KKKSQJ / DeepLearning

A deep learning code base, mainly for paper replication, in the areas of image recognition, object detection, image segmentation, self-supervision, etc. Each project can be run independently, and there are corresponding articles to explain.
MIT License
313 stars 97 forks source link

关于用IP102数据集训练遇到的问题 #5

Closed Aoshika123 closed 1 year ago

Aoshika123 commented 1 year ago

您好作者,我在用TransFG跑IP102数据集的时候,在训练阶段loss一直保持在4点多不变,Valid Accuracy只有百分之0.4,经过多轮训练一直不变化,请问您有遇到过这类的问题么?

KKKSQJ commented 1 year ago

您好,我没有试过训练IP102数据集。您可以自己定义数据集的读取方式以适应您的数据集,并且确保您的数据、参数配置等信息正确。您可以在https://github.com/KKKSQJ/DeepLearning/tree/master/classification/TransFG/logs 找到我训练CUB200数据集的全部配置信息以及log日志。