deepgen
Deep convolutional neural networks for generative modeling (i.e. VAEs, GANs, etc.)
Requirements
Tested with:
- python==3.7
- torch==1.3
- torchvision==0.4
- numpy==1.16
- matplotlib==3.1
- pandas==0.25
All dependencies are listed in environment.yml.
Installation
- Install dependencies:
conda env create -f environment.yml - Install with setup.py:
python setup.py install(replaceinstallwithdevelopto have code changes take immediate effect)
Usage
Run python cnn/run_vae.py -h to see a list of available arguments.
Train ResNet-style (Conditional) Variational Autoencoder:
python cnn/run_vae.py --train --save_examples
Test Trained Model:
python cnn/run_vae.py --test --save_examples --model_path /path/to/model
Note: be careful to set --report_freq and --batch_size properly. The former specifies how often to save examples and report losses (i.e. controls how spammy the output is), while the latter specifies how many examples will be batched together and must be appropriate for your computational and memory resources (I like batch size 1 for local development and use batch size 32 on a GPU, but for performance the largest possible batch size is best).
Play with Conditional Attributes
CelebA gives us access to 40 attributes (e.g. '5_o_Clock_Shadow', 'Arched_Eyebrows') which we have concatenated with the latent vector during training. By changing these attributes manually, we can have a bit of fun tweaking faces. To try this, run:
python cnn/toggle_attributes.py --img_path ... --model_path ...
Note that I have set defaults for all of the attributes because checking every combination would require 2^40 forward passes and is thus prohibitively expensive. The script iterates over each attribute and flips them on/off and outputs the results side-by-side. Below is an example I generated using my own profile photo.
I have compiled all of the positive changes (i.e. attributes = True) into the figure below; the only change between the two frames of the GIF is the value of the default attribute of "Smiling" (to show how default attributes affect the image globally).
Note: there are many different ways of including the attributes in the training process. Here I have simply concatenated them with the latent vector. Other approaches could be, for example:
- direct learning (i.e. as regression from image to latent space)
- computing a transformation by taking the difference in mean latent vectors for examples with and without an attribute