Skip to content

P
PaddleClas

PaddlePaddle / PaddleClas
接近 2 年 前同步成功

通知 116
Star 4999
Fork 1114
P
PaddleClas
提交 a3c154fa 编写于 作者: W weishengyu
浏览文件
操作

add car_config

上级 6432a7c4
隐藏空白更改
内联 并排
Showing with 866 addition and 48 deletion
docs/zh_CN/PULC/PULC_car_exists.md
浏览文件 @ a3c154fa
......@@ -41,19 +41,19 @@
该案例提供了用户使用 PaddleClas 的超轻量图像分类方案（PULC，Practical Ultra Lightweight Classification）快速构建轻量级、高精度、可落地的有人/无人的分类模型。该模型可以广泛应用于如监控场景、人员进出管控场景、海量数据过滤场景等。
下表列出了判断图片中是否有人的二分类模型的相关指标，前两行展现了使用 SwinTranformer_tiny 和 MobileNetV3_small_x0_35 作为 backbone 训练得到的模型的相关指标，第三行至第六行依次展现了替换 backbone 为 PPLCNet_x1_0、使用 SSLD 预训练模型、使用 SSLD 预训练模型 + EDA 策略、使用 SSLD 预训练模型 + EDA 策略 + SKL-UGI 知识蒸馏策略训练得到的模型的相关指标。
下表列出了判断图片中是否有车的二分类模型的相关指标，前两行展现了使用 SwinTranformer_tiny 和 MobileNetV3_small_x0_35 作为 backbone 训练得到的模型的相关指标，第三行至第六行依次展现了替换 backbone 为 PPLCNet_x1_0、使用 SSLD 预训练模型、使用 SSLD 预训练模型 + EDA 策略、使用 SSLD 预训练模型 + EDA 策略 + SKL-UGI 知识蒸馏策略训练得到的模型的相关指标。
| 模型 | Tpr（%） | 延时（ms） | 存储（M） | 策略 |
|-------|-----------|----------|---------------|---------------|
| SwinTranformer_tiny | 95.69 | 95.30 | 107 | 使用 ImageNet 预训练模型 |
| MobileNetV3_small_x0_35 | 68.25 | 2.85 | 1.6 | 使用 ImageNet 预训练模型 |
| PPLCNet_x1_0 | 89.57 | 2.12 | 6.5 | 使用 ImageNet 预训练模型 |
| PPLCNet_x1_0 | 92.10 | 2.12 | 6.5 | 使用 SSLD 预训练模型 |
| PPLCNet_x1_0 | 93.43 | 2.12 | 6.5 | 使用 SSLD 预训练模型+EDA 策略|
| <b>PPLCNet_x1_0<b> | <b>95.60<b> | <b>2.12<b> | <b>6.5<b> | 使用 SSLD 预训练模型+EDA 策略+SKL-UGI 知识蒸馏策略|
| 模型 | Tpr（%）@Fpr0.01 | 延时（ms） | 存储（M） | 策略 |
|-------|----------------|----------|---------------|---------------|
| SwinTranformer_tiny | 97.71 | 95.30 | 107 | 使用 ImageNet 预训练模型 |
| MobileNetV3_small_x0_35 | 81.23 | 2.85 | 1.6 | 使用 ImageNet 预训练模型 |
| PPLCNet_x1_0 | 81.23 | 2.12 | 6.5 | 使用 ImageNet 预训练模型 |
| PPLCNet_x1_0 | 95.48 | 2.12 | 6.5 | 使用 SSLD 预训练模型 |
| PPLCNet_x1_0 | 95.48 | 2.12 | 6.5 | 使用 SSLD 预训练模型+EDA 策略|
| <b>PPLCNet_x1_0<b> | <b>95.72<b> | <b>2.12<b> | <b>6.5<b> | 使用 SSLD 预训练模型+EDA 策略+SKL-UGI 知识蒸馏策略|
从表中可以看出，backbone 为 SwinTranformer_tiny 时精度较高，但是推理速度较慢。将 backboone 替换为轻量级模型 MobileNetV3_small_x0_35 后，速度可以大幅提升，但是会导致精度大幅下降。将 backbone 替换为速度更快的 PPLCNet_x1_0 时，精度较 MobileNetV3_small_x0_35 高 20 多个百分点，与此同时速度依旧可以快 20% 以上。在此基础上，使用 SSLD 预训练模型后，在不改变推理速度的前提下，精度可以提升约 2.6 个百分点，进一步地，当融合EDA策略后，精度可以再提升 1.3 个百分点，最后，在使用 SKL-UGI 知识蒸馏后，精度可以继续提升 2.2 个百分点。此时，PPLCNet_x1_0 达到了 SwinTranformer_tiny 模型的精度，但是速度快 40 多倍。关于 PULC 的训练方法和推理部署方法将在下面详细介绍。
从表中可以看出，backbone 为 SwinTranformer_tiny 时精度较高，但是推理速度较慢。将 backboone 替换为轻量级模型 MobileNetV3_small_x0_35 后，速度可以大幅提升，但是会导致精度大幅下降。将 backbone 替换为速度更快的 PPLCNet_x1_0 时，精度较 MobileNetV3_small_x0_35 高 20 多个百分点，与此同时速度依旧可以快 20% 以上。在此基础上，使用 SSLD 预训练模型后，在不改变推理速度的前提下，精度可以提升约 2.6 个百分点，进一步地，在使用 SKL-UGI 知识蒸馏后，精度可以继续提升 0.24 个百分点。此时，PPLCNet_x1_0 达到了 SwinTranformer_tiny 模型的精度，但是速度快 40 多倍。关于 PULC 的训练方法和推理部署方法将在下面详细介绍。
