Commit
•
7439e48
1
Parent(s):
391c329
Upload folder using huggingface_hub
Browse filesThis view is limited to 50 files because it contains too many changes.
See raw diff
- .dockerignore +4 -0
- .gitattributes +3 -0
- .gitignore +140 -0
- .pre-commit-config.yaml +46 -0
- .streamlit/config.toml +5 -0
- CODE_OF_CONDUCT.md +128 -0
- Comparisons.md +24 -0
- Dockerfile +51 -0
- FAQ.md +7 -0
- LICENSE +351 -0
- MANIFEST.in +8 -0
- PaperModel.md +76 -0
- README 2.md +202 -0
- README_CN.md +7 -0
- VERSION +1 -0
- about.py +33 -0
- assets/gfpgan_logo.png +0 -0
- cog.yaml +22 -0
- cog_predict.py +161 -0
- experiments/pretrained_models/README.md +7 -0
- gfpgan.egg-info/PKG-INFO +220 -0
- gfpgan.egg-info/SOURCES.txt +34 -0
- gfpgan.egg-info/dependency_links.txt +1 -0
- gfpgan.egg-info/not-zip-safe +1 -0
- gfpgan.egg-info/requires.txt +13 -0
- gfpgan.egg-info/top_level.txt +1 -0
- gfpgan/__init__.py +7 -0
- gfpgan/__pycache__/__init__.cpython-311.pyc +0 -0
- gfpgan/__pycache__/utils.cpython-311.pyc +0 -0
- gfpgan/archs/__init__.py +10 -0
- gfpgan/archs/__pycache__/__init__.cpython-311.pyc +0 -0
- gfpgan/archs/__pycache__/arcface_arch.cpython-311.pyc +0 -0
- gfpgan/archs/__pycache__/gfpgan_bilinear_arch.cpython-311.pyc +0 -0
- gfpgan/archs/__pycache__/gfpganv1_arch.cpython-311.pyc +0 -0
- gfpgan/archs/__pycache__/gfpganv1_clean_arch.cpython-311.pyc +0 -0
- gfpgan/archs/__pycache__/restoreformer_arch.cpython-311.pyc +0 -0
- gfpgan/archs/__pycache__/stylegan2_bilinear_arch.cpython-311.pyc +0 -0
- gfpgan/archs/__pycache__/stylegan2_clean_arch.cpython-311.pyc +0 -0
- gfpgan/archs/arcface_arch.py +245 -0
- gfpgan/archs/gfpgan_bilinear_arch.py +312 -0
- gfpgan/archs/gfpganv1_arch.py +439 -0
- gfpgan/archs/gfpganv1_clean_arch.py +324 -0
- gfpgan/archs/restoreformer_arch.py +658 -0
- gfpgan/archs/stylegan2_bilinear_arch.py +613 -0
- gfpgan/archs/stylegan2_clean_arch.py +368 -0
- gfpgan/data/__init__.py +10 -0
- gfpgan/data/__pycache__/__init__.cpython-311.pyc +0 -0
- gfpgan/data/__pycache__/ffhq_degradation_dataset.cpython-311.pyc +0 -0
- gfpgan/data/ffhq_degradation_dataset.py +230 -0
- gfpgan/models/__init__.py +10 -0
.dockerignore
ADDED
|
@@ -0,0 +1,4 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
__pycache__
|
| 2 |
+
.eggs
|
| 3 |
+
.github
|
| 4 |
+
.vscode
|
.gitattributes
CHANGED
|
@@ -33,3 +33,6 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
|
|
| 33 |
*.zip filter=lfs diff=lfs merge=lfs -text
|
| 34 |
*.zst filter=lfs diff=lfs merge=lfs -text
|
| 35 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
|
|
|
|
|
|
|
|
|
|
|
| 33 |
*.zip filter=lfs diff=lfs merge=lfs -text
|
| 34 |
*.zst filter=lfs diff=lfs merge=lfs -text
|
| 35 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
| 36 |
+
*.png filter=lfs diff=lfs merge=lfs -text
|
| 37 |
+
*.jpeg filter=lfs diff=lfs merge=lfs -text
|
| 38 |
+
*.jpg filter=lfs diff=lfs merge=lfs -text
|
.gitignore
ADDED
|
@@ -0,0 +1,140 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# ignored folders
|
| 2 |
+
datasets/*
|
| 3 |
+
experiments/*
|
| 4 |
+
results/*
|
| 5 |
+
tb_logger/*
|
| 6 |
+
wandb/*
|
| 7 |
+
tmp/*
|
| 8 |
+
|
| 9 |
+
version.py
|
| 10 |
+
|
| 11 |
+
# Byte-compiled / optimized / DLL files
|
| 12 |
+
__pycache__/
|
| 13 |
+
*.py[cod]
