Add files

.gitignore

@@ -0,0 +1,164 @@
+training_data/
+results/
+
+
+# Byte-compiled / optimized / DLL files
+__pycache__/
+*.py[cod]
+*$py.class
+
+# C extensions
+*.so
+
+# Distribution / packaging
+.Python
+build/
+develop-eggs/
+dist/
+downloads/
+eggs/
+.eggs/
+lib/
+lib64/
+parts/
+sdist/
+var/
+wheels/
+share/python-wheels/
+*.egg-info/
+.installed.cfg
+*.egg
+MANIFEST
+
+# PyInstaller
+#  Usually these files are written by a python script from a template
+#  before PyInstaller builds the exe, so as to inject date/other infos into it.
+*.manifest
+*.spec
+
+# Installer logs
+pip-log.txt
+pip-delete-this-directory.txt
+
+# Unit test / coverage reports
+htmlcov/
+.tox/
+.nox/
+.coverage
+.coverage.*
+.cache
+nosetests.xml
+coverage.xml
+*.cover
+*.py,cover
+.hypothesis/
+.pytest_cache/
+cover/
+
+# Translations
+*.mo
+*.pot
+
+# Django stuff:
+*.log
+local_settings.py
+db.sqlite3
+db.sqlite3-journal
+
+# Flask stuff:
+instance/
+.webassets-cache
+
+# Scrapy stuff:
+.scrapy
+
+# Sphinx documentation
+docs/_build/
+
+# PyBuilder
+.pybuilder/
+target/
+
+# Jupyter Notebook
+.ipynb_checkpoints
+
+# IPython
+profile_default/
+ipython_config.py
+
+# pyenv
+#   For a library or package, you might want to ignore these files since the code is
+#   intended to run in multiple environments; otherwise, check them in:
+# .python-version
+
+# pipenv
+#   According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.
+#   However, in case of collaboration, if having platform-specific dependencies or dependencies
+#   having no cross-platform support, pipenv may install dependencies that don't work, or not
+#   install all needed dependencies.
+#Pipfile.lock
+
+# poetry
+#   Similar to Pipfile.lock, it is generally recommended to include poetry.lock in version control.
+#   This is especially recommended for binary packages to ensure reproducibility, and is more
+#   commonly ignored for libraries.
+#   https://python-poetry.org/docs/basic-usage/#commit-your-poetrylock-file-to-version-control
+#poetry.lock
+
+# pdm
+#   Similar to Pipfile.lock, it is generally recommended to include pdm.lock in version control.
+#pdm.lock
+#   pdm stores project-wide configurations in .pdm.toml, but it is recommended to not include it
+#   in version control.
+#   https://pdm.fming.dev/#use-with-ide
+.pdm.toml
+
+# PEP 582; used by e.g. github.com/David-OConnor/pyflow and github.com/pdm-project/pdm
+__pypackages__/
+
+# Celery stuff
+celerybeat-schedule
+celerybeat.pid
+
+# SageMath parsed files
+*.sage.py
+
+# Environments
+.env
+.venv
+env/
+venv/
+ENV/
+env.bak/
+venv.bak/
+
+# Spyder project settings
+.spyderproject
+.spyproject
+
+# Rope project settings
+.ropeproject
+
+# mkdocs documentation
+/site
+
+# mypy
+.mypy_cache/
+.dmypy.json
+dmypy.json
+
+# Pyre type checker
+.pyre/
+
+# pytype static type analyzer
+.pytype/
+
+# Cython debug symbols
+cython_debug/
+
+# PyCharm
+#  JetBrains specific template is maintained in a separate JetBrains.gitignore that can
+#  be found at https://github.com/github/gitignore/blob/main/Global/JetBrains.gitignore
+#  and can be added to the global gitignore or merged into this file.  For a more nuclear
+#  option (not recommended) you can uncomment the following to ignore the entire idea folder.
+#.idea/

.gitmodules

@@ -0,0 +1,3 @@
+[submodule "lora"]
+	path = lora
+	url = https://github.com/cloneofsimo/lora

