Run ScratchGAN in Docker
- Clone this repository to your computer.
git clone https://github.com/pulipulichen/docker-ScratchGAN.git
. - Start docker service.
- Copy the files from
input.example
toinput
. - Run
docker-compose build
. - Run
docker-compose run app bash
to enter the console of container. - Run
bash scratchgan/train.sh
. - Wait for training and evaluating.
- The first training result will disappear in
./checkpoints/scratchgan-2002_sentences.txt
. - Keep waiting for
scratchgan-4002_sentences.txt
andscratchgan-6002_sentences.txt
. It takes at least three hours to perform each iteration.
The sentences written by ScratchGAN will become more and more like humans after iteration.
Following are the sentences from scratchgan-2002_sentences.txt:
- What had everything when on of police has running of to winds testing all when am have
- Those I was women to sick to share move 3 have
- It ' s more only decent of ( apart much Those to should so
Following are the sentences from scratchgan-4002_sentences.txt:
- Greg bomb said the principle is still because , done that it is a
- He took the time on a family property to appropriate in 2010 together something , and the Protection
- Although it has said that I just like the after old wonderful ending and opened
And then, following are the sentences from scratchgan-4002_sentences.txt:
- Instead that the same time before I would decided to it held , his lives , than badly like it at the
- And the fans moved the game off open that the middle voting of government , fairly now as those economies country increased at its
- The second year of ABC Florida 250 seconds about the third Trust , certainly ' t show again to it down that every
But the setences still seem meaningless.
I have tried to replace universal-sentence-encoder with universal-sentence-encoder-multilingual
in eval_metrics.py
but failed. It seem like got module version conflicting in tensorflow_gan and tensorflow_text, which is required by universal-sentence-encoder-multilingual. Maybe you can try another sentence representation encoder in the futrue.