Tensor Component Analysis for Interpreting the Latent Space of GANs

[ paper | project page ]

Code to reproduce the results in the paper “Tensor Component Analysis for Interpreting the Latent Space of GANs”.



Firstly, to install the required packages, please run:

$ pip install -r requirements.txt

Pretrained weights

To replicate the results in the paper, you’ll need to first download the pre-trained weights. To do so, simply run this from the command line:


Quantitative results

building the prediction matrices

To reproduce Fig. 5, one can then run the ./quant.ipynb notebook using the pre-computed classification scores (please see this notebook for more details).

manually computing predictions

To call the Microsoft Azure Face API to generate the predictions again from scratch, one can run the shell script in ./quant/classify.sh. Firstly however, you need to generate our synthetic images to classify, which we detail below.

Qualitative results

generating the images

Reproducing the qualitative results (i.e. in Fig. 6) involves generating synthetic faces and 3 edited versions with the 3 attributes of interest (hair colour, yaw, and pitch). To generate these images (which are also used for the quantitative results), simply run:

$ ./generate_quant_edits.sh

mode-wise edits

./images/116-blonde.gif ./images/116-yaw.gif ./images/116-pitch.gif

Manual edits along individual modes of the tensor are made by calling main.py with the --mode edit_modewise flag. For example, one can reproduce the images from Fig. 3 with:

$ python main.py --cp_rank 0 --tucker_ranks "4,4,4,512" --model_name pggan_celebahq1024 --penalty_lam 0.001 --resume_iters 1000
  --n_to_edit 10 \
  --mode edit_modewise \
  --attribute_to_edit male

multilinear edits


Edits achieved with the ‘multilinear mixing’ are achieved instead by loading the relevant weights and supplying the --mode edit_multilinear flag. For example, the images in Fig. 4 are generated with:

$ python main.py --cp_rank 0 --tucker_ranks "256,4,4,512" --model_name pggan_celebahq1024 --penalty_lam 0.001 --resume_iters 200000
  --n_to_edit 10 \
  --mode edit_multilinear \
  --attribute_to_edit thick

Please feel free to get in touch at: [email protected], where x=oldfield


All the code in ./architectures/ and utils.py is directly imported from , only lightly modified to support performing the forward pass through the models partially, and returning the intermediate tensors.

The structure of the codebase follows https://github.com/yunjey/stargan, and hence we use their code as a template to build off. For this reason, you will find small helper functions (e.g. the first few lines of main.py) are borrowed from the StarGAN codebase.


View Github