|
| 14 |
+
*$py.class
|
| 15 |
+
|
| 16 |
+
# C extensions
|
| 17 |
+
*.so
|
| 18 |
+
|
| 19 |
+
# Distribution / packaging
|
| 20 |
+
.Python
|
| 21 |
+
build/
|
| 22 |
+
develop-eggs/
|
| 23 |
+
dist/
|
| 24 |
+
downloads/
|
| 25 |
+
eggs/
|
| 26 |
+
.eggs/
|
| 27 |
+
lib/
|
| 28 |
+
lib64/
|
| 29 |
+
parts/
|
| 30 |
+
sdist/
|
| 31 |
+
var/
|
| 32 |
+
wheels/
|
| 33 |
+
pip-wheel-metadata/
|
| 34 |
+
share/python-wheels/
|
| 35 |
+
*.egg-info/
|
| 36 |
+
.installed.cfg
|
| 37 |
+
*.egg
|
| 38 |
+
MANIFEST
|
| 39 |
+
|
| 40 |
+
# PyInstaller
|
| 41 |
+
# Usually these files are written by a python script from a template
|
| 42 |
+
# before PyInstaller builds the exe, so as to inject date/other infos into it.
|
| 43 |
+
*.manifest
|
| 44 |
+
*.spec
|
| 45 |
+
|
| 46 |
+
# Installer logs
|
| 47 |
+
pip-log.txt
|
| 48 |
+
pip-delete-this-directory.txt
|
| 49 |
+
|
| 50 |
+
# Unit test / coverage reports
|
| 51 |
+
htmlcov/
|
| 52 |
+
.tox/
|
| 53 |
+
.nox/
|
| 54 |
+
.coverage
|
| 55 |
+
.coverage.*
|
| 56 |
+
.cache
|
| 57 |
+
nosetests.xml
|
| 58 |
+
coverage.xml
|
| 59 |
+
*.cover
|
| 60 |
+
*.py,cover
|
| 61 |
+
.hypothesis/
|
| 62 |
+
.pytest_cache/
|
| 63 |
+
|
| 64 |
+
# Translations
|
| 65 |
+
*.mo
|
| 66 |
+
*.pot
|
| 67 |
+
|
| 68 |
+
# Django stuff:
|
| 69 |
+
*.log
|
| 70 |
+
local_settings.py
|
| 71 |
+
db.sqlite3
|
| 72 |
+
db.sqlite3-journal
|
| 73 |
+
|
| 74 |
+
# Flask stuff:
|
| 75 |
+
instance/
|
| 76 |
+
.webassets-cache
|
| 77 |
+
|
| 78 |
+
# Scrapy stuff:
|
| 79 |
+
.scrapy
|
| 80 |
+
|
| 81 |
+
# Sphinx documentation
|
| 82 |
+
docs/_build/
|
| 83 |
+
|
| 84 |
+
# PyBuilder
|
| 85 |
+
target/
|
| 86 |
+
|
| 87 |
+
# Jupyter Notebook
|
| 88 |
+
.ipynb_checkpoints
|
| 89 |
+
|
| 90 |
+
# IPython
|
| 91 |
+
profile_default/
|
| 92 |
+
ipython_config.py
|
| 93 |
+
|
| 94 |
+
# pyenv
|
| 95 |
+
.python-version
|
| 96 |
+
|
| 97 |
+
# pipenv
|
| 98 |
+
# According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.
|
| 99 |
+
# However, in case of collaboration, if having platform-specific dependencies or dependencies
|
| 100 |
+
# having no cross-platform support, pipenv may install dependencies that don't work, or not
|
| 101 |
+
# install all needed dependencies.
|
| 102 |
+
#Pipfile.lock
|
| 103 |
+
|
| 104 |
+
# PEP 582; used by e.g. github.com/David-OConnor/pyflow
|
| 105 |
+
__pypackages__/
|
| 106 |
+
|
| 107 |
+
# Celery stuff
|
| 108 |
+
celerybeat-schedule
|
| 109 |
+
celerybeat.pid
|
| 110 |
+
|
| 111 |
+
# SageMath parsed files
|
| 112 |
+
*.sage.py
|
| 113 |
+
|
| 114 |
+
# Environments
|
| 115 |
+
.env
|
| 116 |
+
.venv
|
| 117 |
+
env/
|
| 118 |
+
venv/
|
| 119 |
+
ENV/
|
| 120 |
+
env.bak/
|
| 121 |
+
venv.bak/
|
| 122 |
+
|
| 123 |
+
# Spyder project settings
|
| 124 |
+
.spyderproject
|
| 125 |
