nst-vgg19 0.1.6

Neural Style Transfer using VGG19

NST_VGG19

Neural Style Transfer using VGG19.

Original paper link.

VGG19 weights from torchvision.

Installation

pip install nst_vgg19

Usage

from nst_vgg19 import NST_VGG19

# images must be Numpy arrays. Use np.array(pil_image)

style_image = load_image('style.png')
content_image_1 = load_image('img1.jpg')
content_image_2 = load_image('img2.png')

nst = NST_VGG19(style_image)

result_1 = nst(content_image_1)
result_2 = nst(content_image_2)

NST_VGG19 constructor options

• style_image_numpy: Numpy array of the style image in format (Heght, Width, Channels). This is a default Numpy image array.
• content_layers_weights: Dictionary of weights for content losses.
• style_layers_weights: Dictionary of weights for style losses.
• quality_loss_weight: Weight for quality loss.
• delta_loss_threshold: Loss change threshold for stopping optimization.

If you do not specify weights of loss, the folowing parameters will be used:

DEFAULT_CONTENT_WEIGHTS = {
    'conv_1': 35000,  # Shape?
    'conv_2': 28000,
    'conv_4': 30000,
}
DEFAULT_STYLE_WEIGHTS = {
    'conv_2': 0.000001,  # Light/shadow?
    'conv_4': 0.000009,  # Contrast?
    'conv_5': 0.000006,  # Volume?
    'conv_7': 0.000003,
    'conv_8': 0.000002,  # Dents?
    'conv_9': 0.000003
}
quality_loss_weight=2e-4

If optimization delta becomes less than delta_loss_threshold then style transfer stops.

nst = NST_VGG19(style_image, style_layers_weights=my_weights, delta_loss_threshold=0.001)

result = nst(content_image) # no params except of image