Demo on using deep signal priors for inverse problems. To try it out, click here.
What this is doing is solving argmin 1/2 || y - (x_1 + x_2) || - logp(x_1) -logp(x_2). The moving lines indicate the flow of the gradient of log p.
Commands:
- space: start/stop animation
- =: increase the standard deviation of the data fidelity term
- -: decrease the standard deviation of the data fidelity term
- click: Moves the data point "y" to a new location
First step is to follow the notebook in notebooks to train a model, next we have to export it in a format that TFJS can read:
$ tensorflowjs_converter --skip_op_check InvertPermutation --input_format=tf_hub models/two_moons_realnvp_d models/js/export3