+
.spyproject
|
| 126 |
+
|
| 127 |
+
# Rope project settings
|
| 128 |
+
.ropeproject
|
| 129 |
+
|
| 130 |
+
# mkdocs documentation
|
| 131 |
+
/site
|
| 132 |
+
|
| 133 |
+
# mypy
|
| 134 |
+
.mypy_cache/
|
| 135 |
+
.dmypy.json
|
| 136 |
+
dmypy.json
|
| 137 |
+
|
| 138 |
+
# Pyre type checker
|
| 139 |
+
.pyre/
|
| 140 |
+
hf.py
|
.pre-commit-config.yaml
ADDED
|
@@ -0,0 +1,46 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
repos:
|
| 2 |
+
# flake8
|
| 3 |
+
- repo: https://github.com/PyCQA/flake8
|
| 4 |
+
rev: 3.8.3
|
| 5 |
+
hooks:
|
| 6 |
+
- id: flake8
|
| 7 |
+
args: ["--config=setup.cfg", "--ignore=W504, W503"]
|
| 8 |
+
|
| 9 |
+
# modify known_third_party
|
| 10 |
+
- repo: https://github.com/asottile/seed-isort-config
|
| 11 |
+
rev: v2.2.0
|
| 12 |
+
hooks:
|
| 13 |
+
- id: seed-isort-config
|
| 14 |
+
|
| 15 |
+
# isort
|
| 16 |
+
- repo: https://github.com/timothycrosley/isort
|
| 17 |
+
rev: 5.2.2
|
| 18 |
+
hooks:
|
| 19 |
+
- id: isort
|
| 20 |
+
|
| 21 |
+
# yapf
|
| 22 |
+
- repo: https://github.com/pre-commit/mirrors-yapf
|
| 23 |
+
rev: v0.30.0
|
| 24 |
+
hooks:
|
| 25 |
+
- id: yapf
|
| 26 |
+
|
| 27 |
+
# codespell
|
| 28 |
+
- repo: https://github.com/codespell-project/codespell
|
| 29 |
+
rev: v2.1.0
|
| 30 |
+
hooks:
|
| 31 |
+
- id: codespell
|
| 32 |
+
|
| 33 |
+
# pre-commit-hooks
|
| 34 |
+
- repo: https://github.com/pre-commit/pre-commit-hooks
|
| 35 |
+
rev: v3.2.0
|
| 36 |
+
hooks:
|
| 37 |
+
- id: trailing-whitespace # Trim trailing whitespace
|
| 38 |
+
- id: check-yaml # Attempt to load all yaml files to verify syntax
|
| 39 |
+
- id: check-merge-conflict # Check for files that contain merge conflict strings
|
| 40 |
+
- id: double-quote-string-fixer # Replace double quoted strings with single quoted strings
|
| 41 |
+
- id: end-of-file-fixer # Make sure files end in a newline and only a newline
|
| 42 |
+
- id: requirements-txt-fixer # Sort entries in requirements.txt and remove incorrect entry for pkg-resources==0.0.0
|
| 43 |
+
- id: fix-encoding-pragma # Remove the coding pragma: # -*- coding: utf-8 -*-
|
| 44 |
+
args: ["--remove"]
|
| 45 |
+
- id: mixed-line-ending # Replace or check mixed line ending
|
| 46 |
+
args: ["--fix=lf"]
|
.streamlit/config.toml
ADDED
|
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[server]
|
| 2 |
+
maxUploadSize = 15
|
| 3 |
+
port = 7860
|
| 4 |
+
enableXsrfProtection = false
|
| 5 |
+
enableCORS = false
|
CODE_OF_CONDUCT.md
ADDED
|
@@ -0,0 +1,128 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Contributor Covenant Code of Conduct
|
| 2 |
+
|
| 3 |
+
## Our Pledge
|
| 4 |
+
|
| 5 |
+
We as members, contributors, and leaders pledge to make participation in our
|
| 6 |
+
community a harassment-free experience for everyone, regardless of age, body
