Why did I not train well according to your method?

[zsh1223]

what is the difference between mono_fm and mono_fm-joint?

[Jamesgender]

Hello, can you tell me where are the training weights? I cant find them while training.

[zsh1223]

你好，请问哪里有训练砝码？我在训练时找不到它们。

Hello, can you tell me where are the training weights? I cant find them while training.

You can find the pre trained weight here:
https://drive . Google . com/file/d/1 bimxnb 9 pegv 3 mzczb 3 UBW 3 EDI 4d pocxf/view

[Jamesgender]

Sorry, I mean the training weights not pretained weights. The common situation is that when you finished one epoch, the model evaluates on the val dataset and saves a weight. 1 2 3 and 20. But I cant find the weights saveing folder while training. Where the weights save?

[zsh1223]

Sorry, I mean the training weights not pretained weights. The common situation is that when you finished one epoch, the model evaluates on the val dataset and saves a weight. 1 2 3 and 20. But I cant find the weights saveing folder while training. Where the weights save?

You can set the path where you want to save the training model in run. py. In author run.py,--work_dir results.That means,your finished one epoch is in the /GCNDepth/results.
I hope my answer is helpful to you

[Jamesgender]

Oh, I know my mistake. The first epoch is epoch1 rather than epoch0. So the model actually have not finished 1 epoch and saved weights. Thanks for your timely reply.

[Jamesgender]

I trained one epoch and got similar results to yours. Then the training was interrupted.

The first epoch's loss seems to be abnormal. When I trained MonoDepth2, the loss decreased rapidly. Have you ever met the same problem?

[lh-study]

Hello, may I ask how your mmcv=0.4.4 was installed? I have encountered the same problem as you, which is that the generated depth map is completely black, but I did not reproduce it. Instead, I used the author's decoder to modify and run it. So, how was your mmcv=0.4.4 installed?

what is the difference between mono_fm and mono_fm-joint?

[ArminMasoumian]

Hello, may I ask how your mmcv=0.4.4 was installed? I have encountered the same problem as you, which is that the generated depth map is completely black, but I did not reproduce it. Instead, I used the author's decoder to modify and run it. So, how was your mmcv=0.4.4 installed?

what is the difference between mono_fm and mono_fm-joint?

You can download the mmcv through this link:
https://pypi.org/project/mmcv/0.4.4/

[zsh1223]

我训练了一个纪元，得到了与你相似的结果。然后训练被打断了。 第一个纪元的损失似乎是不正常的。当我训练 MonoDepth2 时，损失迅速下降。你有没有遇到过同样的问题？

May I ask if you have successfully reproduced this project. In the process of reproduction，the evaluation results of the model after training for 15 epochs are consistent with the epochs_1.

[Jamesgender]

我训练了一个纪元，得到了与你相似的结果。然后训练被打断了。 第一个纪元的损失似乎是不正常的。当我训练 MonoDepth2 时，损失迅速下降。你有没有遇到过同样的问题？

May I ask if you have successfully reproduced this project. In the process of reproduction，the evaluation results of the model after training for 15 epochs are consistent with the epochs_1.

My results are same as yours. The loss is abnormal and consistent with the first epoch 1.