Skip to content

Latest commit

 

History

History

po

README.md

Pseudo Optimizer

We reduce the optimization process into a prediction problem by training a feed-forward network that maps the per-layer gradients of the objective function to the optimal solution.

fig-po

Teaser

Interpolation among three random noise inputs.

Interpolation Target Interpoation Target
interp1 target1 interp2 target2
interp3 target3 interp4 target4
interp5 target5 interp6 target6
interp7 target7 interp8 target8
interp9 target9 interp10 target10

Usage

Prepare VGG model

Download VGG model and convert it to an npz file following the steps in ../texture_utils.

Train

Prepare your training images as follows:

DATA_FOLDER/
|-- train/
    |-- image_123.jpg
    |-- image_456.jpg
|-- test/
    |-- image_789.jpg

Run the following command to train a model and save the logs and checkpoints to SAVE_FOLDER

CUDA_VISIBLE_DEVICES=0 python improved_model.py --data-folder DATA_FOLDER --save-folder SAVE_FOLDER

For the details of more arguments, run the following command

python xxx_model.py --help

Test

Run the following command to test a pretrained model from the checkpoint PRETRAINED_MODEL_CKPT and save the synthesized images to TEST_FOLDER

CUDA_VISIBLE_DEVICES=0 python improved_model.py --test-only --data-folder DATA_FOLDER --test-ckpt PRETRAINED_MODEL_CKPT --test-folder TEST_FOLDER

A checkpoint file is named like train_log/folder/model-ITERATION.

Models

  • improved_model.py The latest model of the ProPO architecture with improved performance. Use this model to reproduce the results. The other models may not be up-to-date and may cause import errors.
  • progressive_model.py The model of the ProPO architecture without tuning.
  • adaptive_model.py The model of the AdaPO architecture.
  • model.py The base ModelDesc class used by tensorpack.