Monai Label Training

[stevenagl12]

I have a folder for imagesTr with my training images, and a subfolder for labels with nrrd files for my labelmaps. How can I train a model with this using monai label?

[diazandr3s]

Hi @stevenagl12,

This is a good question.

If working on a local PC, MONAI Label studies should follow this structure: https://youtu.be/wtiEe_jiUzg?t=3267

You may find this video useful as well: https://www.youtube.com/watch?v=3HTh2dqZqew&list=PLtoSVSQ2XzyD4lc-lAacFBzOdv5Ou-9IA&index=3

Hope this helps

[stevenagl12]

I apologize for the late reply, but neither of these really help to solve the question. So I have my files in the correct format already from the first video. How do I train a new model now given that I have this labeled data? All the videos I can find say how to train amodel from scratch with unlabeled data using the segment editor process of the scribbles processes. Or viceversa how to download a trained model and infer with the testing data, but how do I run training with the data that is in the desired format?

Another way to ask this is how do we train a new model after editing the deepedit config file with the appropriate class dictionary and turning off the pretrained model option (while removing the files for any models from the model folder) from the Task01 or Task04 training data?

[diazandr3s]

Thanks for the update, @stevenagl12.

Another way to ask this is how do we train a new model after editing the deepedit config file with the appropriate class dictionary and turning off the pretrained model option (while removing the files for any models from the model folder) from the Task01 or Task04 training data?

You've already done the most difficult part. To trigger the training process, just click on the Train button in the 3D Slicer module shown here:

You can find more information here: https://docs.monai.io/projects/label/en/latest/quickstart.html

Hope that helps,

[stevenagl12]

How would you get the training and testing curves from the process then?

…

On Wed, Oct 26, 2022, 7:13 PM Andres Diaz-Pinto ***@***.***> wrote: Thanks for the update, @stevenagl12 <https://nam12.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.com%2Fstevenagl12&data=05%7C01%7Csalewis%40g-mail.buffalo.edu%7Cf55f67827bf449e7b81f08dab7a7ad1b%7C96464a8af8ed40b199e25f6b50a20250%7C0%7C0%7C638024228030455236%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000%7C%7C%7C&sdata=ZXXiuLY2dgeJzB9TlIl0FRMF2clcg%2F9ThIghm1kMgto%3D&reserved=0> . Another way to ask this is how do we train a new model after editing the deepedit config file with the appropriate class dictionary and turning off the pretrained model option (while removing the files for any models from the model folder) from the Task01 or Task04 training data? You've already done the most difficult part. To trigger the training process, just click on the Train button in the 3D Slicer module shown here: [image: train-] <https://nam12.safelinks.protection.outlook.com/?url=https%3A%2F%2Fuser-images.githubusercontent.com%2F11991079%2F198155652-63a916cc-fe9b-4329-b464-dd431369006f.png&data=05%7C01%7Csalewis%40g-mail.buffalo.edu%7Cf55f67827bf449e7b81f08dab7a7ad1b%7C96464a8af8ed40b199e25f6b50a20250%7C0%7C0%7C638024228030455236%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000%7C%7C%7C&sdata=XiFEfeuORlGK%2Fa%2FRGDYJ9aluZh9Li%2BCH3%2BCW2Hmm5qo%3D&reserved=0> You can find more information here: https://docs.monai.io/projects/label/en/latest/quickstart.html <https://nam12.safelinks.protection.outlook.com/?url=https%3A%2F%2Fdocs.monai.io%2Fprojects%2Flabel%2Fen%2Flatest%2Fquickstart.html&data=05%7C01%7Csalewis%40g-mail.buffalo.edu%7Cf55f67827bf449e7b81f08dab7a7ad1b%7C96464a8af8ed40b199e25f6b50a20250%7C0%7C0%7C638024228030611033%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000%7C%7C%7C&sdata=3a5SAU74KK0qd41DDdso682GL0u1KozEzFhf25PXGoU%3D&reserved=0> Hope that helps, — Reply to this email directly, view it on GitHub <https://nam12.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.com%2FProject-MONAI%2FMONAILabel%2Fdiscussions%2F1055%23discussioncomment-3975343&data=05%7C01%7Csalewis%40g-mail.buffalo.edu%7Cf55f67827bf449e7b81f08dab7a7ad1b%7C96464a8af8ed40b199e25f6b50a20250%7C0%7C0%7C638024228030611033%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000%7C%7C%7C&sdata=TWUR1sTFjWCYAnOqryO97oWcD26AzDiy%2BQ5HPbHsnPA%3D&reserved=0>, or unsubscribe <https://nam12.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.com%2Fnotifications%2Funsubscribe-auth%2FAHHKCLPV5VUQV4EJ3PO6BODWFG3JDANCNFSM6AAAAAARAAE4FM&data=05%7C01%7Csalewis%40g-mail.buffalo.edu%7Cf55f67827bf449e7b81f08dab7a7ad1b%7C96464a8af8ed40b199e25f6b50a20250%7C0%7C0%7C638024228030611033%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000%7C%7C%7C&sdata=lN8yflq0luXbiQdAfP31HabNJIQ923NCrOSlno5Caj0%3D&reserved=0> . You are receiving this because you were mentioned.Message ID: ***@***.*** com>