.pre-commit-config.yaml

@@ -0,0 +1,35 @@
+repos:
+- repo: https://github.com/pre-commit/pre-commit-hooks
+  rev: v4.2.0
+  hooks:
+  - id: check-executables-have-shebangs
+  - id: check-json
+  - id: check-merge-conflict
+  - id: check-shebang-scripts-are-executable
+  - id: check-toml
+  - id: check-yaml
+  - id: double-quote-string-fixer
+  - id: end-of-file-fixer
+  - id: mixed-line-ending
+    args: ['--fix=lf']
+  - id: requirements-txt-fixer
+  - id: trailing-whitespace
+- repo: https://github.com/myint/docformatter
+  rev: v1.4
+  hooks:
+  - id: docformatter
+    args: ['--in-place']
+- repo: https://github.com/pycqa/isort
+  rev: 5.10.1
+  hooks:
+    - id: isort
+- repo: https://github.com/pre-commit/mirrors-mypy
+  rev: v0.991
+  hooks:
+    - id: mypy
+      args: ['--ignore-missing-imports']
+- repo: https://github.com/google/yapf
+  rev: v0.32.0
+  hooks:
+  - id: yapf
+    args: ['--parallel', '--in-place']

.style.yapf

@@ -0,0 +1,5 @@
+[style]
+based_on_style = pep8
+blank_line_before_nested_class_or_def = false
+spaces_before_comment = 2
+split_before_logical_operator = true

LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2022 hysts
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