|
| 7 |
+
size, visible or invisible disability, ethnicity, sex characteristics, gender
|
| 8 |
+
identity and expression, level of experience, education, socio-economic status,
|
| 9 |
+
nationality, personal appearance, race, religion, or sexual identity
|
| 10 |
+
and orientation.
|
| 11 |
+
|
| 12 |
+
We pledge to act and interact in ways that contribute to an open, welcoming,
|
| 13 |
+
diverse, inclusive, and healthy community.
|
| 14 |
+
|
| 15 |
+
## Our Standards
|
| 16 |
+
|
| 17 |
+
Examples of behavior that contributes to a positive environment for our
|
| 18 |
+
community include:
|
| 19 |
+
|
| 20 |
+
* Demonstrating empathy and kindness toward other people
|
| 21 |
+
* Being respectful of differing opinions, viewpoints, and experiences
|
| 22 |
+
* Giving and gracefully accepting constructive feedback
|
| 23 |
+
* Accepting responsibility and apologizing to those affected by our mistakes,
|
| 24 |
+
and learning from the experience
|
| 25 |
+
* Focusing on what is best not just for us as individuals, but for the
|
| 26 |
+
overall community
|
| 27 |
+
|
| 28 |
+
Examples of unacceptable behavior include:
|
| 29 |
+
|
| 30 |
+
* The use of sexualized language or imagery, and sexual attention or
|
| 31 |
+
advances of any kind
|
| 32 |
+
* Trolling, insulting or derogatory comments, and personal or political attacks
|
| 33 |
+
* Public or private harassment
|
| 34 |
+
* Publishing others' private information, such as a physical or email
|
| 35 |
+
address, without their explicit permission
|
| 36 |
+
* Other conduct which could reasonably be considered inappropriate in a
|
| 37 |
+
professional setting
|
| 38 |
+
|
| 39 |
+
## Enforcement Responsibilities
|
| 40 |
+
|
| 41 |
+
Community leaders are responsible for clarifying and enforcing our standards of
|
| 42 |
+
acceptable behavior and will take appropriate and fair corrective action in
|
| 43 |
+
response to any behavior that they deem inappropriate, threatening, offensive,
|
| 44 |
+
or harmful.
|
| 45 |
+
|
| 46 |
+
Community leaders have the right and responsibility to remove, edit, or reject
|
| 47 |
+
comments, commits, code, wiki edits, issues, and other contributions that are
|
| 48 |
+
not aligned to this Code of Conduct, and will communicate reasons for moderation
|
| 49 |
+
decisions when appropriate.
|
| 50 |
+
|
| 51 |
+
## Scope
|
| 52 |
+
|
| 53 |
+
This Code of Conduct applies within all community spaces, and also applies when
|
| 54 |
+
an individual is officially representing the community in public spaces.
|
| 55 |
+
Examples of representing our community include using an official e-mail address,
|
| 56 |
+
posting via an official social media account, or acting as an appointed
|
| 57 |
+
representative at an online or offline event.
