 'train'
ARCHITECTURE:
name: 'ResNet50'
pretrained_model: ""
model_save_dir: "./output/"
classes_num: 1000
total_images: 1281167
save_interval: 1
validate: True
valid_interval: 1
epochs: 300
topk: 5
image_shape: [3, 224, 224]
use_mix: False
ls_epsilon: -1
LEARNING_RATE:
function: 'Cosine'
params:
lr: 0.1
OPTIMIZER:
function: 'Momentum'
params:
momentum: 0.9
regularizer:
function: 'L2'
factor: 0.000100
TRAIN:
batch_size: 256
num_workers: 4
file_list: "./dataset/ILSVRC2012/train_list.txt"
data_dir: "./dataset/ILSVRC2012/"
shuffle_seed: 0
transforms:
- DecodeImage:
to_rgb: True
to_np: False
channel_first: False
- RandCropImage:
size: 224
- RandFlipImage:
flip_code: 1
- NormalizeImage:
scale: 1./255.
mean: [0.485, 0.456, 0.406]
std: [0.229, 0.224, 0.225]
order: ''
- RandomErasing:
EPSILON: 0.5
sl: 0.02
sh: 0.4
r1: 0.3
mean: [0., 0., 0.]
- ToCHWImage:
VALID:
batch_size: 64
num_workers: 4
file_list: "./dataset/ILSVRC2012/val_list.txt"
data_dir: "./dataset/ILSVRC2012/"
shuffle_seed: 0
transforms:
- DecodeImage:
to_rgb: True
to_np: False
channel_first: False
- ResizeImage:
resize_short: 256
- CropImage:
size: 224
- NormalizeImage:
scale: 1.0/255.0
mean: [0.485, 0.456, 0.406]
std: [0.229, 0.224, 0.225]
order: ''
- ToCHWImage:
configs/Distillation/R50_vd_distill_MV3_large_x1_0.yaml 0 → 100644
浏览文件 @ 19d03a58
mode: 'train'
ARCHITECTURE:
name: 'ResNet50_vd_distill_MobileNetV3_large_x1_0'
pretrained_model: "./pretrained/ResNet50_vd_ssld_pretrained/"
model_save_dir: "./output/"
classes_num: 1000
total_images: 1281167
save_interval: 1
validate: True
valid_interval: 1
epochs: 120
topk: 5
image_shape: [3, 224, 224]
use_distillation: True
LEARNING_RATE:
function: 'CosineWarmup'
params:
lr: 1.3
OPTIMIZER:
function: 'Momentum'
params:
momentum: 0.9
regularizer:
function: 'L2'
factor: 0.00001
TRAIN:
batch_size: 2048
num_workers: 4
file_list: "./dataset/ILSVRC2012/train_list.txt"
data_dir: "./dataset/ILSVRC2012/"
shuffle_seed: 0
transforms:
- DecodeImage:
to_rgb: True
to_np: False
channel_first: False
- RandCropImage:
size: 224
- RandFlipImage:
flip_code: 1
- NormalizeImage:
scale: 1./255.
mean: [0.485, 0.456, 0.406]
std: [0.229, 0.224, 0.225]
order: ''
- ToCHWImage:
VALID:
batch_size: 64
num_workers: 4
file_list: "./dataset/ILSVRC2012/val_list.txt"
data_dir: "./dataset/ILSVRC2012/"
shuffle_seed: 0
transforms:
- DecodeImage:
to_rgb: True
to_np: False
channel_first: False
- ResizeImage:
resize_short: 256
- CropImage:
size: 224
- NormalizeImage:
scale: 1.0/255.0
mean: [0.485, 0.456, 0.406]
std: [0.229, 0.224, 0.225]
order: ''
- ToCHWImage:
configs/Distillation/ResX101_32x16d_wsl_distill_R50_vd.yaml 0 → 100644
浏览文件 @ 19d03a58
mode: 'train'
ARCHITECTURE:
name: 'ResNeXt101_32x16d_wsl_distill_ResNet50_vd'
pretrained_model: "./pretrained/ResNeXt101_32x16d_wsl_pretrained/"
model_save_dir: "./output/"
classes_num: 1000
total_images: 1281167
save_interval: 1
validate: True
valid_interval: 1
epochs: 120
topk: 5
image_shape: [3, 224, 224]
use_distillation: True
LEARNING_RATE:
function: 'CosineWarmup'
params:
lr: 0.4
OPTIMIZER:
function: 'Momentum'
params:
momentum: 0.9
regularizer:
function: 'L2'
factor: 0.00007
TRAIN:
batch_size: 1024
num_workers: 4
file_list: "./dataset/ILSVRC2012/train_list.txt"
data_dir: "./dataset/ILSVRC2012/"
shuffle_seed: 0
transforms:
- DecodeImage:
to_rgb: True
to_np: False
channel_first: False
- RandCropImage:
size: 224
- RandFlipImage:
flip_code: 1
- NormalizeImage:
scale: 1./255.
