rentainhe / TRAR-VQA

[ICCV 2021] Official implementation of the paper "TRAR: Routing the Attention Spans in Transformers for Visual Question Answering"
MIT License
64 stars 18 forks source link

关于精度 #6

Open Whhhzc opened 2 months ago

Whhhzc commented 2 months ago

你好 我想请问一下 为什么划分数据集为'train+val', 'train+val+vg' 时,模型的输出都是nowTime: 2024-07-05 16:33:36 Epoch: 10, Loss: 2.6711035106781544, Lr: 0.0001 Elapsed time: 2118, Speed(s/batch): 0.11792277398014418

而没有显示类似于这样的精度Overall Accuracy is: 59.30 other : 51.37 yes/no : 76.85 number : 38.89

rentainhe commented 1 month ago

你好 我想请问一下 为什么划分数据集为'train+val', 'train+val+vg' 时,模型的输出都是nowTime: 2024-07-05 16:33:36 Epoch: 10, Loss: 2.6711035106781544, Lr: 0.0001 Elapsed time: 2118, Speed(s/batch): 0.11792277398014418

而没有显示类似于这样的精度Overall Accuracy is: 59.30 other : 51.37 yes/no : 76.85 number : 38.89

使用 train + val 或者 train + val + vg的时候需要把测试结果提交到 eval test集的网站,我们的codebase是基于open-vqa的,其只包含了对val的训练时测试