Harry/tfworld-hackathon
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Update README.md

Browse Source
This commit is contained in:
Harry Stuart 2020-01-01 15:18:14 +11:00 committed by GitHub
parent 726b62f304
commit affe105d11
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
1 changed files with 1 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
README.md
View File

@ -24,7 +24,7 @@ I took a highly systematic and methodical approach, since a lot of my work was w
Firstly, I trained a regular GAN on the MNIST dataset using WGAN-GP to ensure that I had implemented the training algorithm correctly. Next, I used an old Tensorflow 1.x WaveGAN implementation with my architecture to be certain that my generator and critic models were correct. Then I inserted my generator and critic models into my WGAN-GP infrastructure, replacing the MNIST GAN. Lastly, I tested the script on the same audio datasets used in the WaveGAN paper to make sure everything was ready to go. Finally, I started running my script on piano audio, adjusting hyperparameters and optimizing my models' architecture (trying to avoid mode collapse and failure to converge). Firstly, I trained a regular GAN on the MNIST dataset using WGAN-GP to ensure that I had implemented the training algorithm correctly. Next, I used an old Tensorflow 1.x WaveGAN implementation with my architecture to be certain that my generator and critic models were correct. Then I inserted my generator and critic models into my WGAN-GP infrastructure, replacing the MNIST GAN. Lastly, I tested the script on the same audio datasets used in the WaveGAN paper to make sure everything was ready to go. Finally, I started running my script on piano audio, adjusting hyperparameters and optimizing my models' architecture (trying to avoid mode collapse and failure to converge).
## Challenges I ran into ## Challenges I ran into
I spent quite a bit of time getting used to tensorflow.GradientTape and watching tensors etc. This was new to me since this project was my first shot at using Tensorflow 2.0. The majority of errors I faced were due to implementation/import mistakes, which I scoured GitHub to solve. In particular, finding elegant workarounds for functions contained in tensorflow.contrib proved to be challenging. Annoyingly, many solutions made use of tf.compat.v1, which I steered clear of. I spent quite a bit of time getting used to `tensorflow.GradientTape` and watching tensors etc. This was new to me since this project was my first shot at using Tensorflow 2.0. The majority of errors I faced were due to implementation/import mistakes, which I scoured GitHub to solve. In particular, finding elegant workarounds for functions contained in `tensorflow.contrib` proved to be challenging. Annoyingly, many solutions made use of `tf.compat.v1`, so I had to circumvent the problem some other way.
Additionally, I had to maintain constant consideration with regards to my computation capacity. My PC has a Nvidia RTX 2060, but training still took many many hours, and I had to use small batch sizes. Additionally, I had to maintain constant consideration with regards to my computation capacity. My PC has a Nvidia RTX 2060, but training still took many many hours, and I had to use small batch sizes.