code and paper about md4all HOT 5 CLOSED

Hi @RuiTianHIT,

Thank you for your interest in our work. We are glad you could reproduce our method.

The console output you shared indicates the number of scenes in the dataset used for training and validation. It does not show the number of images (which are much more), and it does not include the translated images. As intended, the number of original night scenes is 0. This is because the night inputs we use during training are only the translated ones. Specifically, the exact number of translated inputs depends on the randomness of each iteration (according to x%), so it can only be estimated, and we do not print it out. Nevertheless, the amount of available translated scenes is the same as the number of scenes used for training (i.e., train-day-clear).

As we reported in issue #8, the random mix of translated and original easy inputs is regulated by the x%, and in the code that happens in the function linked here. The amount is specified in the configuration files (e.g., here).

We hope this helps clear up your doubts. Feel free to ask us follow-up questions.

from md4all.

@sgasperini @morbi25 Thank you very much for your reply. We will study your work according to your help！

from md4all.

@sgasperini @morbi25 We are working on your code now, but there are a few problems, and the attached debug procedure shows our confusion
(1) During the training of md4all(distillation stage), we found that is_train=False, which is shown in the attachment
(2) At the same time, every input is day-clear, why is there no image representation after mix
(3) In the attachment, we also record our configuration file, which should be the input based on probability



Looking forward to your reply, thank you very much！

We set a breakpoint in the process of running the mix, but the breakpoint has been running for a long time without entering it. We are very confused about this problem and look forward to your reply. We are very interested in this work.

from md4all.

Hi @RuiTianHIT,

From your screenshots, it looks like you are debugging a validation step and not a training step.

You are seeing is_train set to False without any translated input, likely due to a default feature of PyTorch Lightning (PL), that executes a few validation iterations before the actual training begins. You can prevent this behavior via the PL Trainer parameters or wait for it to end.

Once the initial validation checks are done, the training steps should become more apparent, and you should see the mix of original and translated inputs that you are referring to. In PyCharm, you can set breakpoints to activate under specific conditions (via a right-click on the breakpoint), e.g., to focus on the training process.

We hope this helps explain things.

from md4all.

Closing for inactivity. Feel free to reopen this issue or a new one if you have other questions.

from md4all.