app.py

@@ -0,0 +1,225 @@
+#!/usr/bin/env python
+"""Unofficial demo app for https://github.com/cloneofsimo/lora.
+
+The code in this repo is partly adapted from the following repository:
+https://huggingface.co/spaces/multimodalart/dreambooth-training/tree/a00184917aa273c6d8adab08d5deb9b39b997938
+The license of the original code is MIT, which is specified in the README.md.
+"""
+
+from __future__ import annotations
+
+import os
+import pathlib
+
+import gradio as gr
+import torch
+
+from inference import InferencePipeline
+from trainer import Trainer
+
+TITLE = '# LoRA + StableDiffusion Training UI'
+DESCRIPTION = 'This is an unofficial demo for [https://github.com/cloneofsimo/lora](https://github.com/cloneofsimo/lora).'
+
+USAGE_INFO = '''You can train and download models in the "Training" tab, and test them in the "Test" tab.
+
+You can also test the pretrained models in the [original repo](https://github.com/cloneofsimo/lora).
+Models with names starting with "lora/" are the pretrained models and the ones with names starting with "results/" are your trained models.
+After training, you can press "Reload Weight List" button to load your trained model names.
+
+Note that your trained models will be deleted when the second training is started.
+'''
+
+SPACE_ID = os.getenv('SPACE_ID', 'hysts/LoRA-SD-training')
+SHARED_UI_WARNING = f'''# Attention - This Space doesn't work in this shared UI. You can duplicate and use it with a paid private T4 GPU.
+
+<center><a class="duplicate-button" style="display:inline-block" target="_blank" href="https://huggingface.co/spaces/{SPACE_ID}?duplicate=true"><img src="https://img.shields.io/badge/-Duplicate%20Space-blue?labelColor=white&style=flat&logo=data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABAAAAAQCAYAAAAf8/9hAAAAAXNSR0IArs4c6QAAAP5JREFUOE+lk7FqAkEURY+ltunEgFXS2sZGIbXfEPdLlnxJyDdYB62sbbUKpLbVNhyYFzbrrA74YJlh9r079973psed0cvUD4A+4HoCjsA85X0Dfn/RBLBgBDxnQPfAEJgBY+A9gALA4tcbamSzS4xq4FOQAJgCDwV2CPKV8tZAJcAjMMkUe1vX+U+SMhfAJEHasQIWmXNN3abzDwHUrgcRGmYcgKe0bxrblHEB4E/pndMazNpSZGcsZdBlYJcEL9Afo75molJyM2FxmPgmgPqlWNLGfwZGG6UiyEvLzHYDmoPkDDiNm9JR9uboiONcBXrpY1qmgs21x1QwyZcpvxt9NS09PlsPAAAAAElFTkSuQmCC&logoWidth=14" alt="Duplicate Space"></a></center>
+'''
+CUDA_NOT_AVAILABLE_WARNING = '# Attention - CUDA is not available in this environment.'
+
+
+def show_warning(warning_text: str) -> gr.Blocks:
+    with gr.Blocks() as demo:
+        with gr.Box():
+            gr.Markdown(warning_text)
+    return demo
+
+
+def update_output_files() -> dict:
+    paths = sorted(pathlib.Path('results').glob('*.pt'))
+    paths = [path.as_posix() for path in paths]  # type: ignore
+    return gr.update(value=paths or None)
+
+
+def create_training_demo(trainer: Trainer,
+                         pipe: InferencePipeline) -> gr.Blocks:
+    with gr.Blocks() as demo:
+        base_model = gr.Dropdown(
+            choices=['stabilityai/stable-diffusion-2-1-base'],
+            value='stabilityai/stable-diffusion-2-1-base',
+            label='Base Model',
+            visible=False)
+        resolution = gr.Dropdown(choices=['512'],
+                                 value='512',
+                                 label='Resolution',
+                                 visible=False)
+
+        with gr.Row():
+            with gr.Box():
+                gr.Markdown('Training Data')
+                concept_images = gr.Files(label='Images for your concept')
+                concept_prompt = gr.Textbox(label='Concept Prompt',
+                                            value='sks',
+                                            max_lines=1)
+                gr.Markdown('''
+                    - Upload images of the style you are planning on training on.
+                    - For a concept prompt, use a unique, made up word to avoid collisions.
+                    ''')
+            with gr.Box():
+                gr.Markdown('Training Parameters')
+                num_training_steps = gr.Number(
+                    label='Number of Training Steps', value=1000, precision=0)
+                learning_rate = gr.Number(label='Learning Rate', value=0.0001)
+                gr.Markdown('''
+                    - It will take about 15-20 minutes to train for 1000 steps with a T4 GPU.
+                    - You may want to try a small number of steps first, like 1, to see if everything works fine in your environment.
+                    ''')
+
+        run_button = gr.Button('Start Training')
+        with gr.Box():
+            with gr.Row():
+                check_status_button = gr.Button('Check Training Status')
+                with gr.Column():
+                    training_status = gr.Markdown()
+                    output_files = gr.Files(label='Trained Weight Files')
+
+        run_button.click(fn=pipe.clear)
+        run_button.click(fn=trainer.run,
+                         inputs=[
+                             base_model,
+                             resolution,
+                             concept_images,
+                             concept_prompt,
+                             num_training_steps,
+                             learning_rate,
+                         ],
+                         outputs=[
+                             training_status,
+                             output_files,
+                         ],
+                         queue=False)
+        check_status_button.click(fn=trainer.check_if_running,
+                                  inputs=None,
+                                  outputs=training_status,
+                                  queue=False)
+        check_status_button.click(fn=update_output_files,
+                                  inputs=None,
+                                  outputs=output_files,
+                                  queue=False)
+    return demo
+
+
+def find_weight_files() -> list[str]:
+    curr_dir = pathlib.Path(__file__).parent
+    paths = sorted(curr_dir.rglob('*.pt'))
+    return [path.relative_to(curr_dir).as_posix() for path in paths]
+
+
+def reload_lora_weight_list() -> dict:
+    return gr.update(choices=find_weight_files())
+
+
+def create_inference_demo(pipe: InferencePipeline) -> gr.Blocks:
+    with gr.Blocks() as demo:
+        with gr.Row():
+            with gr.Column():
+                base_model = gr.Dropdown(
+                    choices=['stabilityai/stable-diffusion-2-1-base'],
+                    value='stabilityai/stable-diffusion-2-1-base',
+                    label='Base Model',
+                    visible=False)
+                reload_button = gr.Button('Reload Weight List')
+                lora_weight_name = gr.Dropdown(choices=find_weight_files(),
+                                               value='lora/lora_disney.pt',
+                                               label='LoRA Weight File')
+                prompt = gr.Textbox(
+                    label='Prompt',
+                    max_lines=1,
+                    placeholder='Example: "style of sks, baby lion"')
+                alpha = gr.Slider(label='Alpha',
+                                  minimum=0,
+                                  maximum=2,
+                                  step=0.05,
+                                  value=1)
+                seed = gr.Slider(label='Seed',
+                                 minimum=0,
+                                 maximum=100000,
+                                 step=1,
+                                 value=1)
+                with gr.Accordion('Other Parameters', open=False):
+                    num_steps = gr.Slider(label='Number of Steps',
+                                          minimum=0,
+                                          maximum=100,
+                                          step=1,
+                                          value=50)
+                    guidance_scale = gr.Slider(label='CFG Scale',
+                                               minimum=0,
+                                               maximum=50,
+                                               step=0.1,
+                                               value=7)
+
+                run_button = gr.Button('Generate')
+            with gr.Column():
+                result = gr.Image(label='Result')
+
+        reload_button.click(fn=reload_lora_weight_list,
+                            inputs=None,
+                            outputs=lora_weight_name)
+        prompt.submit(fn=pipe.run,
+                      inputs=[
+                          base_model,
+                          lora_weight_name,
+                          prompt,
+                          alpha,
+                          seed,
+                          num_steps,
+                          guidance_scale,
+                      ],
+                      outputs=result,
+                      queue=False)
+        run_button.click(fn=pipe.run,
+                         inputs=[
+                             base_model,
+                             lora_weight_name,
+                             prompt,
+                             alpha,
+                             seed,
+                             num_steps,
+                             guidance_scale,
+                         ],
+                         outputs=result,
+                         queue=False)
+    return demo
+
+
+pipe = InferencePipeline()
+trainer = Trainer()
+
+with gr.Blocks(css='style.css') as demo:
+    if os.getenv('IS_SHARED_UI'):
+        show_warning(SHARED_UI_WARNING)
+    if not torch.cuda.is_available():
+        show_warning(CUDA_NOT_AVAILABLE_WARNING)
+
+    gr.Markdown(TITLE)
+    gr.Markdown(DESCRIPTION)
+
+    with gr.Tabs():
+        with gr.TabItem('Training'):
+            create_training_demo(trainer, pipe)
+        with gr.TabItem('Test'):
+            create_inference_demo(pipe)
+
+    with gr.Accordion('Usage', open=False):
+        gr.Markdown(USAGE_INFO)
+
+demo.queue(default_enabled=False).launch(share=False)

