Junyu Gao

Results 83 comments of Junyu Gao

I does not conduct the reproduction of CSRNet on WorldExpo. However, I train other pretrained models on it, such as ResNet, VGG, deeplab v3 and so on. Their results can...

@liwenxi Thanks for your summary!

@all Please visit this issue: https://github.com/gjy3035/C-3-Framework/issues/22#issuecomment-494287652

有人说,选取不同epoch的模型,然后取在不同场景的最优值,才能复现到8.6.这个我就不清楚了。我自己的实验中,都是单一模型在avg上最优这种取法。当然结果离SOTA很远。

抱歉,我这里没有Win的环境可以测试。或许,需要将loading_data.py中的num_workers改为0.

- [ ] 部分数据路径的一致性修改。

- [ ] **是否**替换csv或者mat的gt存储方式,改为pgm, tiff等图像格式以增加读取速度。

- [ ] **是否**提供数据预存储到内存中,以增加训练速度。

- [x] 增加断点训练功能,增加optimizer状态保存及加载。 (目前完成trainer和trainer_for_M2TCC的断点训练。其他trainer在完成其他相应模型后更新。)

C3F's experiments are on small-scale datasets. The overfitting problem is caused by a small data volume. If you want to generalize other images (web or your own), you should select...