Meeting #2 - 06 Nov 2023

Images dimension: reduce 512 is too much at the beginning, so that the model is able to learn from different levels of resolution, start from small images, crop images from the ones provided instead of padding at the beginning; don’t go below 128x128 idea: cropping and rotating to obtain images that are completely different from the initial ones from the pov of the NN nb: see statistically if the cropout still includes 3 faces, if it only includes one is not good small crop out should still reflect the contex

In longer term: devil is in the details, have a good understanding on how to optimize NN for segmentation of images, doesn’t matter the level of accuracy, more the process of critical thinking to make it better and critical thinking on how to evaluate the performance

We don’t have to implement anything new, he strongly advices to not change the architectures

be creative in creating dataset, training and fine-tuning

Bachelor thesis will be on slack

HPC: from client you access to login note → from here you can go two paths

1. GPU node
2. HPN CPU

modules will stay local, every time they have to be uploaded again

Until Friday the 17 November

• Check IOU, maybe implement it in the script (Gloria)
• Break it:
  • using a smaller training set (~5 images) (nico)
  • different loss functions. MSELoss, NLLLoss, L1Loss, SmoothL1Loss(william)
  • input noise in the input (nico)
  • increase batch size, from 1 to 16 (william)