inference.py

@@ -0,0 +1,73 @@
+from __future__ import annotations
+
+import gc
+import pathlib
+import sys
+
+import gradio as gr
+import PIL.Image
+import torch
+from diffusers import StableDiffusionPipeline
+
+sys.path.insert(0, 'lora')
+from lora_diffusion import monkeypatch_lora, tune_lora_scale
+
+
+class InferencePipeline:
+    def __init__(self):
+        self.pipe = None
+        self.device = torch.device(
+            'cuda:0' if torch.cuda.is_available() else 'cpu')
+        self.weight_path = None
+
+    def clear(self) -> None:
+        self.weight_path = None
+        del self.pipe
+        self.pipe = None
+        torch.cuda.empty_cache()
+        gc.collect()
+
+    @staticmethod
+    def get_lora_weight_path(name: str) -> pathlib.Path:
+        curr_dir = pathlib.Path(__file__).parent
+        return curr_dir / name
+
+    def load_pipe(self, model_id: str, lora_filename: str) -> None:
+        weight_path = self.get_lora_weight_path(lora_filename)
+        if weight_path == self.weight_path:
+            return
+        self.weight_path = weight_path
+        lora_weight = torch.load(self.weight_path, map_location=self.device)
+
+        if self.device.type == 'cpu':
+            pipe = StableDiffusionPipeline.from_pretrained(model_id)
+        else:
+            pipe = StableDiffusionPipeline.from_pretrained(
+                model_id, torch_dtype=torch.float16)
+            pipe = pipe.to(self.device)
+
+        monkeypatch_lora(pipe.unet, lora_weight)
+        self.pipe = pipe
+
+    def run(
+        self,
+        base_model: str,
+        lora_weight_name: str,
+        prompt: str,
+        alpha: float,
+        seed: int,
+        n_steps: int,
+        guidance_scale: float,
+    ) -> PIL.Image.Image:
+        if not torch.cuda.is_available():
+            raise gr.Error('CUDA is not available.')
+
+        self.load_pipe(base_model, lora_weight_name)
+
+        generator = torch.Generator(device=self.device).manual_seed(seed)
+        tune_lora_scale(self.pipe.unet, alpha)  # type: ignore
+        out = self.pipe(prompt,
+                        num_inference_steps=n_steps,
+                        guidance_scale=guidance_scale,
+                        generator=generator)  # type: ignore
+        return out.images[0]

