使用PaddleClas训练的分类模型在pphuman中准确率极低

[cskkx1]

问题确认 Search before asking

• 我已经搜索过问题，但是没有找到解答。I have searched the question and found no related answer.

请提出你的问题 Please ask your question

我在PaddleClas中使用safety_helmet/PPLCNet_x1_0.yaml 配置文件训练自己的数据集（一共两类），

global configs

Global:
checkpoints: null
pretrained_model: null
output_dir: ./output/
device: gpu
save_interval: 1
eval_during_train: True
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

model architecture

Arch:
name: PPLCNet_x1_0
pretrained: True
use_ssld: True
class_num: 2
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: 0.025
warmup_epoch: 5
regularizer:
name: 'L2'
coeff: 0.00003

data loader for train and eval

DataLoader:
Train:
dataset:
name: ImageNetDataset
image_root: ./dataset/normal_sleep
cls_label_path: ./dataset/normal_sleep/train_list.txt
transform_ops:
- DecodeImage:
to_rgb: True
channel_first: False
- RandCropImage:
size: 176
- RandFlipImage:
flip_code: 1
- TimmAutoAugment:
prob : 0.5
config_str: rand-m9-mstd0.5-inc1
interpolation: bicubic
img_size : 176
- 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
r1 : 0.3
sh : 1.0/3.0
sl : 0.02
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/normal_sleep
cls_label_path: ./dataset/normal_sleep/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: sleep3.jpg
batch_size: 1
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: sleep
label_1: normal

Metric:
Train:
- TopkAcc:
topk: [1]
Eval:
- TprAtFpr:
max_fpr: 0.0001
- TopkAcc:
topk: [1]

直接在PaddleClas中使用python tools/train.py -c ppcls/configs/PULC/safety_helmet/PPLCNet_x1_0.yaml -o Arch.pretrained=True进行训练，因为图片较少，所以只训练了20轮top1就很高了，直接使用训练获得的best model进行预测python3 tools/infer.py -c ./ppcls/configs/PULC/safety_helmet/PPLCNet_x1_0.yaml -o Global.pretrained_model=output/best_model，图片为配置文件中的默认图片，结果是正确的，但是当我将这个模型导出python tools/export_model.py -c ppcls/configs/PULC/safety_helmet/PPLCNet_x1_0.yaml -o weights=output/best_model，再应用到pphuman中时，修改配置文件

使用pipeline进行测试，效果很差，而且测试用的视频就是训练集，但仍然预测不对。

[TingquanGao]

导出后的模型，使用PaddleClas的deploy进行推理预测，结果正确吗？