指标问题

[largerwxt]

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

您好，我在训练ppmatting时，指标到了300左右就不降了，最后的结果也很差，请问我应该怎么解决呢

[Stinky-Tofu]

@largerwxt 这个问题太笼统了，请自行分析数据集标注、学习率等是否有问题

[largerwxt]

您好，我已经检查过数据集了，我使用的Distinctions-646数据集，并且安装官方文档准备好了。我的config如下batch_size: 4
iters: 300000

train_dataset:
type: MattingDataset
dataset_root: /home/wxt/pp-matting/work/PaddleSeg/Matting/data/Distinctions-646/train
train_file: train.txt
transforms:
- type: LoadImages
- type: Padding
target_size: [512, 512]
- type: ResizeByShort
short_size: 512
- type: RandomCrop
crop_size: [[512, 512],[640, 640], [800, 800]]
- type: Resize
target_size: [512, 512]
- type: RandomDistort
- type: RandomBlur
prob: 0.1
- type: RandomHorizontalFlip
- type: Normalize
mode: train
separator: '|'

val_dataset:
type: MattingDataset
dataset_root: /home/wxt/pp-matting/work/PaddleSeg/Matting/data/Distinctions-646/test
val_file: test.txt
transforms:
- type: LoadImages
- type: LimitShort
max_short: 1536
- type: ResizeToIntMult
mult_int: 32
- type: Normalize
mode: val
get_trimap: False
separator: '|'

model:
type: PPMatting
backbone:
type: HRNet_W48
pretrained: https://bj.bcebos.com/paddleseg/dygraph/hrnet_w48_ssld.tar.gz
pretrained: Null

optimizer:
type: sgd
momentum: 0.9
weight_decay: 4.0e-5

lr_scheduler:
type: PolynomialDecay
learning_rate: 0.01
end_lr: 0
power: 0.9
但是在训练时，语义损失一直不下去，并且sad在200左右就不下降了。请问这是什么原因呢