lora

@@ -0,0 +1 @@
+Subproject commit ba349e56e23e92e3b128c7c67ae58d3067540daa

requirements.txt

@@ -0,0 +1,10 @@
+accelerate==0.15.0
+bitsandbytes==0.35.4
+diffusers==0.10.2
+ftfy==6.1.1
+Pillow==9.3.0
+torch==1.13.0
+torchvision==0.14.0
+transformers==4.25.1
+triton==2.0.0.dev20220701
+xformers==0.0.13

style.css

@@ -0,0 +1,3 @@
+h1 {
+  text-align: center;
+}

trainer.py

@@ -0,0 +1,109 @@
+from __future__ import annotations
+
+import os
+import pathlib
+import shlex
+import shutil
+import subprocess
+
+import gradio as gr
+import PIL.Image
+import torch
+
+os.environ['PYTHONPATH'] = f'lora:{os.getenv("PYTHONPATH", "")}'
+
+
+def pad_image(image: PIL.Image.Image) -> PIL.Image.Image:
+    w, h = image.size
+    if w == h:
+        return image
+    elif w > h:
+        new_image = PIL.Image.new(image.mode, (w, w), (0, 0, 0))
+        new_image.paste(image, (0, (w - h) // 2))
+        return new_image
+    else:
+        new_image = PIL.Image.new(image.mode, (h, h), (0, 0, 0))
+        new_image.paste(image, ((h - w) // 2, 0))
+        return new_image
+
+
+class Trainer:
+    def __init__(self):
+        self.is_running = False
+        self.is_running_message = 'Another training is in progress.'
+
+        self.instance_data_dir = pathlib.Path('training_data')
+        self.output_dir = pathlib.Path('results')
+
+    def check_if_running(self) -> dict:
+        if self.is_running:
+            return gr.update(value=self.is_running_message)
+        else:
+            return gr.update(value='No training is running.')
+
+    def cleanup_dirs(self) -> None:
+        shutil.rmtree(self.instance_data_dir, ignore_errors=True)
+        shutil.rmtree(self.output_dir, ignore_errors=True)
+
+    def prepare_dataset(self, concept_images: list, resolution: int) -> None:
+        self.instance_data_dir.mkdir()
+        for i, temp_path in enumerate(concept_images):
+            image = PIL.Image.open(temp_path.name)
+            image = pad_image(image)
+            image = image.resize((resolution, resolution))
+            image = image.convert('RGB')
+            out_path = self.instance_data_dir / f'{i:03d}.jpg'
+            image.save(out_path, format='JPEG', quality=100)
+
+    def run(
+        self,
+        base_model: str,
+        resolution_s: str,
+        concept_images: list | None,
+        concept_prompt: str,
+        n_steps: int,
+        learning_rate: float,
+    ) -> tuple[dict, str]:
+        if not torch.cuda.is_available():
+            raise gr.Error('CUDA is not available.')
+
+        out_path = ''
+        if self.is_running:
+            return gr.update(value=self.is_running_message), out_path
+
+        if concept_images is None:
+            raise gr.Error('You need to upload images.')
+        if not concept_prompt:
+            raise gr.Error('The concept prompt is missing.')
+
+        resolution = int(resolution_s)
+
+        self.cleanup_dirs()
+        self.prepare_dataset(concept_images, resolution)
+
+        self.is_running = True
+        command = f'''
+        accelerate launch lora/train_lora_dreambooth.py \
+          --pretrained_model_name_or_path={base_model}  \
+          --instance_data_dir={self.instance_data_dir} \
+          --output_dir={self.output_dir} \
+          --instance_prompt="style of {concept_prompt}" \
+          --resolution={resolution} \
+          --train_batch_size=1 \
+          --gradient_accumulation_steps=1 \
+          --learning_rate={learning_rate} \
+          --lr_scheduler=constant \
+          --lr_warmup_steps=0 \
+          --max_train_steps={n_steps}
+        '''
+        res = subprocess.run(shlex.split(command))
+        self.is_running = False
+
+        if res.returncode == 0:
+            result_message = 'Training Completed!'
+            weight_path = self.output_dir / 'lora_weight.pt'
+            if weight_path.exists():
+                out_path = weight_path.as_posix()
+        else:
+            result_message = 'Training Failed!'
+        return gr.update(value=result_message), out_path