**备注：**
......@@ -81,13 +81,13 @@ pip3 install paddlepaddle paddleclas
* 使用命令行快速预测
```bash
paddleclas --model_name=car_exists --infer_imgs=deploy/images/PULC/person_exists/objects365_01780782.jpg
paddleclas --model_name=car_exists --infer_imgs=deploy/images/PULC/car_exists/objects365_00001507.jpeg
```
结果如下：
```
>>> result
class_ids: [0], scores: [0.9955421453341842], label_names: ['nobody'], filename: deploy/images/PULC/person_exists/objects365_01780782.jpg
class_ids: [1], scores: [0.9740616], label_names: ['contains_vehicle'], filename: deploy/images/PULC/car_exists/objects365_00001507.jpeg
Predict complete!
```
......@@ -98,7 +98,7 @@ Predict complete!
```python
import paddleclas
model = paddleclas.PaddleClas(model_name="car_exists")
result = model.predict(input_data="deploy/images/PULC/person_exists/objects365_01780782.jpg")
result = model.predict(input_data="deploy/images/PULC/car_exists/objects365_00001507.jpeg")
print(next(result))
```
......@@ -106,7 +106,7 @@ print(next(result))
```
>>> result
[{'class_ids': [0], 'scores': [0.9955421453341842], 'label_names': ['nobody'], 'filename': 'PaddleClas/deploy/images/PULC/person_exists/objects365_01780782.jpg'}]
[{'class_ids': [1], 'scores': [0.9740616], 'label_names': ['contains_vehicle'], 'filename': 'deploy/images/PULC/car_exists/objects365_00001507.jpeg'}]
```
<a name="3"></a>
......@@ -135,13 +135,15 @@ print(next(result))
在公开数据集的基础上经过后处理即可得到本案例需要的数据，具体处理方法如下：
- 训练集合，本案例处理了 Objects365 数据训练集的标注文件，如果某张图含有“人”的标签，且这个框的面积在整张图中的比例大于 10%，即认为该张图中含有车，如果某张图中没有“车”的标签，则认为该张图中不含有车。经过处理后，得到 92964 条可用数据，其中有人的数据有 39813 条，无人的数据 53151 条。
- 训练集合，本案例处理了 Objects365 数据训练集的标注文件，如果某张图含有“car”的标签，且这个框的面积在整张图中的比例大于 10%，即认为该张图中含有车，如果某张图中没有任何与交通工具，例如car、bus等相关的的标签，则认为该张图中不含有车。经过处理后，得到 108629 条可用数据，其中有车的数据有 27422 条，无车的数据 81207 条。
- 验证集合，从 Object365 数据中随机抽取一小部分数据，使用在 MS-COCO 上训练得到的较好的模型预测这些数据，将预测结果和数据的标注文件取交集，将交集的结果按照得到训练集的方法筛选出验证集合。经过处理后，得到 27820 条可用数据。其中有人的数据有 2255 条，无人的数据有 25565 条。
- 验证集合，处理方法与训练集相同，数据来源与 Objects365 数据集的验证集。为了测试结果准确，验证集经过人工校正，去除了一些可能存在标注错误的图像。
* 注：由于objects365的标签并不是完全互斥的，例如F1赛车可能是 "F1 Formula"，也可能被标称"car"。为了减轻干扰，我们仅保留"car"标签作为有车，而将不含任何交通工具的图作为无车。
处理后的数据集部分数据可视化如下：
![](../../images/PULC/docs/car_exists_data_demo.png)
![](../../images/PULC/docs/car_exists_data_demo.jpeg)
此处提供了经过上述方法处理好的数据，可以直接下载得到。
......@@ -156,23 +158,18 @@ cd path_to_PaddleClas
```shell
cd dataset
wget https://paddleclas.bj.bcebos.com/data/PULC/person_exists.tar
tar -xf person_exists.tar
wget https://paddleclas.bj.bcebos.com/data/PULC/car_exists.tar
tar -xf car_exists.tar
cd ../
```
执行上述命令后，`dataset/` 下存在 `person_exists` 目录，该目录中具有以下数据：
执行上述命令后，`dataset/` 下存在 `car_exists` 目录，该目录中具有以下数据：
```
├── train
│   ├── 000000000009.jpg
│   ├── 000000000025.jpg
...