mean: [0.485, 0.456, 0.406]
std: [0.229, 0.224, 0.225]
order: ''
- ToCHWImage:
VALID:
batch_size: 64
num_workers: 4
file_list: "./dataset/ILSVRC2012/val_list.txt"
data_dir: "./dataset/ILSVRC2012/"
shuffle_seed: 0
transforms:
- DecodeImage:
to_rgb: True
to_np: False
channel_first: False
- ResizeImage:
resize_short: 256
- CropImage:
size: 224
- NormalizeImage:
scale: 1.0/255.0
mean: [0.485, 0.456, 0.406]
std: [0.229, 0.224, 0.225]
order: ''
- ToCHWImage:
docs/zh_CN/advanced_tutorials/distillation/distillation.md
浏览文件 @ 19d03a58
......@@ -185,9 +185,73 @@ for var in ./*_student; do cp "$var" "../student_model/${var%_student}"; done #
# 五、SSLD实战
* 该部分内容正在持续更新中，敬请期待。
本节将介绍训练ImageNet-1k的SSLD蒸馏内容，用户可以通过此处的讲解来实战SSLD蒸馏。如果想快速体验此方法，可以阅读快读开始中的训练Flowers102的SSLD蒸馏章节。
## 7.1 参数配置
实战部分提供了SSLD蒸馏的示例，在`ppcls/modeling/architectures/distillation_models.py`中提供了`ResNeXt101_32x16d_wsl`蒸馏`ResNet50_vd`与`ResNet50_vd_ssld`蒸馏`MobileNetV3_large_x1_0`的示例，`configs/Distillation`里分别提供了二者的配置文件，用户可以在`tools/run.sh`里直接替换配置文件的路径即可使用。
### ResNeXt101_32x16d_wsl蒸馏ResNet50_vd
`ResNeXt101_32x16d_wsl`蒸馏`ResNet50_vd`的配置如下，其中`pretrained model`指定了`ResNeXt101_32x16d_wsl`（教师模型）的预训练模型的路径，该路径也可以同时指定教师模型与学生模型的预训练模型的路径，用户只需要同时传入二者预训练的路径即可（配置中的注释部分）。
```yaml
ARCHITECTURE:
name: 'ResNeXt101_32x16d_wsl_distill_ResNet50_vd'
pretrained_model: "./pretrained/ResNeXt101_32x16d_wsl_pretrained/"
# pretrained_model:
# - "./pretrained/ResNeXt101_32x16d_wsl_pretrained/"
# - "./pretrained/ResNet50_vd_pretrained/"
use_distillation: True
```
### ResNet50_vd_ssld蒸馏MobileNetV3_large_x1_0
类似于`ResNeXt101_32x16d_wsl`蒸馏`ResNet50_vd`，`ResNet50_vd_ssld`蒸馏`MobileNetV3_large_x1_0`的配置如下:
```yaml
ARCHITECTURE:
name: 'ResNet50_vd_distill_MobileNetV3_large_x1_0'
pretrained_model: "./pretrained/ResNet50_vd_ssld_pretrained/"
# pretrained_model:
# - "./pretrained/ResNet50_vd_ssld_pretrained/"
# - "./pretrained/ResNet50_vd_pretrained/"
use_distillation: True
```
## 7.2 启动命令
当用户配置完训练环境后，类似于训练其他分类任务，只需要将`tools/run.sh`中的配置文件替换成为相应的蒸馏配置文件即可。
其中`run.sh`中的内容如下：
```bash
export PYTHONPATH=path_to_PaddleClas:$PYTHONPATH
python -m paddle.distributed.launch \
