Some notes from my first run of the script.

[jrittvo]

The script is very cool, especially how it does everything including the folder structure and the zipping!

These are notes about a run without --cn_unet.

1. If the file name has a . in it besides for the extension, like stable-diff-1.5.ckpt, will the split command still split it at the right place? Does it look for the last . or the first . or does it split at all of them? It would be fine to have a rule when you explain the script to someone that files names can only have a . at the extension, if that helps keep the code simple.

2. The xxxxx_original_512x512.zip needs to drop the _512x512. If it is not in a subfolder with a size like 512x768, the CoreML repo says that means the model is a default 512x512. The same way a split_einsum model does not get a size in the file name, or a 512x512 subfolder inside of the split-einsum folder.

3. I think the script should delete working folders as it completes each zip, and delete the xxxx_diffusers folder at the end of everything. To use up less space on the working drive while it is running.

I will do a run with --cn_unet tomorrow night. I think that one will need some minor changes, but I want to actually run it because my code reading skills are not that good.

[jrittvo]

I asked ChatGPT about the split('.') issue:

👱
Please explain what this line of python code does: fileNameArr = fileName.split('.')

🤖
This line of Python code takes a string variable fileName and splits it into an array of strings using the period (.) as the separator. The resulting array is assigned to the variable fileNameArr. This is useful when you need to extract the file extension from a file name, as the last element of the array will be the extension.

👱
The code: fileName = fileNameArr[0] is used to define the fileName. If the file name includes a . besides the extension separator . does this code use the parts of the file name up to the . at the extension, or only the part of the file name before a first . is encountered?

🤖
The line of code fileName = fileNameArr[0] takes the first element of the fileNameArr array and assigns it to the fileName variable. Therefore, if the file name includes a period (".") besides the extension separator, this code will only use the part of the file name before the first period (".") is encountered as the file name.

For example, if the original fileName variable is "my.file.name.txt", the split() method will create an array ['my', 'file', 'name', 'txt']. The line fileName = fileNameArr[0] will assign the first element of the array, "my", to the fileName variable, effectively ignoring the rest of the file name.

👱
How would I tell the code to combine all of the elements in the array back into the fileName, just omitting the last array element which is the file extension?

🤖
To combine all of the elements in the array back into the fileName, except for the last element which is the file extension, you can use the join() method to join the elements of the fileNameArr array, except for the last one, with a period separator. Here's an example code snippet:

# assuming fileNameArr is already defined
fileName = ".".join(fileNameArr[:-1])

In this code, fileNameArr[:-1] selects all the elements of the fileNameArr array except for the last element (which contains the file extension). The join() method then joins these elements together using a period separator to create the fileName.

[zhuHearthstone]

Thank you for your notes. I'm glad to discuss this script.
'--cn_unet' : I totally forgot that parameter. I think I should add it and try again.
'stable-diff-1.5.ckpt': You are right. I thought that was not normal.
I will delete xxxx_diffusers folder. But I think I should retain the working folders, the model can use in Mochi diffusion.I usually test the converted model after it has finished converting.

[zhuHearthstone]

I asked ChatGPT about the split('.') issue:

👱 Please explain what this line of python code does: fileNameArr = fileName.split('.')

🤖 This line of Python code takes a string variable fileName and splits it into an array of strings using the period (.) as the separator. The resulting array is assigned to the variable fileNameArr. This is useful when you need to extract the file extension from a file name, as the last element of the array will be the extension.

👱 The code: fileName = fileNameArr[0] is used to define the fileName. If the file name includes a . besides the extension separator . does this code use the parts of the file name up to the . at the extension, or only the part of the file name before a first . is encountered?

🤖 The line of code fileName = fileNameArr[0] takes the first element of the fileNameArr array and assigns it to the fileName variable. Therefore, if the file name includes a period (".") besides the extension separator, this code will only use the part of the file name before the first period (".") is encountered as the file name.

