cannot reproduce comparable results in the paper

[felix-yuxiang]

Hi, I ran the code with a single GPU NVIDIA GeForce RTX 3090 with the given config file listed in the paper. Here is my reproduced result which is significantly different from the results provided in README.md file. Can you guide me through and specify what could be the issue? Can you provide more info on how to train a model with the same performance of your pre-trained model you provided in /checkpoints. Any help will be appreciated.

[Pranav-chib]

I also have the similar result while reproducing the training.

[fangzl123]

I think it's due to the hyperparameters setting. In the paper it's mentioned "With a frozen denoising module, we then train the leapfrog initializer for 200 epochs with an initial learning rate of 10−4, decaying by 0.9 every 32 epochs", but in the default led_augment.yml it is not like this.

Set them based on the paper and I've got

[Frank-Star-fn]

I think it's due to the hyperparameters setting. In the paper it's mentioned "With a frozen denoising module, we then train the leapfrog initializer for 200 epochs with an initial learning rate of 10−4, decaying by 0.9 every 32 epochs", but in the default led_augment.yml it is not like this.

Set them based on the paper and I've got

Hello, did you use the pre trained model provided by him for the diffusion model in the first stage, or did you train yourself for one stage according to the settings in the paper?

[fangzl123]

I think it's due to the hyperparameters setting. In the paper it's mentioned "With a frozen denoising module, we then train the leapfrog initializer for 200 epochs with an initial learning rate of 10−4, decaying by 0.9 every 32 epochs", but in the default led_augment.yml it is not like this.
Set them based on the paper and I've got

Hello, did you use the pre trained model provided by him for the diffusion model in the first stage, or did you train yourself for one stage according to the settings in the paper?

Hi, I use the provided pre-trained model as the first stage.

[fangzl123]

Hi, I double checked my training configuration, and found that my learning rate setting is 1e-3 (same with the default code setting). So, I think you can try once with lr=1e-3 decaying by 0.9 every 32 epochs? The hyperparameter settings in this repo are quite confusing, some parts are consistent with the paper but some are not... Best,

…

On Wed, 22 Nov 2023 at 19:08, Frank Star ***@***.***> wrote: I think it's due to the hyperparameters setting. In the paper it's mentioned "With a frozen denoising module, we then train the leapfrog initializer for 200 epochs with an initial learning rate of 10−4, decaying by 0.9 every 32 epochs", but in the default led_augment.yml it is not like this. Set them based on the paper and I've got [image: image] <https://user-images.githubusercontent.com/58146879/278015992-f6145560-0ebe-441c-8bd2-846dff8b996d.png> Hello, did you use the pre trained model provided by him for the diffusion model in the first stage, or did you train yourself for one stage according to the settings in the paper? Hi, I use the provided pre-trained model as the first stage. Hello, I used the same hyperparameter settings as in the paper (200 epochs with an initial learning rate of 10−4, decaying by 0.9 every 32 epochs) on the RTX4090 server, but got [image: 1111] <https://user-images.githubusercontent.com/72490620/284882130-3917c2a5-48a6-475b-ab3e-04160f66aaa1.png> This result is much lower than the results in the paper, and there is a slight improvement after increasing the epoch to 400, but it is still not as good as the results in the paper [image: 400epo] <https://user-images.githubusercontent.com/72490620/284882634-6a129d5e-c8de-412c-a4f8-e57f49fa9fb5.png> Do you have any clue about this? What do you think is the reason for the inability to reproduce? — Reply to this email directly, view it on GitHub <#5 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AN3UAP72PZKJLV4TCMVIEX3YFXML3AVCNFSM6AAAAAA2O3BXUGVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMYTQMRSGU3DIMBZGU> . You are receiving this because you commented.Message ID: ***@***.***>

[Pranav-chib]

I think 0.83/1.69 was the only reproduced result

[felix-yuxiang]

Now, I am able to reproduce their stageone and LED stagetwo results. The answer from @woyoudian2gou helped me a lot. But I would say it requires an non-trivial amount of engineering work to tune this well.

[Pranav-chib]

Could you share with us some insight, it would me helpful.

[kkk00714]

Now, I am able to reproduce their stageone and LED stagetwo results. The answer from @woyoudian2gou helped me a lot. But I would say it requires an non-trivial amount of engineering work to tune this well.

Yes, and the whole implementation is difficult to explain, I think the original
author may have used a different method to get the pre-trained model.

[ZY-Ren]

@woyoudian2gou Hi, I have implemented your mentioned hyperparameters setting, but still can't get a reasonable result. So could you share your config.yml with us? Thank you very much.

[kkk00714]

@woyoudian2gou Hi, I have implemented your mentioned hyperparameters setting, but still can't get a reasonable result. So could you share your config.yml with us? Thank you very much.

See https://github.com/MediaBrain-SJTU/LED/issues/6

[ZY-Ren]

@kkk00714 Thank you for your prompt reply, I would also like to know the hyperparameters of Phase 2 training, could you share that? I would appreciate it

[kkk00714]

@kkk00714 Thank you for your prompt reply, I would also like to know the hyperparameters of Phase 2 training, could you share that? I would appreciate it

The hyperparameters of stage 2 are same as original implement of author (batchsiaze = 10, lr = 1e-4...).

