neural-style-transfer

Neural Style Transfer in Rust

This example shows how to implement Neural Style Transfer using PyTorch and Rust, it is very close to the Neural Transfer using PyTorch tutorial which contains far more details.

The Neural Style Transfer algorithm takes as input a content image and a style image and produces some new images looking similar to the content image but using the style from the second image.

Running the Code

We use a pre-trained VGG-16 convolutional network, the weights can be found in vgg16.ot. Running the example can be done with the following command, this will create some new images in the current directory.

cargo run --example neural-style-transfer style.jpg content.jpg vgg16.ot

Here is some example content and output images using the Starry Night as a style image.

(original image source)

Neural Style Transfer

The Neural Style Transfer algorithm works by gradient descent starting from an initial image -- we use the content image as a starting point but could also use random noise.

This gradient descent aims at minimizing a loss composed of two parts: the style loss evaluates how much our current image is similar to the style image in terms of style; the content loss focuses on whether our current image has a content close to the reference content image.

In order to define these distances we use a pre-trained convolutional neural network. This network is run on the current image as well as on the style and content images. Intuitively the features extracted by the lower layers are more about style and the features extracted by the upper layers about content.

The content loss is computed using the mean-square error on the extracted features for some upper layers. This error helps ensuring that the extracted features happen at the same on-screen places between the content and the current images.

For the style loss we do not really care about spatial location of the style features. So rather than the mean-square error we use some Gram matrix to define a loss that is location independent.

fn gram_matrix(m: &Tensor) -> Tensor {
    let (a, b, c, d) = m.size4().unwrap();
    let m = m.view(&[a * b, c * d]);
    let g = m.matmul(&m.tr());
    g / (a * b * c * d)
}

fn style_loss(m1: &Tensor, m2: &Tensor) -> Tensor {
    gram_matrix(m1).mse_loss(&gram_matrix(m2), 1)
}

Loading the Model and Images

We start by creating a device which can be either a CPU or a GPU depending on what is available.

    let device = Device::cuda_if_available();

Then we create a variable store on this device, this variable store will be used to hold the pre-trained variables for the VGG-16 model. Using this variable store we create the model and load the weights from the weight file. Finally we freeze the variable store ensuring that the model weights are not modified when running the optimizer later in the process.

    let mut net_vs = tch::nn::VarStore::new(device);
    let net = vgg::vgg16(&net_vs.root(), imagenet::CLASS_COUNT);
    net_vs.load(weights_file)?;
    net_vs.freeze();

Next we load the style and content image, and move them to the device. Note that the unsqueeze method is used to add a batch-dimension of size 1 so that the model can be run on these images.

    let style_img = imagenet::load_image(style_img)?
        .unsqueeze(0)
        .to_device(device);
    let content_img = imagenet::load_image(content_img)?
        .unsqueeze(0)
        .to_device(device);

Running the Model on the Content and Style Images

We now run the model on the style and content images. These calls returns some vector containing the extracted features for each of the model layers.

    let style_layers = net.forward_all_t(&style_img, false, Some(max_layer));
    let content_layers = net.forward_all_t(&content_img, false, Some(max_layer));

As we will use gradient descent to optimize an image, we create a second variable store to hold this image as a variable. The initial value for this variable is obtained by copying the content image.

    let vs = nn::VarStore::new(device);
    let input_var = vs.root().var_copy("img", &content_img);

Gradient Descent Loop

We use Adam for optimization: creating an optimizer requires to pass it the variable store that we want to optimize on.

    let opt = nn::Adam::default().build(&vs, LEARNING_RATE)?;

In the gradient descent loop the style and content losses are computed, summed together, and run an Adam optimization step is performed. We also regularly print the current loss value and write the current image to a file.

    for step_idx in 1..(1 + TOTAL_STEPS) {
        // ... compute style_loss and content_loss ...
        let loss = style_loss * STYLE_WEIGHT + content_loss;
        opt.backward_step(&loss);
        if step_idx % 1000 == 0 {
            println!("{} {}", step_idx, f64::from(loss));
            imagenet::save_image(&input_var, &format!("out{}.jpg", step_idx))?;
        }
    }

In order to compute the losses, the model is evaluated on the current image. The style loss is then extracted from the style layers and the content loss from the content layers.

        let input_layers = net.forward_all_t(&input_var, false, Some(max_layer));
        let style_loss: Tensor = STYLE_INDEXES
            .iter()
            .map(|&i| style_loss(&input_layers[i], &style_layers[i]))
            .sum();
        let content_loss: Tensor = CONTENT_INDEXES
            .iter()
            .map(|&i| input_layers[i].mse_loss(&content_layers[i], 1))
            .sum();