Implementation of Neural Style Transfer in Tensorflow + experimental feature for audio style transfer.
This app takes a 'content' (image / audio) and adapts it with texture elements from a second source (image / audio).
EXPERIMENTAL FEATURE: "audio" style transfer is now implemented in ipynb code but being tuned to produce more pleasing audio results.)
- Try my web app implementation at www.communicatemission.com/ml-projects#style_transfer. (audio style not yet implemented in app)
Credits / Attributions:
Using Audio as Style source:
- I have not seen this done before. The idea of representing audio as an image in order to apply visual network comes from Deep Learning for Coders with fastai & Pytorch by Howard and Gugger (2020). (Original inverntor unknwon?)
Image Content & Style:
-
This has been been coded heavily relying on utilizing techniques from DeepLearning.ai specialization on Coursera (www.coursera.org/learn/nlp-sequence-models) and the Tensorflow style transfer documentation (www.tensorflow.org/tutorials/generative/style_transfer), which are in turn based on a paper by Gatys, Ecker and Bethge (https://arxiv.org/abs/1508.06576).
-
Pretrained image models used are loaded from the tf.keras.applications model. This model currently ustilizes VGG19 (citation: Very Deep Convolutional Networks for Large-Scale Image Recognition (ICLR 2015))
Web app:
- The web app is built on the Anvil platform.
Images from Unpslash.com