For part 1: Environment Preparationdirectory:https://github.com/bnsreenu/python_for_microscopists/tree/master/228_semantic_segmentation_of_aerial_imagery_using_unetpseudocode for github repository...

For part 1: Environment Preparation



You are given a seed repo for this project located here - look under the directory 228_semantic_segmentation_of aerial imagery _using_unet. 1. Create a new repo under one Github account of a member of your team. 2. If you are working on own laptops/desktop: Checkout the code of the seed repo under src/ and incorporate poetry to allow your virtual environment to set up properly. 3. If you are working in Colab: use pip to install all dependencies bearing in mind that Colab has built in TF/Pytorch installations. 4. Install NNI. For Colab users: ensure that you have followed the instructions here to allow the Ul to be visible. 5. Submit the github repository URL with a branch titled ‘milestone-1" with the README.md file containing the installation instructions you followed and a screenshot of your computer of the NNI Ul screen. Add as collaborator the TA. 1. Watch this video that introduces the dataset and the approach ( gy, li VLEET El iT JEN Ke &: CIr: I O \ ad ; Watch later ~~ Share Testing Image VECO JT 1. Run the segmentation and produce plots of (a) 10 segmented images from the validation set (b) training and validation loss vs epochs, (b) Precision and Recall values 2. Submit the github repository URL with a branch titled ‘milestone-2" and with the docs/ folder containing the markdown file baseline e.md where you explain in 2 pages including figures UNet and in separate pages the results obtained. 1. Incorporate the Hyperparameter Optimization (HPO) method in your code and run the code to obtain the best (a) 10 segmented images from the validation set (b) training and validation loss vs epochs, (b) Precision and Recall values 2. Submit the github repository URL with a branch titled ‘milestone-3" and with the docs/ folder containing the markdown file hyperparameter-optimization.md where you explain in 2 pages including figures the method and in separate pages the results obtained. If you run out of compute resources i.e in Colab the GPU is taken away after few hours or your dedicated machine runs for ever or very slow, reduce the search space to produce results and achieve the milestone. If you reduce the space without any reason you will be deducted points. 1. Incorporate the NAS method in your code and run the code to obtain the best (a) 10 segmented images from the validation set (b) training and validation loss vs epochs, (b) Precision and Recall values 2. Submit the github repository URL with a branch titled ‘milestone-3’ and with the docs/ folder containing the markdown file neural-arch-search.md where you explain in 2 pages including figures the method and in separate pages the results obtained. If you run out of compute resources i.e in Colab the GPU is taken away after few hours or your dedicated machine runs for ever or very slow, reduce the search space to produce results and achieve the milestone. You are now tasked to do compress your model to fit a computer that may not have accelerator (GPUs) or enough memory to fit the original baseline model. Dont underestimate the impact of model compression - managing inference costs are the ranked consistently Number 1 goal for every enterprise that uses ML/AL 1. Read about the seminal paper 2. Implement Knowledge Distillation (KD) 3. Run the code to obtain the student's network (a) 10 segmented images from the validation set (b) training and validation loss vs epochs, (b) Precision and Recall values 4. Submit the github repository URL with a branch titled ‘milestone-4’ and with the docs/ folder containing the markdown file knowledge-distillation.md where you explain in 2 pages including figures the method and in separate pages the results obtained.