Daily_notes.txt

19/1/2021
Wrote code for squeeze unet from a paper (first time i transulated a unreleased gitcode and found accuracy) wow!

20/1/2021
ALso ajith this last paper which is out of office right ? spend more time in office papers okay? thats also good u
r spending too much time with easier things
Write inference code - done
Write accuracy checking code             - done
Test the speed of squeezenet inference
Understand why squeezeunet works
use augumentation - done
21/1/2021
Compare with hourglass model, Eff Unet, Squeeze unet , Form a new architecture - best method is to draw and see
Compare with segnet, see other hour glass models also
You get 10 days finish it
0.70 is the target accuracy for camvid - Unet
There are two things Ajith - HRNEt+hour glass/squeezenet to new model
                           - Squeezenet to new model
    both fastness and accuracy should be good

HR-Net
https://arxiv.org/pdf/1908.07919v2.pdf
https://github.com/HRNet/HRNet-Semantic-Segmentation

22/1/2021

Training log file should be generated with - dataset name, model name, number of train and val images , training epochs inos - done
Write code for full camvid testing - done
Write model predict and generate output - done
check if mean iou score and mean f1 score is correct or not  with model predict output - important because we would know
    whether our test is correct or not . -

Ajith finsh all above in this weekend okay ? be fast
squeezenet , HRNet training with camvid - done
Original HRNet with LIP dataset - working check the iou with original repo in LIP and in paperswithcode 

Make the new model to get above all accuracies in camvid dataset -

Dont change from UNet - it hard to get world accuacy within a small time, you also dont have gpus- reproduce HRNet then modying HRnet and family
    of HRNet is the only way

model should be best UNet in this world !! okay ? it should also have the speed

Ajith the percentage of SOTA for human parsing is less because of varying size of images, you can fill black color in it
    like before
Squeezenet - having the speed - done

Ajith competitors are - (for camvid)
PSP net with 0.69
HRNet with 0.74

23/1/2021
Testings :
code is not using GPU - use tensorflow-gpu
drone - working & speed is good
camvid_full - test again - do this today -working to a level
lookintoperon - rewrite code and test
write color.txt for segmentation after prediction

dont worry about hrnet


26/1/2021-
option 1 -  new hour glass model with speed of inference and training for real time usage  - squeezenet with Unet++ or UnetResNext
finish above one first today
New model:
only 1 max pooling like red net -use attenation gates like Att-net - try - 26/1
Make ResUnet like how squeeze unet was made from unet
FeedbackUnet based Squeezenet - try - 26/1
Adding Enhanced rotation invariant with squeezenet - try - 26/1

Checklist :
see whether the model is correct with batch normalization , relus
check DA's
check learning rates


now you have four different models. form it and see which wins use it .

option 2 - above HR net but it should beat HRNet accuracy in LIP