|
| 58 |
+
|
| 59 |
+
## Enforcement
|
| 60 |
+
|
| 61 |
+
Instances of abusive, harassing, or otherwise unacceptable behavior may be
|
| 62 |
+
reported to the community leaders responsible for enforcement at
|
| 63 |
+
xintao.wang@outlook.com or xintaowang@tencent.com.
|
| 64 |
+
All complaints will be reviewed and investigated promptly and fairly.
|
| 65 |
+
|
| 66 |
+
All community leaders are obligated to respect the privacy and security of the
|
| 67 |
+
reporter of any incident.
|
| 68 |
+
|
| 69 |
+
## Enforcement Guidelines
|
| 70 |
+
|
| 71 |
+
Community leaders will follow these Community Impact Guidelines in determining
|
| 72 |
+
the consequences for any action they deem in violation of this Code of Conduct:
|
| 73 |
+
|
| 74 |
+
### 1. Correction
|
| 75 |
+
|
| 76 |
+
**Community Impact**: Use of inappropriate language or other behavior deemed
|
| 77 |
+
unprofessional or unwelcome in the community.
|
| 78 |
+
|
| 79 |
+
**Consequence**: A private, written warning from community leaders, providing
|
| 80 |
+
clarity around the nature of the violation and an explanation of why the
|
| 81 |
+
behavior was inappropriate. A public apology may be requested.
|
| 82 |
+
|
| 83 |
+
### 2. Warning
|
| 84 |
+
|
| 85 |
+
**Community Impact**: A violation through a single incident or series
|
| 86 |
+
of actions.
|
| 87 |
+
|
| 88 |
+
**Consequence**: A warning with consequences for continued behavior. No
|
| 89 |
+
interaction with the people involved, including unsolicited interaction with
|
| 90 |
+
those enforcing the Code of Conduct, for a specified period of time. This
|
| 91 |
+
includes avoiding interactions in community spaces as well as external channels
|
| 92 |
+
like social media. Violating these terms may lead to a temporary or
|
| 93 |
+
permanent ban.
|
| 94 |
+
|
| 95 |
+
### 3. Temporary Ban
|
| 96 |
+
|
| 97 |
+
**Community Impact**: A serious violation of community standards, including
|
| 98 |
+
sustained inappropriate behavior.
|
| 99 |
+
|
| 100 |
+
**Consequence**: A temporary ban from any sort of interaction or public
|
| 101 |
+
communication with the community for a specified period of time. No public or
|
| 102 |
+
private interaction with the people involved, including unsolicited interaction
|
| 103 |
+
with those enforcing the Code of Conduct, is allowed during this period.
|
| 104 |
+
Violating these terms may lead to a permanent ban.
|
| 105 |
+
|
| 106 |
+
### 4. Permanent Ban
|
| 107 |
+
|
| 108 |
+
**Community Impact**: Demonstrating a pattern of violation of community
|
| 109 |
+
standards, including sustained inappropriate behavior, harassment of an
|
| 110 |
+
individual, or aggression toward or disparagement of classes of individuals.
|
| 111 |
+
|
| 112 |
+
**Consequence**: A permanent ban from any sort of public interaction within
|
| 113 |
+
the community.
|
| 114 |
+
|
| 115 |
+
## Attribution
|
| 116 |
+
|
| 117 |
+
This Code of Conduct is adapted from the [Contributor Covenant][homepage],
|
| 118 |
+
version 2.0, available at
|
| 119 |
+
https://www.contributor-covenant.org/version/2/0/code_of_conduct.html.
|
| 120 |
+
|
| 121 |
+
Community Impact Guidelines were inspired by [Mozilla's code of conduct
|
| 122 |
+
enforcement ladder](https://github.com/mozilla/diversity).
|
| 123 |
+
|
| 124 |
+
[homepage]: https://www.contributor-covenant.org
|
| 125 |
+
|
| 126 |
+
For answers to common questions about this code of conduct, see the FAQ at
|
| 127 |
+
https://www.contributor-covenant.org/faq. Translations are available at
|
| 128 |
+
https://www.contributor-covenant.org/translations.