├── val
│   ├── objects365_01780637.jpg
│   ├── objects365_01780640.jpg
...
├── objects365_car
│   ├── objects365_00000039.jpg
│   ├── objects365_00000099.jpg
├── ImageNet_val
│   ├── ILSVRC2012_val_00000001.JPEG
│   ├── ILSVRC2012_val_00000002.JPEG
......@@ -205,14 +202,14 @@ export CUDA_VISIBLE_DEVICES=0,1,2,3
python3 -m paddle.distributed.launch \
--gpus="0,1,2,3" \
tools/train.py \
-c ./ppcls/configs/PULC/person_exists/PPLCNet_x1_0.yaml
-c ./ppcls/configs/PULC/car_exists/PPLCNet_x1_0.yaml
```
验证集的最佳指标在 `0.94-0.95` 之间（数据集较小，容易造成波动）。
验证集的最佳指标在 `0.95-0.96` 之间（数据集较小，容易造成波动）。
**备注：**
* 此时使用的指标为Tpr，该指标描述了在假正类率（Fpr）小于某一个指标时的真正类率（Tpr），是产业中二分类问题常用的指标之一。在本案例中，Fpr 为千分之一。关于 Fpr 和 Tpr 的更多介绍，可以参考[这里](https://baike.baidu.com/item/AUC/19282953)。
* 此时使用的指标为Tpr，该指标描述了在假正类率（Fpr）小于某一个指标时的真正类率（Tpr），是产业中二分类问题常用的指标之一。在本案例中，Fpr 为 1/100 。关于 Fpr 和 Tpr 的更多介绍，可以参考[这里](https://baike.baidu.com/item/AUC/19282953)。
* 在eval时，会打印出来当前最佳的 TprAtFpr 指标，具体地，其会打印当前的 `Fpr`、`Tpr` 值，以及当前的 `threshold`值，`Tpr` 值反映了在当前 `Fpr` 值下的召回率，该值越高，代表模型越好。`threshold` 表示当前最佳 `Fpr` 所对应的分类阈值，可用于后续模型部署落地等。
......@@ -224,7 +221,7 @@ python3 -m paddle.distributed.launch \
```bash
python3 tools/eval.py \
-c ./ppcls/configs/PULC/person_exists/PPLCNet_x1_0.yaml \
-c ./ppcls/configs/PULC/car_exists/PPLCNet_x1_0.yaml \
-o Global.pretrained_model="output/PPLCNet_x1_0/best_model"
```
......@@ -238,21 +235,21 @@ python3 tools/eval.py \
```python
python3 tools/infer.py \
-c ./ppcls/configs/PULC/person_exists/PPLCNet_x1_0.yaml \
-c ./ppcls/configs/PULC/car_exists/PPLCNet_x1_0.yaml \
-o Global.pretrained_model=output/PPLCNet_x1_0/best_model \
```
输出结果如下：
```
[{'class_ids': [0], 'scores': [0.9878496769815683], 'label_names': ['nobody'], 'file_name': './dataset/person_exists/val/objects365_01780637.jpg'}]
[{'class_ids': [1], 'scores': [0.9740616], 'label_names': ['contains_vehicle'], 'filename': 'deploy/images/PULC/car_exists/objects365_00001507.jpeg'}]
```
**备注：**
* 这里`-o Global.pretrained_model="output/PPLCNet_x1_0/best_model"` 指定了当前最佳权重所在的路径，如果指定其他权重，只需替换对应的路径即可。
* 默认是对 `deploy/images/PULC/person_exists/objects365_02035329.jpg` 进行预测，此处也可以通过增加字段 `-o Infer.infer_imgs=xxx` 对其他图片预测。
* 默认是对 `deploy/images/PULC/car_exists/objects365_00001507.jpeg` 进行预测，此处也可以通过增加字段 `-o Infer.infer_imgs=xxx` 对其他图片预测。
* 二分类默认的阈值为0.5， 如果需要指定阈值，可以重写 `Infer.PostProcess.threshold` ，如`-o Infer.PostProcess.threshold=0.9794`，该值需要根据实际场景来确定，此处的 `0.9794` 是在该场景中的 `val` 数据集在千分之一 Fpr 下得到的最佳 Tpr 所得到的。
......@@ -278,7 +275,7 @@ export CUDA_VISIBLE_DEVICES=0,1,2,3
python3 -m paddle.distributed.launch \
--gpus="0,1,2,3" \
tools/train.py \
-c ./ppcls/configs/PULC/person_exists/PPLCNet/PPLCNet_x1_0.yaml \
-c ./ppcls/configs/PULC/car_exists/PPLCNet/PPLCNet_x1_0.yaml \
-o Arch.name=ResNet101_vd
```
......@@ -295,7 +292,7 @@ export CUDA_VISIBLE_DEVICES=0,1,2,3
python3 -m paddle.distributed.launch \
--gpus="0,1,2,3" \
tools/train.py \
-c ./ppcls/configs/PULC/person_exists/PPLCNet_x1_0_distillation.yaml \