[diazandr3s]

MONAI Label uses tensorboard to keep track of training and validation curves. You can find those tensorboard files in the model folder once training starts.

[stevenagl12]

Ok, so another question is for the Task01 Brain Tumour segmentation, is there preprocessing that should take place in order to get the system to work? I am getting a size mismatch error that says: "RuntimeError: Given groups=1, weight of size [32, 5, 3, 3, 3], expected input[1, 8, 128, 128, 128] to have 5 channels, but got 8 channels instead."

[SachidanandAlle]

You might be using pretrained model that was done over 8 channels.. but your input has only 5 of them..

Fix the number of channels based on your data and discard existing pretrained model and train from scratch.. or if you have only subset of 8.. add empty channels to for the missing labels and fine tune model for your data

[stevenagl12]

I'm very confused with what I am doing wrong here. When I first connect to Monailabel for the server, the Train button is blacked out. As I said before, I have edited the config file, and removed all of the models. How am I supposed to train any new models from existing data when the train button isn't even active?

…

On Thu, Oct 27, 2022 at 12:59 PM SACHIDANAND ALLE ***@***.***> wrote: You might be using pretrained model that was done over 8 channels.. but your input has only 5 of them.. Fix the number of channels based on your data and discard existing pretrained model and train from scratch.. or if you have only subset of 8.. add empty channels to for the missing labels and fine tune model for your data — Reply to this email directly, view it on GitHub <https://nam12.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.com%2FProject-MONAI%2FMONAILabel%2Fdiscussions%2F1055%23discussioncomment-3984173&data=05%7C01%7Csalewis%40g-mail.buffalo.edu%7C84cce36b7e764a225f3d08dab83ca5e4%7C96464a8af8ed40b199e25f6b50a20250%7C0%7C0%7C638024867855200980%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000%7C%7C%7C&sdata=UuV2cNiemh1dLhcdO%2Bu3d5Knu70klVLGoIFcQPWz9rI%3D&reserved=0>, or unsubscribe <https://nam12.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.com%2Fnotifications%2Funsubscribe-auth%2FAHHKCLLP33R3F4GLKMHDOBDWFKYH7ANCNFSM6AAAAAARAAE4FM&data=05%7C01%7Csalewis%40g-mail.buffalo.edu%7C84cce36b7e764a225f3d08dab83ca5e4%7C96464a8af8ed40b199e25f6b50a20250%7C0%7C0%7C638024867855200980%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000%7C%7C%7C&sdata=in%2Ft0TtJpTlBdQpbNb9bcbL2e6%2FXXbIQnYN5%2BfR5%2Bkw%3D&reserved=0> . You are receiving this because you were mentioned.Message ID: ***@***.*** com>

-- Sincerely, Steven A. Lewis

[stevenagl12]

Ok, this wasn't in any of the videos I had reviewed so hopefully this is the problem.

[SachidanandAlle]

May be later you can help so improve the documentation part.. may be a wiki etc is also fine.. or some section on how to...?

[stevenagl12]

If this works, I will absolutely do that.

[diazandr3s]

Ok, this wasn't in any of the videos I had reviewed so hopefully this is the problem.

Hi @stevenagl12,

This folder structure was exactly the one I suggested in the first comment: #1055 (comment)

Here is again for your reference: https://youtu.be/wtiEe_jiUzg?t=3275

Hope this helps,

[stevenagl12]

The problem wasn't the folder structure. It was knowing that the data structure that is downloaded from the published tasks has the data in the incorrect folder to be able to train. The necessary labels for the data are in the labelsTr folder outside of the imagesTr or imagesTs folders, while the labels folder with the final and original folders in the imagesTr folder were empty for both of these tasks. I thought the dataset.json file that was downloaded that contained all of the necessary paths for the labels was sufficient to be able to run the training. Still having training errors but less so and working on those now.