--selected_gpus="0,1,2,3" \
--log_dir=R50_vd_distill_MV3_large_x1_0 \
tools/train.py \
-c ./configs/Distillation/R50_vd_distill_MV3_large_x1_0.yaml
```
运行`run.sh`
```bash
sh tools/run.sh
```
## 7.3 注意事项
* 用户在使用SSLD蒸馏之前，首先需要在目标数据集上训练一个教师模型，该教师模型用于指导学生模型在该数据集上的训练。
* 在用户使用SSLD蒸馏的时候需要将配置文件中的`use_distillation`设置为`True`，另外由于学生模型学习带有知识信息的soft-label，所以需要关掉label_smoothing选项，即将`ls_epsilon`中的值设置在[0,1]之外。
* 如果学生模型没有加载预训练模型，训练的其他超参数可以参考该学生模型在ImageNet-1k上训练的超参数，如果学生模型加载了预训练模型，学习率可以调整到原来的1/10或者1/100。
* 在SSLD蒸馏的过程中，学生模型只学习soft-label导致训练目标变的更加复杂，建议可以适当的调小`l2_decay`的值来获得更高的验证集准确率。
* 若用户准备添加无标签的训练数据，只需要将新的训练数据放置在原本训练数据的路径下，生成新的数据list即可，另外，新生成的数据list需要将无标签的数据添加伪标签（只是为了统一读数据）。
**此处插播一条硬广~**
> 如果您觉得此文档对您有帮助，欢迎star、watch、fork，三连我们的项目：[https://github.com/PaddlePaddle/PaddleClas](https://github.com/PaddlePaddle/PaddleClas)
......
docs/zh_CN/advanced_tutorials/image_augmentation/ImageAugment.md
浏览文件 @ 19d03a58
......@@ -438,9 +438,119 @@ new_batch = cutmix_op(batch)
# 七、数据增广分类实战
* 该部分内容正在持续更新中，敬请期待。
本节将介绍训练ImageNet-1k的数据增广内容，用户可以通过此处的介绍来进行数据增广的实战训练。如果想快速体验各种数据增广，可以阅读快读开始章节中的训练Flowers102的数据增广内容。
## 7.1 参数配置
由于不同的数据增广方式含有不同的超参数，为了便于理解和使用，我们在`configs/DataAugment`里分别列举了8种训练ResNet50的数据增广方式的参数配置文件，用户可以在`tools/run.sh`里直接替换配置文件的路径即可使用。此处分别挑选了图像变换、图像裁剪、图像混叠中的一个示例展示，其他参数配置用户可以自查配置文件。
### RandAugment
`RandAugment`的图像增广方式的配置如下，其中用户需要指定其中的参数`num_layers`与`magnitude`，默认的数值分别是`2`和`5`。`RandAugment`是在uint8的数据格式上转换的，所以其处理过程应该放在归一化操作（`NormalizeImage`）之前。
```yaml
transforms:
- DecodeImage:
to_rgb: True
to_np: False
channel_first: False
- RandCropImage:
size: 224
- RandFlipImage:
flip_code: 1
- RandAugment:
num_layers: 2
magnitude: 5
- NormalizeImage:
scale: 1./255.
mean: [0.485, 0.456, 0.406]
std: [0.229, 0.224, 0.225]
order: ''
- ToCHWImage:
```
### Cutout
`Cutout`的图像增广方式的配置如下，其中用户需要指定其中的参数`n_holes`与`length`，默认的数值分别是`1`和`112`。类似其他图像裁剪类的数据增广方式，`Cutout`既可以在uint8格式的数据上操作，也可以在归一化（`NormalizeImage`）后的数据上操作，此处给出的是在归一化后的操作。
```yaml
transforms:
- DecodeImage:
to_rgb: True
to_np: False
channel_first: False
- RandCropImage:
size: 224
- RandFlipImage:
flip_code: 1
- NormalizeImage:
scale: 1./255.
