06_assignment.md

Fashion MNIST VAE encoder (Tensorflow.js demo)

https://github.com/tensorflow/tfjs-examples/tree/master/fashion-mnist-vae

// Need nodejs, yarn
npm install -g yarn

git clone https://github.com/tensorflow/tfjs-examples.git
cd tfjs-examples/fashion-mnist-vae

# Make sure you switched to Python 2.7 (for some depedencies)
conda activate python_2.7
yarn install

yarn download-data
yarn train
yarn serve-model

# separate terminal
yarn serve-client

# Open in a browser
http://localhost:1234/

Deep Image Prior (using Pytorch)

https://dmitryulyanov.github.io/deep_image_prior

• JPEG artifacts removal

• Inpainting (painting blank/corrupted sections)

• Super-resolution

• Denoising

• Inpainting (watermark removal)

    git clone https://github.com/DmitryUlyanov/deep-image-prior
    cd deep-image-prior
    conda activate nanos
    conda install jupyter
    conda env create -f environment.yml
    # conda install --yes --file requirements.txt
  
    # Share the kernel for the Jupyter notebook
    jupyter notebook .
  
    # Go to localhost:8000/<project>.ypnb
    # Change kernel to `nanos`
    # Execute each cell
  

On your local laptop or computer, if you don't have a CUDA GPU, replace the following three lines

    torch.backends.cudnn.enabled = True
    torch.backends.cudnn.benchmark =True
    dtype = torch.cuda.FloatTensor

With the following:

    # torch.backends.cudnn.enabled = True
    # torch.backends.cudnn.benchmark =True
    dtype = torch.FloatTensor

And re-run the experiments, the experience is about 10 times slower than with GPU.

If you want to run on Google's Colab:

    !git clone hhttps://github.com/DmitryUlyanov/deep-image-prior.git
    !cd deep-image-prior

Use Deep Image Prior to fix your family photos.

Style Transfer with webcam