search

mirthAI / Fast-DDPM

MIT License
52 stars 7 forks source link

Hello #6

Closed superY688 closed 1 week ago

superY688 commented 1 month ago

Hello,First of all, thanks for the great work and making it available to everyone.

I am interested in image denoising,I want to know to how to train my own datasets for image denoising.

I would appreciate it if you could tell me.Thanks

Sebastianjhx commented 1 month ago

Hi there, thanks for your interest in our work.

For the purpose of training a customized dataset for an image denoising task, first ensure that your dataset with paired images is ready in the 'data' folder. Then, update the data root directory in the configuration file. And you also need to write you own dataset scripts in the 'dataset' folder.

Feel free to leave more comments regarding any other issues.

superY688 commented 1 month ago

Hi there, thanks for your interest in our work.

For the purpose of training a customized dataset for an image denoising task, first ensure that your dataset with paired images is ready in the 'data' folder. Then, update the data root directory in the configuration file. And you also need to write you own dataset scripts in the 'dataset' folder.

Feel free to leave more comments regarding any other issues.

Thank you so much for your reply.I have trained a customized dataset through the method presented in the paper. However，model denoising is not good on my customized dataset. The loss has been fluctuating，the psnr value of the denoised image is only 15 ，I do not know how to improve it. Can you provide me with the LDFDCT‘s training logs？ I would appreciate it if you could reply to me.

Sebastianjhx commented 1 month ago

In this case, is your dataset medical related? How many time steps out of 1000 are you using and how many iterations have you trained the model?

superY688 commented 1 month ago

In this case, is your dataset medical related? How many time steps out of 1000 are you using and how many iterations have you trained the model?

Many thanks for your prompt response

My dataset is not medical related.I use 10 time steps out of 1000.I trained the model for 110000steps.The length of my training datasets is 2500 and the datasets have 4 catogories.Evey training image has a noised version and a groundtruth.

Do datasets have to be medical related, and what part of the code should I change if I want to use datasets from other fields.

Thank you very much for your kindness，your assistance means the world to me.

Sebastianjhx commented 4 weeks ago

So far all datasets used in our experiments are medical related, and we haven't extend our model to non-medical fields yet. However, we suppose our model should apply to all grayscale data in this scenario without changes. My suggestion would be training the models for longer periods to observe the changes and experimenting with different numbers of timesteps (20, 50, 100 out of 1000). And please make sure your data is properly preprocessed before feeding into the model.

superY688 commented 3 weeks ago

So far all datasets used in our experiments are medical related, and we haven't extend our model to non-medical fields yet. However, we suppose our model should apply to all grayscale data in this scenario without changes. My suggestion would be training the models for longer periods to observe the changes and experimenting with different numbers of timesteps (20, 50, 100 out of 1000). And please make sure your data is properly preprocessed before feeding into the model.

Many thanks for your prompt response.My datasets are gray,but the loss value fluctuates and does not converge when i train the moel.Can you provide me some information about the loss value When you train the model on CT datasets.

Thank you very much for your kindness，your assistance means the world to me.