TechTrash-OstrisTraining 0.1.0

A package to use the OstrisTraining

Libs-Ostris (TechTrash_OstrisTraining)

This package provides a small Python wrapper around an Ostris training workflow.

The main entrypoint is OstrisTraining.train().

What train() does

• Download + prepare dataset from a public ZIP URL
  • ZIP root must contain image files + matching caption files (.txt)
  • Example pair: photobelle.png + photobelle.txt
  • Output: files are normalized to image_0.<ext>, image_0.txt, image_1.<ext>, image_1.txt, ...
• Write config YAML from a string into a deterministic file:
  • {absolute_path_racine}/config-{user_name}.yaml
  • Before writing, we patch the YAML automatically:
    • config.name is forced to user_name
    • config.process[*].datasets[*].folder_path is forced to the prepared dataset folder
    • config.process[*].training_folder is forced to absolute_path_output
    • config.process[*].trigger_word is forced to:
      • "ohwx woman" if gender="woman"
      • "ohwx man" if gender="man"
• Run training
  • Calls run.py inside absolute_path_ostris with the generated config path
• Cleanup on success
  • Deletes the prepared dataset folder + generated config yaml
  • Keeps the produced .safetensors model file

Install (local dev)

From this folder:

pip install -e .

Usage (recommended)

The best reference is src/ostristraining/example.py.

You can run it directly after installing:

python3 -m ostristraining.example

Or copy/paste this minimal usage:

from ostristraining.main import OstrisTraining

trainer = OstrisTraining(
    user_name="demo_user",
    absolute_path_ostris="/tmp/ostris_project",
    absolute_path_racine="/tmp/ostris_runs",
    gender="woman",  # "woman" or "man"
    absolute_path_output="/tmp/ostris_output",
)

# Public ZIP URL containing images + captions at the ZIP root.
url_zip_dataset = "https://example.com/dataset.zip"

# YAML config must be a STRING. Newlines + indentation matter in YAML.
# Tip: start from `src/ostristraining/example_config.yaml` and customize it.
config_yaml_content = """
job: "extension"
config:
  name: "will_be_overwritten"
  process:
    - type: "diffusion_trainer"
      trigger_word: "will_be_overwritten"
      datasets:
        - folder_path: "will_be_overwritten"
"""

trained_model_path = trainer.train(
    url_zip_dataset=url_zip_dataset,
    config_yaml_content=config_yaml_content,
)

print(trained_model_path)

Logging

This library uses Python logging (logger name: ostristraining).

In your app/handler you typically enable logs like:

import logging
logging.basicConfig(level=logging.INFO)

Dataset ZIP format (important)

At the root of the ZIP, you must have:

• images: .png, .jpg, .jpeg, .webp, .bmp, .gif
• captions: .txt

And each image must have a caption with the same base name:

• photo001.png + photo001.txt
• photo002.jpg + photo002.txt

If an image has no matching .txt, it is skipped.

Config patching rules (quick recap)

If you send a config similar to src/ostristraining/example_config.yaml, the library will ensure:

• config.name == user_name
• config.process[*].datasets[*].folder_path == {absolute_path_ostris}/dataset/{user_name}
• config.process[*].training_folder == absolute_path_output
• config.process[*].trigger_word == "ohwx woman" | "ohwx man" based on gender