Thanks for sharing your code! I try to reproduce the results in the paper and I failed.

[yangpan22223]

Thanks for sharing your code! I try to reproduce the results in the paper and I failed.
I have a few questions to ask:

1. How many GPUs do you use?
2. What's your GPU configuration?
3. What is the batch size?
4. Does training from scratch need to load a pretrained model?
5. Can you provide your training log?

[XINRUII]

Thank you for your interest in our work~
As for 1.2.3. We use a single NVIDIA RTX 2080Ti to train our network and the batch size is 8.
For 4. It does not.
For 5. We will upload our training log soon.

[Yeah2333]

I have run this code with default config provided by authors. After 100 epoch, the RPE is
`2022-06-01 11:16:06.565856
---- EPOCH 103 EVALUATION ----
---- batch 000 in evaluation ----
---- batch 100 in evaluation ----
seq07 Average_t_error 2.21 Average_r_error 1.42

2022-06-01 11:17:23.734339
---- EPOCH 103 EVALUATION ----
---- batch 000 in evaluation ----
---- batch 100 in evaluation ----
---- batch 200 in evaluation ----
---- batch 300 in evaluation ----
---- batch 400 in evaluation ----
---- batch 500 in evaluation ----
seq08 Average_t_error 2.82 Average_r_error 1.35

2022-06-01 11:22:05.404721
---- EPOCH 103 EVALUATION ----
---- batch 000 in evaluation ----
---- batch 100 in evaluation ----
seq09 Average_t_error 2.07 Average_r_error 0.85

2022-06-01 11:23:51.822465
---- EPOCH 103 EVALUATION ----
---- batch 000 in evaluation ----
---- batch 100 in evaluation ----
seq10 Average_t_error 2.38 Average_r_error 2.09`

[Joechencc]

It works if you train from scratch, but using their pretrained model. It diverges after L2

[JunyuanDeng]

I have run this code with default config provided by authors. After 100 epoch, the RPE is `2022-06-01 11:16:06.565856 ---- EPOCH 103 EVALUATION ---- ---- batch 000 in evaluation ---- ---- batch 100 in evaluation ---- seq07 Average_t_error 2.21 Average_r_error 1.42

2022-06-01 11:17:23.734339 ---- EPOCH 103 EVALUATION ---- ---- batch 000 in evaluation ---- ---- batch 100 in evaluation ---- ---- batch 200 in evaluation ---- ---- batch 300 in evaluation ---- ---- batch 400 in evaluation ---- ---- batch 500 in evaluation ---- seq08 Average_t_error 2.82 Average_r_error 1.35

2022-06-01 11:22:05.404721 ---- EPOCH 103 EVALUATION ---- ---- batch 000 in evaluation ---- ---- batch 100 in evaluation ---- seq09 Average_t_error 2.07 Average_r_error 0.85

2022-06-01 11:23:51.822465 ---- EPOCH 103 EVALUATION ---- ---- batch 000 in evaluation ---- ---- batch 100 in evaluation ---- seq10 Average_t_error 2.38 Average_r_error 2.09`

@Yeah2333 Did you use the pre-trained model or train from scratch? I use the pre-trained model, the result seems corresponding to the article.

[tom-bu]

Which file is their pre-trained model? I tried all of them and none of them can be loaded.

[Joechencc]

We do not use any pertained model, DeepMapping and DeepMapping2 are designed to train from scratch. But we are working on using the pertrained model to speed up the process on our next work.

[tom-bu]

@Joechencc Are you responding to the right repo? This if for PWCLONet

[tom-bu]

I realized the way to load the pretrained model is like this: --checkpoint_path pretrained_model/pretrained_model.ckpt
tensorflow/models#2675

[Joechencc]

Sorry. Reply to wrong repo..