search

QData / C-Tran

General Multi-label Image Classification with Transformers
MIT License
244 stars 43 forks source link

Weird same prediction with large dataset #15

Closed shinu335 closed 2 years ago

shinu335 commented 2 years ago

Thank you for your wonderful project!

I have been doing research with your C-Tran project, and conducting training with several autonomous driving dataset, such as nuScenes and METEOR.

It seems that when the number of images is larger than 100,000, all of the images tend to have almost same predictions. For example, [car: 0.96, pedestrian: 0.01, truck: 0.48] for all of the images, even though labels of different images are different. However, when I shrink the number of images to 3000 or less, different images will have different predictions of labels.

May I ask why this problem happens? Is it because I failed to apply your network correctly, or because the network with Transformers struggled to output satisfying predictions with large data?

Thank you for your reading, and I look forward to your reply.

jacklanchantin commented 2 years ago

Are you seeing the same results on the datasets from our experiments? If not, it's likely something wrong with your data.

shinu335 commented 2 years ago

Are you seeing the same results on the datasets from our experiments? If not, it's likely something wrong with your data.

Thank you for your reply.

I have not conducted training with datasets from your experiments.

With regards to my data, including nuScenes and METEOR, weird same results are output only when number of data is large. It can output normal predictions when number of data is small.

Anyway I am going to check my experiment again. Thank you again for your reply.