- The created dataset has 132K images
- Time Taken to create 132K images and labels of size 224px is 9mins
- Total 162K images of size 224px of which 128K are glomeruli images
- Something weird is happening transposing img on the axis (2,0,1) rather than reshaping inc the accuracy by 15%.
- Basic Unet
- Classification and then Unet architecture
- Transformer based Semantic segmentation
- can Convolution 3D be added to UNet
- IMPORTANT reduce parameters present in the model in a strategic way because model is always overfitting.
- Do changing the brightness of img is really need?
- Should we change the size of the image?
- Blurring Image with prob of (0.3)
- Flip all glomeruli images
- Mirroring all glomeruli images
- Changes the brightness of image between 0.8 and 1.3
- changing the brightness of image is about (800 micro secs)
- [ X ] the model was overfitting because there was to many parameters to just classify wheter an image has glomeruli or not.
- Current model accuracy is 84% with 3 epochs with Augmentation 3 images.
- [ ] Goal accuracy is 95%.
- Model is not able to recognize Aug 3 images i.e. images that has different brightness.
- can we add conv Nets to VIT? yes.
- Instead of Dot product between Keys and queries there should be matrix multiplication i.e. use vector attention.
- Added Conv to VIT for upsampling and got acc of 88.37% and loss of 0.27 with 2 epochs without Aug 3 Images.
- Which loss function will be the best ?
- BCELossWithLogits
- Focal Loss
- Dice Loss
- Weighted BCELossWithLogits
- Ideas to improve the current VIT-Conv model? -> Nothing work.
- Reduce Number of Filters in Conv2d.
- Add batchNorm2d.
- Change the reduce embedding dim.
- Added More attention layers.
- recognising few pixels in not glomeruli image.
- Make use of Anatomical Segmentation given in the dataset
- Model overfits after 2 epochs best prediction score 79% +- 1