JuliaGRAM: Julia implementation of GRAM-nets

This is the source code for the paper Generative Ratio Matching Networks.

Check out the results interactively right now

You can check our experiment logs at https://app.wandb.ai/xukai92/gram-public, which are logged by WeightsAndBiasLogger.jl, a Julia interface for Weights & Biases.

How to run the code?

1. Install Julia and make julia available in your executable path.
   • This code was developed using Julia 1.3.1 and we suggest using the same version.
2. Download the code in a location which we will refer as GRAM_DIR.
3. Start a Julia REPL by entering julia in your terminal.
   • Press ] button to enter the package manager.
   • Install DrWatson and PyCall by add DrWatson PyCall.
     • Input using PyCall to start install Python depedencies.
     • Input PyCall.Conda.add("matplotlib") to install matplotlib.
     • Input PyCall.Conda.add("tikzplotlib") to install tikzplotlib.
     • Input PyCall.Conda.add("wandb") to install wandb.
   • Activate the project environment by activate $GRAM_DIR.
   • Install all dependencies by instantiate.
   • Exit the REPL.
4. Do julia $GRAM_DIR/scripts/master.jl --exp 1
   • This will produce Figure 1 (and Figure 6 & 9 in the appendix).
   • To produce other plots, change the argument to
     • --exp 2 for Figure 2 (and Figure 7 in the appendix), or
     • --exp 3 for Figure 8 in the appendix.
   • You can also adjust the parameter sweep in master.jl by yourself.
   • If you want to run multiple trainings in parallel, please append JULIA_NUM_THREADS=$N_PARALLEL to the beginning of the command (with a space), where N_PARALLEL is the number of parallel trainings you want to run.
     • By default the script uses 2 GPUs for parallelism. You can add --n_gpus $N_GPUS to the command to control the number.

You can also modify the arguments in scripts/gram.jl and run the file on its own by julia $GRAM_DIR/scripts/gram.jl

Our code by default logs all the training details using Weights & Biases, please install W&B and set it up following here. Or if you don't want to log things, add --nowandb.

Interact with the a pre-trained GRAM network on CIFAR10

We provide a pre-trained model using this Julia version of GRAM networks at demo/cifar10-gramnet.bson. You can use the provided Jupyter notebook at demo/interact.ipynb to interact with our pre-trained GRAM network.

NOTE: This is NOT the exact code we used for the CIFAR10 experiment in our paper. For reproducibility of paper on those experiments, please check GRAMFlow.

A small quiz

We generated 50 images from our GRAM network and mixed them with 50 real images. Can you tell the generated ones from sampled ones?

Click here to see the answers. How many mistakes did you make?

---

Maintainer
Kai Xu