-c ./ppcls/configs/PULC/car_exists/PPLCNet_x1_0_distillation.yaml \
-o Arch.models.0.Teacher.pretrained=output/ResNet101_vd/best_model
```
......@@ -330,14 +327,14 @@ Paddle Inference 是飞桨的原生推理库， 作用于服务器端和云端
```bash
python3 tools/export_model.py \
-c ./ppcls/configs/PULC/person_exists/PPLCNet_x1_0.yaml \
-c ./ppcls/configs/PULC/car_exists/PPLCNet_x1_0.yaml \
-o Global.pretrained_model=output/DistillationModel/best_model_student \
-o Global.save_inference_dir=deploy/models/PPLCNet_x1_0_person_exists_infer
-o Global.save_inference_dir=deploy/models/PPLCNet_x1_0_car_exists_infer
```
执行完该脚本后会在 `deploy/models/` 下生成 `PPLCNet_x1_0_person_exists_infer` 文件夹，`models` 文件夹下应有如下文件结构：
执行完该脚本后会在 `deploy/models/` 下生成 `PPLCNet_x1_0_car_exists_infer` 文件夹，`models` 文件夹下应有如下文件结构：
```
├── PPLCNet_x1_0_person_exists_infer
├── PPLCNet_x1_0_car_exists_infer
│ ├── inference.pdiparams
│ ├── inference.pdiparams.info
│ └── inference.pdmodel
......@@ -354,13 +351,13 @@ python3 tools/export_model.py \
```
cd deploy/models
# 下载 inference 模型并解压
wget https://paddleclas.bj.bcebos.com/models/PULC/person_exists_infer.tar && tar -xf person_exists_infer.tar
wget https://paddleclas.bj.bcebos.com/models/PULC/car_exists_infer.tar && tar -xf car_exists_infer.tar
```
解压完毕后，`models` 文件夹下应有如下文件结构：
```
├── person_exists_infer
├── car_exists_infer
│ ├── inference.pdiparams
│ ├── inference.pdiparams.info
│ └── inference.pdmodel
......@@ -381,13 +378,13 @@ wget https://paddleclas.bj.bcebos.com/models/PULC/person_exists_infer.tar && tar
cd ../
```
运行下面的命令，对图像 `./images/PULC/person_exists/objects365_02035329.jpg` 进行有人/无人分类。
运行下面的命令，对图像 `./images/PULC/car_exists/objects365_00001507.jpeg` 进行有人/无人分类。
```shell
# 使用下面的命令使用 GPU 进行预测
python3.7 python/predict_cls.py -c configs/PULC/person_exists/inference_person_exists.yaml
python3.7 python/predict_cls.py -c configs/PULC/car_exists/inference_car_exists.yaml
# 使用下面的命令使用 CPU 进行预测
python3.7 python/predict_cls.py -c configs/PULC/person_exists/inference_person_exists.yaml -o Global.use_gpu=False
python3.7 python/predict_cls.py -c configs/PULC/car_exists/inference_car_exists.yaml -o Global.use_gpu=False
```
输出结果如下。
......@@ -407,7 +404,7 @@ objects365_02035329.jpg: class id(s): [1], score(s): [1.00], label_name(s): ['so
```shell
# 使用下面的命令使用 GPU 进行预测，如果希望使用 CPU 预测，可以在命令后面添加 -o Global.use_gpu=False
python3.7 python/predict_cls.py -c configs/PULC/person_exists/inference_person_exists.yaml -o Global.infer_imgs="./images/PULC/person_exists/"
python3.7 python/predict_cls.py -c configs/PULC/car_exists/inference_car_exists.yaml -o Global.infer_imgs="./images/PULC/car_exists/"
```
终端中会输出该文件夹内所有图像的分类结果，如下所示。
......@@ -417,7 +414,7 @@ objects365_01780782.jpg: class id(s): [0], score(s): [1.00], label_name(s): ['no
objects365_02035329.jpg: class id(s): [1], score(s): [1.00], label_name(s): ['someone']
```
其中，`someone` 表示该图里存在人，`nobody` 表示该图里不存在人。
其中，`contains_car` 表示该图里存在车，`nocar` 表示该图里不存在车。
<a name="6.3"></a>
......
