Training issues - Githubissues

HenriquesLab / ZeroCostDL4Mic

ZeroCostDL4Mic: A Google Colab based no-cost toolbox to explore Deep-Learning in Microscopy

MIT License

566 stars 130 forks source link

Training issues #218

Open carolche opened 2 years ago

carolche commented 2 years ago

Hi Everyone,

I'm experiencing the same problem as a previous user (mafsousa, March 2022). I tried training my own model on the Cellpose/Omnipose notebook as well as command line (from kevinjohncutler's github), but I've been getting errors. I'm using the following:

OS: Mac Monterey 12.6
Browser: Chrome
Cellpose notebook version 1.0.2, 2.0, 2.1
with or without the most recent Omnipose The shape of my images is (1024, 1024, 1). I created the masks using Fiji adaptive thresholding plugin and are saved as binary images with the same shape as the images.
There was no problem with the paths and I was able to mount the images as they were shown correctly in step 3.1, and subsequent steps were also error-free. However, when I start training the model, I encounter errors as follows:
I also had the same error as mafsousa. Any help would be greatly appreciated. Thanks. Caroline

guijacquemet commented 2 years ago

Hi @carolche

Many thanks for reaching out! It looks like you may have an issue with your training dataset. You indicate that you saved your segmentation ground truth as binary images? Do you have only one structure to segment per image or multiple ? If you have more than one structure (ie multiple cells,), Cellpose would require label images to train.

carolche commented 2 years ago

@guijacquemet Yes, you are entirely right! I forgot to label the images! But alas, I still get an error.......

Thanks for your help.

guijacquemet commented 2 years ago

Hi @carolche

The error message shows you have less than 5 labels per image in your training dataset. Is this the case?

Maybe you could share your dataset?

Cheers

Guillaume

carolche commented 2 years ago

Hello @guijacquemet

so sorry for the late reply! I really appreciate your help. I was able to figure what the problem was. I wasn't labeling the masks appropriately. I tried a few different things and what worked best was to use ImageJ, adaptive threshold to make a binary mask, then threshold again, followed by counting and creating an ROI map. and it's working so far..... Thank you for all your help! Keep up the good work!

Caroline

Burt3 commented 1 year ago

Hi @carolche,

Would you please be able to provide an explanation to how you solved this issue (FileNotFoundError), as I am experiencing the exact same issue! I am currently producing a PNG mask image using the Cellpose GUI to complement my source images. Thanks:)