[Nighttell]

@kkk00714
Hello, I have reproduced the first and second stages of training according to your method. Although the final results are similar to those given in MD, the loss is
[2024-08-05 04:31:00] Epoch: 99 Loss: 5.839539 Loss Dist.: 5.320199 Loss Uncertainty: 0.519340
I think its LOS is too high, or is this normal?
The results of running the test set in the 100th and 75th epochs are similar
[Initialization Model] Trainable/Total: 4634721/4634721
./results/led_augment/try2/models/model_0100.p
--ADE(1s): 0.1775 --FDE(1s): 0.2671
--ADE(2s): 0.3698 --FDE(2s): 0.5612
--ADE(3s): 0.5799 --FDE(3s): 0.8323
--ADE(4s): 0.8115 --FDE(4s): 1.1440
[Core Denoising Model] Trainable/Total: 6568720/6568720
[Initialization Model] Trainable/Total: 4634721/4634721
./results/led_augment/try2/models/model_0075.p
--ADE(1s): 0.1774 --FDE(1s): 0.2671
--ADE(2s): 0.3698 --FDE(2s): 0.5612
--ADE(3s): 0.5799 --FDE(3s): 0.8323
--ADE(4s): 0.8115 --FDE(4s): 1.1437
Is it due to the influence of batch_2? I used a 4070TI 12G for training. The batch size is set to 6 (I have tried training with the default settings and also tried training with the parameters given in the paper, but the results are not significantly different)
Look forward to your reply! thank

[kkk00714]

@kkk00714 Hello, I have reproduced the first and second stages of training according to your method. Although the final results are similar to those given in MD, the loss is [2024-08-05 04:31:00] Epoch: 99 Loss: 5.839539 Loss Dist.: 5.320199 Loss Uncertainty: 0.519340 I think its LOS is too high, or is this normal? The results of running the test set in the 100th and 75th epochs are similar [Initialization Model] Trainable/Total: 4634721/4634721 ./results/led_augment/try2/models/model_0100.p --ADE(1s): 0.1775 --FDE(1s): 0.2671 --ADE(2s): 0.3698 --FDE(2s): 0.5612 --ADE(3s): 0.5799 --FDE(3s): 0.8323 --ADE(4s): 0.8115 --FDE(4s): 1.1440 [Core Denoising Model] Trainable/Total: 6568720/6568720 [Initialization Model] Trainable/Total: 4634721/4634721 ./results/led_augment/try2/models/model_0075.p --ADE(1s): 0.1774 --FDE(1s): 0.2671 --ADE(2s): 0.3698 --FDE(2s): 0.5612 --ADE(3s): 0.5799 --FDE(3s): 0.8323 --ADE(4s): 0.8115 --FDE(4s): 1.1437 Is it due to the influence of batch_2? I used a 4070TI 12G for training. The batch size is set to 6 (I have tried training with the default settings and also tried training with the parameters given in the paper, but the results are not significantly different) Look forward to your reply! thank

Your loss is normal because the author multiplies a coefficient in front of loss_dist to give it a greater weight than loss_uncertainty. You can check it in the paper or code.

[Nighttell]

@kkk00714 Hello, I have reproduced the first and second stages of training according to your method. Although the final results are similar to those given in MD, the loss is [2024-08-05 04:31:00] Epoch: 99 Loss: 5.839539 Loss Dist.: 5.320199 Loss Uncertainty: 0.519340 I think its LOS is too high, or is this normal? The results of running the test set in the 100th and 75th epochs are similar [Initialization Model] Trainable/Total: 4634721/4634721 ./results/led_augment/try2/models/model_0100.p --ADE(1s): 0.1775 --FDE(1s): 0.2671 --ADE(2s): 0.3698 --FDE(2s): 0.5612 --ADE(3s): 0.5799 --FDE(3s): 0.8323 --ADE(4s): 0.8115 --FDE(4s): 1.1440 [Core Denoising Model] Trainable/Total: 6568720/6568720 [Initialization Model] Trainable/Total: 4634721/4634721 ./results/led_augment/try2/models/model_0075.p --ADE(1s): 0.1774 --FDE(1s): 0.2671 --ADE(2s): 0.3698 --FDE(2s): 0.5612 --ADE(3s): 0.5799 --FDE(3s): 0.8323 --ADE(4s): 0.8115 --FDE(4s): 1.1437 Is it due to the influence of batch_2? I used a 4070TI 12G for training. The batch size is set to 6 (I have tried training with the default settings and also tried training with the parameters given in the paper, but the results are not significantly different) Look forward to your reply! thank

Your loss is normal because the author multiplies a coefficient in front of loss_dist to give it a greater weight than loss_uncertainty. You can check it in the paper or code.

Thank you for your answer. This is the first time I have encountered such a loss handling method