For example, if the original fileName variable is "my.file.name.txt", the split() method will create an array ['my', 'file', 'name', 'txt']. The line fileName = fileNameArr[0] will assign the first element of the array, "my", to the fileName variable, effectively ignoring the rest of the file name.

👱 How would I tell the code to combine all of the elements in the array back into the fileName, just omitting the last array element which is the file extension?

🤖 To combine all of the elements in the array back into the fileName, except for the last element which is the file extension, you can use the join() method to join the elements of the fileNameArr array, except for the last one, with a period separator. Here's an example code snippet:

# assuming fileNameArr is already defined
fileName = ".".join(fileNameArr[:-1])

In this code, fileNameArr[:-1] selects all the elements of the fileNameArr array except for the last element (which contains the file extension). The join() method then joins these elements together using a period separator to create the fileName.

I also ask ChatGPT just now. Your method is correct.

[jrittvo]

I should retain the working folders, the model can use in Mochi diffusion.I usually test the converted model after it has finished converting.

Yes, you are right. It makes sense to have it to test.

[jrittvo]

These are my notes from the run with --cn_unet added to the command arguments. I hope they make sense.

With the --cn_unet argument:

After it completes filling the xxxx_split-or-original_size_cn folder, it needs to copy the Unet.mlmodelc file from the resources folder of the xxxx_split-or-original_size folder to the resources folder of the xxxx_split-or-original_size_cn folder.

The xxxx_split-or-original_size_cn/resources folder then has BOTH Unet types. That resources folder is the one that needs to be zipped, with _cn added to the end of the file name before the .zip extension. And then all of the _cn.zip files are used to build the folder tree, with all the folders also having _cn added to their names. None of the resource folders with only the regular Unet get saved as zips. They were just there to create a regular Unet file that could be added to the other set.

The way you have it running twice is the only to get both Unets. So don't change anything there. Just at the end, copy the one Unet type to the resources folder that has the other Unet type, and then zip the 7 pieces into the finished model, with _cn added in the name and _cn added to the names of the folders they get saved to, so we know it is a model that can work with ControlNet.

The logic is that the script makes either a final complete zipped model with the regular Unet or a final complete zipped model with both the regular Unet and the ControlledUnet. It never zips a model with only the ControlledUnet type. I don't think people would want that type. They either use a model that doesn't need to do ControlNet, or a model that works both with and without a ControlNet.

[jrittvo]

If the above set of comments made sense, here is a trick to make that logic simpler to execute.

When the script does a second run for the --cu_unet, it can save all the work to the same destination folder as the first run. All of the files that are duplicated in a second run are exact duplicates of the first run files, and they can and will just overwrite them. The first run's Unet.mlmodelc stays as it was, and the second run's ControledUnet.mlmodelc is an added file to the others. Reusing the destination folder saves needing to move that one Unet from one set to the other, and it saves about half of the holding space needed while the two sets of files are collecting for the final zip file. The files and folders still need to all have _cn added somehow.

(For a while, I was using just the arguments for the Unet pieces in the second run, without the TextEncoder, TextDecoder, etc arguments. That does not seem to save any significant time for the second run. It looked they they all got worked on anyway, even if they were not ever saved to disk. Maybe they are needed to build he Unet file(s). So the way you reuse the conversion command for the second run, with the added argument appended, is the cleanest way to have the code.)

[jrittvo]

For this --cn_unet stuff, I think it means you need to define 2 different zip routines.

The first run alone never calls a zip routine.

if second run (is invoked with the --cn_unet)
call a zip routine that appends _cn to files and folders,
else
call a zip routine that doe not append _cn to files and folders (using the output from the first run)

[zhuHearthstone]

"That's good. I am also trying to find a way to combine the two models: the normal model and the CN model. I will read your comments carefully. We can first convert the two models and then try to find a way to combine them.”
Zip the files at the end.