ppcls/configs/PULC/car_exists/MobileNetV3_small_x0_35.yaml 0 → 100644
浏览文件 @ a3c154fa
# global configs
Global:
checkpoints: null
pretrained_model: null
output_dir: ./output/
device: gpu
save_interval: 1
eval_during_train: True
eval_interval: 1
start_eval_epoch: 10
epochs: 20
print_batch_step: 10
use_visualdl: False
# used for static mode and model export
image_shape: [3, 224, 224]
save_inference_dir: ./inference
# training model under @to_static
to_static: False
use_dali: False
# model architecture
Arch:
name: MobileNetV3_small_x0_35
class_num: 2
pretrained: True
use_sync_bn: True
# loss function config for traing/eval process
Loss:
Train:
- CELoss:
weight: 1.0
epsilon: 0.1
Eval:
- CELoss:
weight: 1.0
Optimizer:
name: Momentum
momentum: 0.9
lr:
name: Cosine
learning_rate: 0.05
warmup_epoch: 5
regularizer:
name: 'L2'
coeff: 0.00001
# data loader for train and eval
DataLoader:
Train:
dataset:
name: ImageNetDataset
image_root: ./dataset/car_exists/
cls_label_path: ./dataset/car_exists/train_list.txt
transform_ops:
- DecodeImage:
to_rgb: True
channel_first: False
- RandCropImage:
size: 224
- RandFlipImage:
flip_code: 1
- NormalizeImage:
scale: 1.0/255.0
mean: [0.485, 0.456, 0.406]
std: [0.229, 0.224, 0.225]
order: ''
sampler:
name: DistributedBatchSampler
batch_size: 512
drop_last: False
shuffle: True
loader:
num_workers: 8
use_shared_memory: True
Eval:
dataset:
name: ImageNetDataset
image_root: ./dataset/car_exists/
cls_label_path: ./dataset/car_exists/val_list.txt
transform_ops:
- DecodeImage:
to_rgb: True
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: ''
sampler:
name: DistributedBatchSampler
batch_size: 64
drop_last: False
shuffle: False
loader:
num_workers: 4
use_shared_memory: True
Infer:
infer_imgs: deploy/images/PULC/car_exists/objects365_00001507.jpeg
batch_size: 10
transforms:
- DecodeImage:
to_rgb: True
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:
PostProcess:
name: ThreshOutput
threshold: 0.5
label_0: nobody
label_1: someone
Metric:
Train:
- TopkAcc:
topk: [1, 2]
Eval:
- TprAtFpr:
max_fpr: 0.01
- TopkAcc:
topk: [1, 2]
ppcls/configs/PULC/car_exists/PPLCNet_x1_0.yaml 0 → 100644
浏览文件 @ a3c154fa
# global configs
Global:
checkpoints: null
pretrained_model: null
output_dir: ./output/
device: gpu
save_interval: 1
eval_during_train: True
eval_interval: 1
start_eval_epoch: 10
epochs: 20
print_batch_step: 10
use_visualdl: False
# used for static mode and model export
image_shape: [3, 224, 224]
save_inference_dir: ./inference
# training model under @to_static
to_static: False
use_dali: False
# model architecture
Arch:
name: PPLCNet_x1_0
class_num: 2
pretrained: True
use_ssld: True
use_sync_bn: True
# loss function config for traing/eval process
Loss:
Train:
- CELoss:
weight: 1.0
Eval:
- CELoss:
weight: 1.0
Optimizer:
name: Momentum
momentum: 0.9
lr:
name: Cosine
learning_rate: 0.0125
warmup_epoch: 5
regularizer:
name: 'L2'
coeff: 0.00004
# data loader for train and eval
DataLoader:
Train:
dataset:
name: ImageNetDataset
image_root: ./dataset/car_exists/
cls_label_path: ./dataset/car_exists/train_list.txt
transform_ops:
- DecodeImage:
to_rgb: True
channel_first: False
- RandCropImage:
size: 192
- RandFlipImage:
flip_code: 1
- TimmAutoAugment:
prob: 0.5
config_str: rand-m9-mstd0.5-inc1
interpolation: bicubic
img_size: 192
- NormalizeImage:
