config # not a good way to do this

unet from scratch # unet

dataloader # add basic dataset class for segmentations

loss

optimizers

training # 1. train step 2. loader 3. save checkpoint 4. save train and val loss (todo : log)

ViT from scratch

ViT from scratch # qvk not clear

add cityscapes dataset for segmentation, can run but net output and target shape confused

loss function pytorch bug # https://clay-atlas.com/blog/2020/05/16/pytorch-cn-error-runtimeerror-ambiguous/

running successfully!! # total label 35， solve CUDA ERROR: device-side assert triggered at problem, reason:number of classes and network out of classes is not identical.

loss is decreasing but very slowly, accurary is not right!,more evaluation.

200 epoch without too much data arguementation

add inference

inference structure and training structure and trace