QualityInspector无监督学习,padim_resnet18_mvtec训练后对good图做预测结果也很差.

[David-dotcom666]

问题确认 Search before asking

• 我已经查询历史issue(包括open与closed)，没有发现相似的bug。I have searched the open and closed issues and found no similar bug report.

Bug描述 Describe the Bug

配置如下:

common arguments

device: gpu
seed: 3

dataset arguments

batch_size: 1
num_workers: 0
category: bottle
resize: [256, 256]
crop_size: [224, 224]
data_path: data/mvtec_anomaly_detection
save_path: output/

train arguments

do_eval: False # After training, whether to eval model;
backbone: resnet18 # Support resnet18, resnet50, wide_resnet50_2

val and predict arguments

save_pic: True # Whether to save one output example picture in val and predict;
model_path: output/resnet18/bottle/bottle.pdparams

predict arguments

img_path: data/mvtec_anomaly_detection/bottle/test/good/001.png
threshold: 0.5
另外除了这个，另外两种算法都无法运行.

复现环境 Environment

paddle 2.6.1
paddleseg 2.8
cuda 11.7,cudnn 8.4.1

Bug描述确认 Bug description confirmation

• 我确认已经提供了Bug复现步骤、代码改动说明、以及环境信息，确认问题是可以复现的。I confirm that the bug replication steps, code change instructions, and environment information have been provided, and the problem can be reproduced.

是否愿意提交PR？ Are you willing to submit a PR?

• 我愿意提交PR！I'd like to help by submitting a PR!