scale: 1.0/255.0
mean: [0.485, 0.456, 0.406]
std: [0.229, 0.224, 0.225]
order: ''
- RandomErasing:
EPSILON: 0.5
sl: 0.02
sh: 1.0/3.0
r1: 0.3
attempt: 10
use_log_aspect: True
mode: pixel
sampler:
name: DistributedBatchSampler
batch_size: 64
drop_last: False
shuffle: True
loader:
num_workers: 8
use_shared_memory: True
Eval:
dataset:
name: ImageNetDataset
image_root: ./dataset/car_exists
cls_label_path: ./dataset/car_exists/val_list.txt
transform_ops:
- DecodeImage:
to_rgb: True
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: ''
sampler:
name: DistributedBatchSampler
batch_size: 64
drop_last: False
shuffle: False
loader:
num_workers: 4
use_shared_memory: True
Infer:
infer_imgs: deploy/images/PULC/car_exists/objects365_00001507.jpeg
batch_size: 10
transforms:
- DecodeImage:
to_rgb: True
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:
PostProcess:
name: ThreshOutput
threshold: 0.9
label_0: nobody
label_1: someone
Metric:
Train:
- TopkAcc:
topk: [1, 2]
Eval:
- TprAtFpr:
max_fpr: 0.01
- TopkAcc:
topk: [1, 2]
ppcls/configs/PULC/car_exists/PPLCNet_x1_0_distillation.yaml 0 → 100644
浏览文件 @ a3c154fa
# global configs
Global:
checkpoints: null
pretrained_model: null
output_dir: ./output
device: gpu
save_interval: 1
eval_during_train: True
start_eval_epoch: 1
eval_interval: 1
epochs: 20
print_batch_step: 10
use_visualdl: False
# used for static mode and model export
image_shape: [3, 224, 224]
save_inference_dir: ./inference
# training model under @to_static
to_static: False
use_dali: False
# model architecture
Arch:
name: "DistillationModel"
class_num: &class_num 2
# if not null, its lengths should be same as models
pretrained_list:
# if not null, its lengths should be same as models
freeze_params_list:
- True
- False
use_sync_bn: True
models:
- Teacher:
name: ResNet101_vd
class_num: *class_num
- Student:
name: PPLCNet_x1_0
class_num: *class_num
pretrained: True
use_ssld: True
infer_model_name: "Student"
# loss function config for traing/eval process
Loss:
Train:
- DistillationDMLLoss:
weight: 1.0
model_name_pairs:
- ["Student", "Teacher"]
Eval:
- CELoss:
weight: 1.0
Optimizer:
name: Momentum
momentum: 0.9
lr:
name: Cosine
learning_rate: 0.01
warmup_epoch: 5
regularizer:
name: 'L2'
coeff: 0.00004
# data loader for train and eval
DataLoader:
Train:
dataset:
name: ImageNetDataset
image_root: ./dataset/car_exists/
cls_label_path: ./dataset/car_exists/train_list_for_distill.txt
transform_ops:
- DecodeImage:
to_rgb: True
channel_first: False
- RandCropImage:
size: 192
- RandFlipImage:
flip_code: 1
- TimmAutoAugment:
prob: 0.0
config_str: rand-m9-mstd0.5-inc1
interpolation: bicubic
img_size: 192
- NormalizeImage:
scale: 1.0/255.0
mean: [0.485, 0.456, 0.406]
std: [0.229, 0.224, 0.225]
order: ''
- RandomErasing:
EPSILON: 0.1
sl: 0.02
sh: 1.0/3.0
r1: 0.3
attempt: 10
use_log_aspect: True
mode: pixel
sampler:
name: DistributedBatchSampler
batch_size: 64
drop_last: False
shuffle: True
loader:
num_workers: 16
use_shared_memory: True
Eval:
dataset:
name: ImageNetDataset
image_root: ./dataset/car_exists/
cls_label_path: ./dataset/car_exists/val_list.txt
transform_ops:
- DecodeImage:
to_rgb: True
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: ''
sampler:
name: DistributedBatchSampler
batch_size: 64
drop_last: False
shuffle: False
loader:
num_workers: 4
use_shared_memory: True
