|
|
@@ -88,11 +88,12 @@ Oh and I swear I'll document the code properly...eventually. Admittedly I'm *on
|
|
|
### Getting Started Yourself
|
|
|
This project is built around the wonderful Fast.AI library. Unfortunately, it's the -old- version and I have yet to upgrade it to the new version. (That's definitely on the agenda.) So prereqs, in summary:
|
|
|
* ***Old* Fast.AI library** **UDATED 11/7/2018** Easiest thing to do in my mind is just to take the fastai/fastai folder and drop it in the root of this project, right next to fasterai's folder. Just today, I found this thread on installing fast.ai 0.7- This is probably your best resource on this subject! https://forums.fast.ai/t/fastai-v0-7-install-issues-thread/24652 . Do this first- this will take you most of the way, including dependencies.
|
|
|
-* **Pytorch 0.4.1** (needs spectral_norm, so latest stable release is needed).
|
|
|
-* **Jupyter Lab**
|
|
|
-* **Tensorboard** (i.e. install Tensorflow) and **TensorboardX** (https://github.com/lanpa/tensorboardX). I guess you don't *have* to but man, life is so much better with it. And I've conveniently provided hooks/callbacks to automatically write all kinds of stuff to tensorboard for you already! The notebooks have examples of these being instantiated (or commented out since I didn't really need the ones doing histograms of the model weights). Noteably, progress images will be written to Tensorboard every 200 iterations by default, so you get a constant and convenient look at what the model is doing.
|
|
|
-* **ImageNet** – It proved to be a great dataset for training.
|
|
|
+* **Pytorch 0.4.1** (needs spectral_norm, so latest stable release is needed). https://pytorch.org/get-started/locally/
|
|
|
+* **Jupyter Lab** [conda install -c conda-forge jupyterlab]
|
|
|
+* **Tensorboard** (i.e. install Tensorflow) and **TensorboardX** (https://github.com/lanpa/tensorboardX). I guess you don't *have* to but man, life is so much better with it. And I've conveniently provided hooks/callbacks to automatically write all kinds of stuff to tensorboard for you already! The notebooks have examples of these being instantiated (or commented out since I didn't really need the ones doing histograms of the model weights). Noteably, progress images will be written to Tensorboard every 200 iterations by default, so you get a constant and convenient look at what the model is doing. [conda install -c anaconda tensorflow-gpu]
|
|
|
+* **ImageNet** – Only if training of course. It proved to be a great dataset. http://www.image-net.org/download-images
|
|
|
* **BEEFY Graphics card**. I'd really like to have more memory than the 11 GB in my GeForce 1080TI (11GB). You'll have a tough time with less. The Unet and Critic are ridiculously large but honestly I just kept getting better results the bigger I made them.
|
|
|
+* **Linux** (I'm using Ubuntu 16.04) is assumed, but nothing from the above precludes Windows 10 support as far as I know. I just haven't tested it and am not going to make it a priority for now.
|
|
|
|
|
|
**For those wanting to start transforming their own images right away:** To start right away with your own images without training the model yourself, download the weights here: https://www.dropbox.com/s/7r2wu0af6okv280/colorize_gen_192.h5 (right click and download from this link). Then open the ColorizationVisualization.ipynb in Jupyter Lab. Make sure that there's this sort of line in the notebook referencing the weights:
|
|
|
|
Jason Antic