|
Comparisons.md
ADDED
|
@@ -0,0 +1,24 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Comparisons
|
| 2 |
+
|
| 3 |
+
## Comparisons among different model versions
|
| 4 |
+
|
| 5 |
+
Note that V1.3 is not always better than V1.2. You may need to try different models based on your purpose and inputs.
|
| 6 |
+
|
| 7 |
+
| Version | Strengths | Weaknesses |
|
| 8 |
+
| :---: | :---: | :---: |
|
| 9 |
+
|V1.3 | ✓ natural outputs<br> ✓better results on very low-quality inputs <br> ✓ work on relatively high-quality inputs <br>✓ can have repeated (twice) restorations | ✗ not very sharp <br> ✗ have a slight change on identity |
|
| 10 |
+
|V1.2 | ✓ sharper output <br> ✓ with beauty makeup | ✗ some outputs are unnatural|
|
| 11 |
+
|
| 12 |
+
For the following images, you may need to **zoom in** for comparing details, or **click the image** to see in the full size.
|
| 13 |
+
|
| 14 |
+
| Input | V1 | V1.2 | V1.3
|
| 15 |
+
| :---: | :---: | :---: | :---: |
|
| 16 |
+
||  |  |  |
|
| 17 |
+
|  |  |  | |
|
| 18 |
+
|  |  |  | |
|
| 19 |
+
|  |  |  | |
|
| 20 |
+
|  |  |  | |
|
| 21 |
+
|  |  |  | |
|
| 22 |
+
|  |  |  | |
|
| 23 |
+
|
| 24 |
+
<!-- | ![]() | ![]() | ![]() | ![]()| -->
|
Dockerfile
ADDED
|
@@ -0,0 +1,51 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
FROM python:3.11-slim
|
| 2 |
+
### Set up user with permissions
|
| 3 |
+
# Set up a new user named "user" with user ID 1000
|
| 4 |
+
|
| 5 |
+
RUN apt-get update && apt-get install -y python3-opencv
|
| 6 |
+
RUN pip install opencv-python
|
| 7 |
+
|
| 8 |
+
RUN useradd -m -u 1000 user
|
| 9 |
+
|
| 10 |
+
# Switch to the "user" user
|
| 11 |
+
USER user
|
| 12 |
+
|
| 13 |
+
# Set home to the user's home directory
|
| 14 |
+
ENV HOME=/home/user \
|
| 15 |
+
PATH=/home/user/.local/bin:$PATH
|
| 16 |
+
|
| 17 |
+
# Set the working directory to the user's home directory
|
| 18 |
+
WORKDIR $HOME/app
|
| 19 |
+
|
| 20 |
+
# Copy the current directory contents into the container at $HOME/app setting the owner to the user
|
| 21 |
+
COPY --chown=user . $HOME/app
|
| 22 |
+
|
| 23 |
+
### Set up app-specific content
|
| 24 |
+
COPY req.txt req.txt
|
| 25 |
+
|
| 26 |
+
# Install basicsr - https://github.com/xinntao/BasicSR
|
| 27 |
+
# We use BasicSR for both training and inference
|
| 28 |
+
RUN pip install basicsr
|
| 29 |
+
|
| 30 |
+
# Install facexlib - https://github.com/xinntao/facexlib
|
| 31 |
+
# We use face detection and face restoration helper in the facexlib package
|
| 32 |
+
RUN pip install facexlib
|
| 33 |
+
|
| 34 |
+
RUN pip3 install -r req.txt
|
| 35 |
+
RUN python setup.py develop
|
| 36 |
+
|
| 37 |
+
# If you want to enhance the background (non-face) regions with Real-ESRGAN,
|
| 38 |
+
# you also need to install the realesrgan package
|
| 39 |
+
RUN pip install realesrgan
|
| 40 |
+
|
| 41 |
+
|
| 42 |
+
COPY . .
