- Need to add the loss in the validation to compare
- adding the dilated convolutions part
- Why would I have less image ( Only 4514 imgs after cropping )?
- Try
Generalized Dice Losson the correct images - check after resample, whether it need to change the affine
- Try the
3D Patchwise U-Net with Transition Layers for MR Brain Segmentation - using BET to do brain stripping
- read other people's patch size (?), and training time, and learning rate
- check what I did in the padding part
- Try dropout
- the filter in each layer might need to change
- try transfer learning to speed up training
- might need more place to overlap with each other(could be a place to dip into) (how to remove borders) (look for the border pharams) (how the torchio deal with the overlapping part?)
- read about the learning rate
- why the matrix is different from the loss??
- why the print network show so less information now?
- tune the patch based parameters to make the train faster
- remember to find the suitable patch size
- make a parameter of whether to compute the background
- to test the dice loss
- add more Fully Connected Layer(???)
- rewrite the test part
- add one more layer without using concat
- force to upload once to github to clear some redundant file
- Reading the paper: Bringing Background into the Foreground: Making All Classes Equal in Weakly-supervised Video Semantic Segmentation (Whether to include the background??)
- Finding more paper to read!
- using leave one out validation
- find out why Clipping input data to the valid range for imshow with RGB data ([0..1] for floats or [0..255] for integers)_. (if still cannot fix, we can go along with the gray image)
- find out the reason why the second row of visualization do not working
- make the data plot according to the original scale
- why the color part isn't right????
- add mp4 part in the validation step end later
- adding the name of each axs
- in utils, there is a squeeze_data.py file which is used for squeeze the label data (in ADNI, label data have an extra dimension like this (1, 192, ))
- loss function and matrixs: all compute the background
- the visualization part:
0. Using the whole brain volume to do the visualization part
- randomly choose 150 images (but every time the image is fixed) from the 1069 baseline images in the ADNI dataset
- using this images to do visualization everytime (with the same order every time)
- if not using resample, using the orginal dice loss ( reduce by mean ), the loss would drop at a special moment to around 0.4, and all predict the background class.
- But after using the resample, cropped, and GeneralizedDiceLoss. The model would not learn at all ( the loss is always around 1). Very weird
- ddp2 cannot be used in Compute Canada
- When I directly using the patches to compute the dice score, it return quite small number (almost equal to 0), but when I use the whole image to compute, it would return the normal number
- Let my model run with only one GPU, ( this isn't helping anything at all )
