search

chervonij / DFL-Colab

DeepFaceLab fork which provides IPython Notebook to use DFL with Google Colab
1.05k stars 265 forks source link

Wording/Phrasing changes, Remove References to multiple Drive accounts, Rename function name in Import Workspace, Refactor Dataset/Model Download #51

Closed camjac251 closed 1 year ago

camjac251 commented 2 years ago

I saw after the change in Google Drive (same account restriction), that a reference to multiple Google accounts was in the text box, so I ended up removing it but then wanted to go further and try helping with some of the other stuff as well. I updated the GPUs that Colab has as well, filled out the VRAM amount for the GPUs I have gotten. I also changed the function name in "import workspace". Also refactored the dataset and model downloading to hopefully make it easier in the future for multiple options and also allow it to be optional and not required.

chervonij commented 2 years ago

@camjac251

Okay, let's see.

  1. Downloading datasets and Q96 model cannot be optional.
  2. "GenericXSeg" and "Copy_to_Drive" must be enabled by default.
  3. More detailed descriptions are helpful, but why so many "the"? Are you a native English speaker or are you using some kind of translator?
camjac251 commented 2 years ago

Hey :)

  1. I was kind of hoping to trim off a minute or so when I continue training after a session expires. It's currently always set to download the pretrain data and model although I'm often training from my own model and only need to resume from the model saved in my workspace.zip on Google Drive. Making it optional, for me at least, is useful just as a time saver.
  2. I am oftentimes unchecking GenericXSeg when running XSeg Apply since I've trained my own model and want to force it to use only the model to apply. I didn't mean to uncheck Copy_to_Drive.
  3. I started working on this when I was a little sleep-deprived so I might've not have been the best at articulation. I'll read over what I wrote again and make sure I didn't goof up. I'm a native speaker though.
camjac251 commented 2 years ago

I gave it another look and saw many mistakes that I originally missed. Sorry about that. Added a commit just now with the latest changes.

chervonij commented 2 years ago

@camjac251

Okay, it works for you, but if someone needs to use built-in pretrain function and doesn't have CelebA or FFHQ downloaded, train won't start. Same for Quick96.

About language, that's fine, I'm just wondering ) I just often see google translator put "the" literally everywhere ))

camjac251 commented 2 years ago

@camjac251

if someone needs to use built-in pretrain function and doesn't have CelebA or FFHQ downloaded, train won't start. Same for Quick96.

The dropdown box for Download_Dataset and Download_Pretrained_Model help this. They're set to None by default since it'll be more convenient for someone to not have to set it to None every time they want to continue training vs. having to change the option every time they load the window. If the dropdown is changed from None to say FFHQ, it'll download that during the installation process. Someone will likely be continuing an existing model more often than starting a new one so None made better sense to me being the default, but it is still working with the download process of the pretraining datasets/model. image

camjac251 commented 2 years ago

I noticed a typo I made with the CelebA dataset URL. Fixed and should work now

chervonij commented 2 years ago

@camjac251

The dropdown box for Download_Dataset and Download_Pretrained_Model help this.

I understand. That's not what I'm talking about. I meant that you underestimate number of dumb people who will just forget to turn this option on, even when they need it. What to do with such people?

Now, pretrain dataset and Quick96 are loaded anyway. It's foolproof. It takes a few minutes, which isn't a lot of time on scale of a Colab session. Even if you save 5 minutes, on those options, it doesn't matter.

camjac251 commented 2 years ago

The entire DFL workflow can be quite a feat of its own, but I guess I could come up with something, as I was hoping this could be merged for everyone to find use in. I was thinking of moving the Manage workspace block above the DFL Installation block since it's more convenient (In my opinion) to get all of the dialogue boxes out of the way if you're wanting to continue training but only have a minute.

I was thinking if someone did import their workspace before the install block would run, there could be a check to see if they have files in their model folder or data_src/data_dst folder and if nothing is detected, then it could run the download of both the dataset and model, regardless of the option above. I'll try and figure something else, but it might be more than what I was hoping for.

chervonij commented 2 years ago

@camjac251

I understand what you want to do, but I have to consider that it will be used by different people, with different ideas of convenience and logic.

In any case. I disagree about optional uploading of datasets and Q96. Also, GenericXSeg must be enabled by default.

chervonij commented 2 years ago

Also, all sections must be collapsed by default