|
| 43 |
+
|
| 44 |
+
### Update permissions for the app
|
| 45 |
+
USER root
|
| 46 |
+
RUN chmod 777 ~/app/*
|
| 47 |
+
USER user
|
| 48 |
+
|
| 49 |
+
EXPOSE 7860 7860
|
| 50 |
+
ENTRYPOINT ["streamlit", "run"]
|
| 51 |
+
CMD ["streamlit-app.py", "--server.port", "7860"]
|
FAQ.md
ADDED
|
@@ -0,0 +1,7 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# FAQ
|
| 2 |
+
|
| 3 |
+
1. **How to finetune the GFPGANCleanv1-NoCE-C2 (v1.2) model**
|
| 4 |
+
|
| 5 |
+
**A:** 1) The GFPGANCleanv1-NoCE-C2 (v1.2) model uses the *clean* architecture, which is more friendly for deploying.
|
| 6 |
+
2) This model is not directly trained. Instead, it is converted from another *bilinear* model.
|
| 7 |
+
3) If you want to finetune the GFPGANCleanv1-NoCE-C2 (v1.2), you need to finetune its original *bilinear* model, and then do the conversion.
|
LICENSE
ADDED
|
@@ -0,0 +1,351 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
Tencent is pleased to support the open source community by making GFPGAN available.
|
| 2 |
+
|
| 3 |
+
Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
|
| 4 |
+
|
| 5 |
+
GFPGAN is licensed under the Apache License Version 2.0 except for the third-party components listed below.
|
| 6 |
+
|
| 7 |
+
|
| 8 |
+
Terms of the Apache License Version 2.0:
|
| 9 |
+
---------------------------------------------
|
| 10 |
+
Apache License
|
| 11 |
+
|
| 12 |
+
Version 2.0, January 2004
|
| 13 |
+
|
| 14 |
+
http://www.apache.org/licenses/
|
| 15 |
+
|
| 16 |
+
TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
|
| 17 |
+
1. Definitions.
|
| 18 |
+
|
| 19 |
+
“License” shall mean the terms and conditions for use, reproduction, and distribution as defined by Sections 1 through 9 of this document.
|
| 20 |
+
|
| 21 |
+
“Licensor” shall mean the copyright owner or entity authorized by the copyright owner that is granting the License.
|
| 22 |
+
|
| 23 |
+
“Legal Entity” shall mean the union of the acting entity and all other entities that control, are controlled by, or are under common control with that entity. For the purposes of this definition, “control” means (i) the power, direct or indirect, to cause the direction or management of such entity, whether by contract or otherwise, or (ii) ownership of fifty percent (50%) or more of the outstanding shares, or (iii) beneficial ownership of such entity.
|
| 24 |
+
|
| 25 |
+
“You” (or “Your”) shall mean an individual or Legal Entity exercising permissions granted by this License.
|
| 26 |
+
|
| 27 |
+
“Source” form shall mean the preferred form for making modifications, including but not limited to software source code, documentation source, and configuration files.
|
| 28 |
+
|
| 29 |
+
“Object” form shall mean any form resulting from mechanical transformation or translation of a Source form, including but not limited to compiled object code, generated documentation, and conversions to other media types.
|
| 30 |
+
|
| 31 |
+
“Work” shall mean the work of authorship, whether in Source or Object form, made available under the License, as indicated by a copyright notice that is included in or attached to the work (an example is provided in the Appendix below).
|
| 32 |
+
|
| 33 |
+
“Derivative Works” shall mean any work, whether in Source or Object form, that is based on (or derived from) the Work and for which the editorial revisions, annotations, elaborations, or other modifications represent, as a whole, an original work of authorship. For the purposes of this License, Derivative Works shall not include works that remain separable from, or merely link (or bind by name) to the interfaces of, the Work and Derivative Works thereof.