mean: [0.485, 0.456, 0.406]
std: [0.229, 0.224, 0.225]
order: ''
- Cutout:
n_holes: 1
length: 112
- ToCHWImage:
```
### Mixup
`Mixup`的图像增广方式的配置如下，其中用户需要指定其中的参数`alpha`，默认的数值是`0.2`。类似其他图像混合类的数据增广方式，`Mixup`是在图像做完数据处理后将每个batch内的数据做图像混叠，将混叠后的图像和标签输入网络中训练，所以其是在图像数据处理（图像变换、图像裁剪）后操作。另外，在配置文件中，需要将`use_mix`参数设置为`True`。
```yaml
transforms:
- DecodeImage:
to_rgb: True
to_np: False
channel_first: False
- RandCropImage:
size: 224
- RandFlipImage:
flip_code: 1
- NormalizeImage:
scale: 1./255.
mean: [0.485, 0.456, 0.406]
std: [0.229, 0.224, 0.225]
order: ''
- ToCHWImage:
mix:
- MixupOperator:
alpha: 0.2
```
## 7.2 启动命令
当用户配置完训练环境后，类似于训练其他分类任务，只需要将`tools/run.sh`中的配置文件替换成为相应的数据增广方式的配置文件即可。
其中`run.sh`中的内容如下：
```bash
export PYTHONPATH=path_to_PaddleClas:$PYTHONPATH
python -m paddle.distributed.launch \
--selected_gpus="0,1,2,3" \
--log_dir=ResNet50_Cutout \
tools/train.py \
-c ./configs/DataAugment/ResNet50_Cutout.yaml
```
运行`run.sh`
```bash
sh tools/run.sh
```
## 7.3 注意事项
* 在使用图像混叠类的数据处理时，需要将配置文件中的`use_mix`设置为`True`，另外由于图像混叠时需对label进行混叠，无法计算训练数据的准确率，所以在训练过程中没有打印训练准确率。
* 在使用数据增广后，由于训练数据更难，所以训练损失函数可能较大，训练集的准确率相对较低，但其有拥更好的泛化能力，所以验证集的准确率相对较高。
* 在使用数据增广后，模型可能会趋于欠拟合状态，建议可以适当的调小`l2_decay`的值来获得更高的验证集准确率。
* 几乎每一类图像增广均含有超参数，我们只提供了基于ImageNet-1k的超参数，其他数据集需要用户自己调试超参数，具体超参数的含义用户可以阅读相关的论文，调试方法也可以参考训练技巧的章节。
**此处插播一条硬广~**
> 如果您觉得此文档对您有帮助，欢迎star、watch、fork，三连我们的项目：[https://github.com/PaddlePaddle/PaddleClas](https://github.com/PaddlePaddle/PaddleClas)
......
ppcls/data/imaug/__init__.py
浏览文件 @ 19d03a58
......@@ -44,12 +44,12 @@ def transform(data, ops=[]):
return data
class ImageNetPolicy(RawImageNetPolicy):
class AutoAugment(RawImageNetPolicy):
""" ImageNetPolicy wrapper to auto fit different img types """
def __init__(self, *args, **kwargs):
if six.PY2:
super(ImageNetPolicy, self).__init__(*args, **kwargs)
super(AutoAugment, self).__init__(*args, **kwargs)
else:
super().__init__(*args, **kwargs)
......@@ -59,7 +59,7 @@ class ImageNetPolicy(RawImageNetPolicy):
img = Image.fromarray(img)
if six.PY2:
img = super(ImageNetPolicy, self).__call__(img)
img = super(AutoAugment, self).__call__(img)
else:
img = super().__call__(img)
......
ppcls/data/imaug/grid.py
浏览文件 @ 19d03a58
......@@ -25,7 +25,7 @@ NUM_EPOCHS = 240
class GridMask(object):
def __init__(self, d1, d2, rotate=1, ratio=0.5, mode=0, prob=1.):
def __init__(self, d1=96, d2=224, rotate=1, ratio=0.5, mode=0, prob=1.):
self.d1 = d1
self.d2 = d2
self.rotate = rotate
......
ppcls/data/imaug/randaugment.py
浏览文件 @ 19d03a58
......@@ -20,7 +20,7 @@ import random
class RandAugment(object):
def __init__(self, num_layers, magnitude, fillcolor=(128, 128, 128)):
def __init__(self, num_layers=2, magnitude=5, fillcolor=(128, 128, 128)):
self.num_layers = num_layers
self.magnitude = magnitude
self.max_level = 10
......
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