Goal: To work on Gaussian NLL on facial keypoint detection task

[nandhini30]

Worked on it and have attached the loss graph in this comment.

[deebuls]

Great job !! Please continue the such kind of pro-active methods to gain faster feedback.

Feedback:

• Which loss are you tracking RMSE or NLL ?
• [ ] Please plot both for Gaussian as you need RMSE to compare with the non-uncertainty method
• Looking at the long tail one can see that you might be overlearning
  • [ ] Read about EarlyStopping and use that to stop the training early when there is no learning
• [ ] Make a boxplot and density plot for both the RMSE and the predicted sigma of the validation/testing set
• [ ] Calculate entropy of the predicted gaussian distribution and also make a Density plot for validation/testing set
  • We need both this plot for comparing uncertainty
• Also take some sample images and plot the predicted points and the uncertainty as heatmap
  • This can be little tricky , just send me a set of images and the predicted output, I will do it fast.
  • [ ] Choose the 5 images with min error and 5 images with max error
  • [ ] Also choose 5 image with min sigma and 5 images with max sigma.
  • [ ] Google and download 5 random face images from the net . convert to the size your model requires and find its output
  • [ ] Send the images and its output(in csv) here , I will plot the heatmap
  • This is for qualitative analysis

[deebuls]

@nandhini30 Dont wait to do all the task . Even if you have done any 1 task please put the corresponding graph here and we can discuss.

[nandhini30]

Okay sure. I’m working on earlystopping now. I’ll upload the graph once i finish that

[nandhini30]

I tried early stopping with patience = 3 for RMSE loss. This is the output i got Total no of epochs = 300 epochs, it stopped at epoch 138 epochs as early stopping. I plotted val loss For RMSE

For gaussian loss, it will upload it before night

[nandhini30]

This for GNLL loss. for 300 epochs, it stopped at 199 epoch at early stopping.

[nandhini30]

[deebuls]

@nandhini30 good work .

The only reason we plot is to compare(think about it !! There is no other purpose). If you agree on the above , then the plots you have plotted what should I compare ? If the answer is nothing to compare then the plot is not serving the purpose.

Maybe plot with val_loss box plots of 2 methods . So that we can compare how 2 different methods loss. Etc. you can plot .

Golden rule for plotting . Is there atleast 2 values to compare ?

Also read at the above comments where I have given a list of comments .

[nandhini30]

Yes..i will send the compared plots today. I just wanted to update my progress. That’s why i uploaded the individual graphs

[nandhini30]

[nandhini30]

I have attached the sprint report for 20 july, 2022 in this comment. Sprint report_20july.pdf

[nandhini30]

Below graph is drawn between |val_difference| vs two methods