Infer:
infer_imgs: deploy/images/PULC/car_exists/objects365_00001507.jpeg
batch_size: 10
transforms:
- DecodeImage:
to_rgb: True
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:
PostProcess:
name: ThreshOutput
threshold: 0.5
label_0: nobody
label_1: someone
Metric:
Train:
- DistillationTopkAcc:
model_key: "Student"
topk: [1, 2]
Eval:
- TprAtFpr:
max_fpr: 0.01
- TopkAcc:
topk: [1, 2]
ppcls/configs/PULC/car_exists/PPLCNet_x1_0_search.yaml 0 → 100644
浏览文件 @ a3c154fa
# global configs
Global:
checkpoints: null
pretrained_model: null
output_dir: ./output/
device: gpu
save_interval: 1
eval_during_train: True
eval_interval: 1
start_eval_epoch: 10
epochs: 20
print_batch_step: 10
use_visualdl: False
# used for static mode and model export
image_shape: [3, 224, 224]
save_inference_dir: ./inference
# training model under @to_static
to_static: False
use_dali: False
# model architecture
Arch:
name: PPLCNet_x1_0
class_num: 2
pretrained: True
use_ssld: True
use_sync_bn: True
# loss function config for traing/eval process
Loss:
Train:
- CELoss:
weight: 1.0
Eval:
- CELoss:
weight: 1.0
Optimizer:
name: Momentum
momentum: 0.9
lr:
name: Cosine
learning_rate: 0.01
warmup_epoch: 5
regularizer:
name: 'L2'
coeff: 0.00004
# data loader for train and eval
DataLoader:
Train:
dataset:
name: ImageNetDataset
image_root: ./dataset/car_exists/
cls_label_path: ./dataset/car_exists/train_list.txt
transform_ops:
- DecodeImage:
to_rgb: True
channel_first: False
- RandCropImage:
size: 224
- RandFlipImage:
flip_code: 1
- TimmAutoAugment:
prob: 0.0
config_str: rand-m9-mstd0.5-inc1
interpolation: bicubic
img_size: 224
- NormalizeImage:
scale: 1.0/255.0
mean: [0.485, 0.456, 0.406]
std: [0.229, 0.224, 0.225]
order: ''
- RandomErasing:
EPSILON: 0.0
sl: 0.02
sh: 1.0/3.0
r1: 0.3
attempt: 10
use_log_aspect: True
mode: pixel
sampler:
name: DistributedBatchSampler
batch_size: 64
drop_last: False
shuffle: True
loader:
num_workers: 8
use_shared_memory: True
Eval:
dataset:
name: ImageNetDataset
image_root: ./dataset/car_exists/
cls_label_path: ./dataset/car_exists/val_list.txt
transform_ops:
- DecodeImage:
to_rgb: True
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: ''
sampler:
name: DistributedBatchSampler
batch_size: 64
drop_last: False
shuffle: False
loader:
num_workers: 4
use_shared_memory: True
Infer:
infer_imgs: deploy/images/PULC/car_exists/objects365_00001507.jpeg
batch_size: 10
transforms:
- DecodeImage:
to_rgb: True
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:
PostProcess:
name: ThreshOutput
threshold: 0.5
label_0: nobody
label_1: someone
Metric:
Train:
- TopkAcc:
topk: [1, 2]
Eval:
- TprAtFpr:
max_fpr: 0.01
- TopkAcc:
topk: [1, 2]
ppcls/configs/PULC/car_exists/SwinTransformer_tiny_patch4_window7_224.yaml 0 → 100644
浏览文件 @ a3c154fa
# global configs
Global:
checkpoints: null
pretrained_model: null
output_dir: ./output/
device: gpu
save_interval: 1
eval_during_train: True
eval_interval: 1
start_eval_epoch: 10
epochs: 20
print_batch_step: 10
use_visualdl: False
# used for static mode and model export
image_shape: [3, 224, 224]
save_inference_dir: ./inference
# training model under @to_static
to_static: False
use_dali: False
# mixed precision training
AMP:
scale_loss: 128.0
use_dynamic_loss_scaling: True
# O1: mixed fp16
level: O1
# model architecture
Arch:
