re-blocking

(1) Image Generation

Instructions

1. Clone our re-blocking Github repo

2. create an .env file in the main folder (hidden file?)

   1. create and enter Mapbox Access Token for the Mapbox Web API
   2. enter your file path to the parcel and building data (NYC is on our Dropbox)

   MAPBOX_ACCESS_TOKEN="$YOUR-API-KEY"
   LOCAL_PATH="$YOUR-DROPBOX-PATH/Million Neighborhoods/"
   

3. Run the Notebook parcels-buildings.ipynb to export images (parcels and buildings)

4. Combine the images with the script combine_images.sh for Pix2Pix HD (leave the individual ones for CycleGan)

---

(2) Model Training

Instructions

1. Clone the model’s Github repo

   1. Read the readme.md
   2. install all dependencies (see readme)
   3. install/update cuda for linux

2. Move the generated images (from step 1 or from Dropbox → Images) into /datasets/parcels

   1. →/test, /train, and /val for Pix2Pix
   2. TestA, TestB, TrainA, TrainB etc. for CycleGan

3. Open the repo folder in your terminal

   cd …/pytorch-CycleGAN-and-pix2pix/

4. Train the model

   1. (I named the output folder according to the parameters I used not to forget them…)

   **Pix2Pix HD**
   python3 [train.py](http://train.py/) --dataroot ./datasets/parcels --name test --model pix2pix --direction AtoB --n_epochs 100 --batch_size 60
   
   **CycleGan**
   python3  --dataroot ./datasets/parcels --name parcels_cycle-gan_25k_e50_b50_A-B --model cycle_gan --direction AtoB --n_epochs 10 --batch_size 1

5. Output

   1. See output in /checkpoints/FOLDERNAME/web/index.html
   2. Monitor the training logs with logs-visualised.ipynb (WIP)
   3. Copy your folder (without the training data) to our Dropbox → Model-Training if you want to share the results 😊

Parameters

• See /options → base_options.py, test_options.py, and train_options.py for the full parameter list (and to set standards)
• So far, the goal has been to increase batch_size to just below memory collapse

Other

watch -n0.1 nvidia-smi → Nvidia GPU monitor

---

(3) Running the trained Model

Instructions

• Find the trained model here
  • CycleGan → Dropbox
    • Trained on 25k Images (Train A/B: 20k, Buffer: 2.5k, Val: 2.5k)
    • 50 Epochs, batch Size 50
    • Direction A→B
  • Pix2Pix → needs to be retrained due to storage issue

---

Notes

Running (Training?) the model on an Apple ARM (M1/M2) Mac

Problem

• Issue: No NVIDIA GPU → NO CUDA
• ARM GPU (MPS) can generally be used with PyTorch but doesn’t seem to work with our model

Solution

• Run on CPU → add to terminal command when running training script: “--gpu_ids -1”

python3 [train.py](http://train.py/) --dataroot ./datasets/parcels --name facades_pix2pix --model pix2pix --direction BtoA **--gpu_ids -1**

Resources

• https://developer.apple.com/metal/pytorch/
• https://pytorch.org/blog/introducing-accelerated-pytorch-training-on-mac/
• https://www.mrdbourke.com/pytorch-apple-silicon/