name: SwinTransformer_tiny_patch4_window7_224
class_num: 2
pretrained: True
# loss function config for traing/eval process
Loss:
Train:
- CELoss:
weight: 1.0
epsilon: 0.1
Eval:
- CELoss:
weight: 1.0
Optimizer:
name: AdamW
beta1: 0.9
beta2: 0.999
epsilon: 1e-8
weight_decay: 0.05
no_weight_decay_name: absolute_pos_embed relative_position_bias_table .bias norm
one_dim_param_no_weight_decay: True
lr:
name: Cosine
learning_rate: 1e-4
eta_min: 2e-6
warmup_epoch: 5
warmup_start_lr: 2e-7
# data loader for train and eval
DataLoader:
Train:
dataset:
name: ImageNetDataset
image_root: ./dataset/car_exists/
cls_label_path: ./dataset/car_exists/train_list.txt
transform_ops:
- DecodeImage:
to_rgb: True
channel_first: False
- RandCropImage:
size: 224
interpolation: bicubic
backend: pil
- RandFlipImage:
flip_code: 1
- TimmAutoAugment:
config_str: rand-m9-mstd0.5-inc1
interpolation: bicubic
img_size: 224
- NormalizeImage:
scale: 1.0/255.0
mean: [0.485, 0.456, 0.406]
std: [0.229, 0.224, 0.225]
order: ''
- RandomErasing:
EPSILON: 0.25
sl: 0.02
sh: 1.0/3.0
r1: 0.3
attempt: 10
use_log_aspect: True
mode: pixel
batch_transform_ops:
- OpSampler:
MixupOperator:
alpha: 0.8
prob: 0.5
CutmixOperator:
alpha: 1.0
prob: 0.5
sampler:
name: DistributedBatchSampler
batch_size: 128
drop_last: False
shuffle: True
loader:
num_workers: 8
use_shared_memory: True
Eval:
dataset:
name: ImageNetDataset
image_root: ./dataset/car_exists/
cls_label_path: ./dataset/car_exists/val_list.txt
transform_ops:
- DecodeImage:
to_rgb: True
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: ''
sampler:
name: DistributedBatchSampler
batch_size: 64
drop_last: False
shuffle: False
loader:
num_workers: 8
use_shared_memory: True
Infer:
infer_imgs: deploy/images/PULC/car_exists/objects365_00001507.jpeg
batch_size: 10
transforms:
- DecodeImage:
to_rgb: True
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:
PostProcess:
name: ThreshOutput
threshold: 0.5
label_0: nobody
label_1: someone
Metric:
Train:
- TopkAcc:
topk: [1, 2]
Eval:
- TprAtFpr:
max_fpr: 0.01
- TopkAcc:
topk: [1, 2]
ppcls/configs/PULC/car_exists/search.yaml 0 → 100644
浏览文件 @ a3c154fa
base_config_file: ppcls/configs/PULC/person_exists/PPLCNet_x1_0_search.yaml
distill_config_file: ppcls/configs/PULC/person_exists/PPLCNet_x1_0_distillation.yaml
gpus: 0,1,2,3
output_dir: output/search_person_cls
search_times: 1
search_dict:
- search_key: lrs
replace_config:
- Optimizer.lr.learning_rate
search_values: [0.0075, 0.01, 0.0125]
- search_key: resolutions
replace_config:
- DataLoader.Train.dataset.transform_ops.1.RandCropImage.size
- DataLoader.Train.dataset.transform_ops.3.TimmAutoAugment.img_size
search_values: [176, 192, 224]
- search_key: ra_probs
replace_config:
- DataLoader.Train.dataset.transform_ops.3.TimmAutoAugment.prob
search_values: [0.0, 0.1, 0.5]
- search_key: re_probs
replace_config:
- DataLoader.Train.dataset.transform_ops.5.RandomErasing.EPSILON
search_values: [0.0, 0.1, 0.5]
- search_key: lr_mult_list
replace_config:
- Arch.lr_mult_list
search_values:
- [0.0, 0.2, 0.4, 0.6, 0.8, 1.0]
- [0.0, 0.4, 0.4, 0.8, 0.8, 1.0]
- [1.0, 1.0, 1.0, 1.0, 1.0, 1.0]
teacher:
rm_keys:
- Arch.lr_mult_list
search_values:
- ResNet101_vd
- ResNet50_vd
final_replace:
Arch.lr_mult_list: Arch.models.1.Student.lr_mult_list
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