|
| 34 |
+
|
| 35 |
+
“Contribution” shall mean any work of authorship, including the original version of the Work and any modifications or additions to that Work or Derivative Works thereof, that is intentionally submitted to Licensor for inclusion in the Work by the copyright owner or by an individual or Legal Entity authorized to submit on behalf of the copyright owner. For the purposes of this definition, “submitted” means any form of electronic, verbal, or written communication sent to the Licensor or its representatives, including but not limited to communication on electronic mailing lists, source code control systems, and issue tracking systems that are managed by, or on behalf of, the Licensor for the purpose of discussing and improving the Work, but excluding communication that is conspicuously marked or otherwise designated in writing by the copyright owner as “Not a Contribution.”
|
| 36 |
+
|
| 37 |
+
“Contributor” shall mean Licensor and any individual or Legal Entity on behalf of whom a Contribution has been received by Licensor and subsequently incorporated within the Work.
|
| 38 |
+
|
| 39 |
+
2. Grant of Copyright License. Subject to the terms and conditions of this License, each Contributor hereby grants to You a perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable copyright license to reproduce, prepare Derivative Works of, publicly display, publicly perform, sublicense, and distribute the Work and such Derivative Works in Source or Object form.
|
| 40 |
+
|
| 41 |
+
3. Grant of Patent License. Subject to the terms and conditions of this License, each Contributor hereby grants to You a perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable (except as stated in this section) patent license to make, have made, use, offer to sell, sell, import, and otherwise transfer the Work, where such license applies only to those patent claims licensable by such Contributor that are necessarily infringed by their Contribution(s) alone or by combination of their Contribution(s) with the Work to which such Contribution(s) was submitted. If You institute patent litigation against any entity (including a cross-claim or counterclaim in a lawsuit) alleging that the Work or a Contribution incorporated within the Work constitutes direct or contributory patent infringement, then any patent licenses granted to You under this License for that Work shall terminate as of the date such litigation is filed.
|
| 42 |
+
|
| 43 |
+
4. Redistribution. You may reproduce and distribute copies of the Work or Derivative Works thereof in any medium, with or without modifications, and in Source or Object form, provided that You meet the following conditions:
|
| 44 |
+
|
| 45 |
+
You must give any other recipients of the Work or Derivative Works a copy of this License; and
|
| 46 |
+
|
| 47 |
+
You must cause any modified files to carry prominent notices stating that You changed the files; and
|
| 48 |
+
|
| 49 |
+
You must retain, in the Source form of any Derivative Works that You distribute, all copyright, patent, trademark, and attribution notices from the Source form of the Work, excluding those notices that do not pertain to any part of the Derivative Works; and
|
| 50 |
+
|
| 51 |
+
If the Work includes a “NOTICE” text file as part of its distribution, then any Derivative Works that You distribute must include a readable copy of the attribution notices contained within such NOTICE file, excluding those notices that do not pertain to any part of the Derivative Works, in at least one of the following places: within a NOTICE text file distributed as part of the Derivative Works; within the Source form or documentation, if provided along with the Derivative Works; or, within a display generated by the Derivative Works, if and wherever such third-party notices normally appear. The contents of the NOTICE file are for informational purposes only and do not modify the License. You may add Your own attribution notices within Derivative Works that You distribute, alongside or as an addendum to the NOTICE text from the Work, provided that such additional attribution notices cannot be construed as modifying the License.
|
| 52 |
+
|
| 53 |
+
You may add Your own copyright statement to Your modifications and may provide additional or different license terms and conditions for use, reproduction, or distribution of Your modifications, or for any such Derivative Works as a whole, provided Your use, reproduction, and distribution of the Work otherwise complies with the conditions stated in this License.
|
| 54 |
+
|
| 55 |
+
5. Submission of Contributions. Unless You explicitly state otherwise, any Contribution intentionally submitted for inclusion in the Work by You to the Licensor shall be under the terms and conditions of this License, without any additional terms or conditions. Notwithstanding the above, nothing herein shall supersede or modify the terms of any separate license agreement you may have executed with Licensor regarding such Contributions.
|
| 56 |
+
|
| 57 |
+
6. Trademarks. This License does not grant permission to use the trade names, trademarks, service marks, or product names of the Licensor, except as required for reasonable and customary use in describing the origin of the Work and reproducing the content of the NOTICE file.
|
| 58 |
+
|
| 59 |
+
7. Disclaimer of Warranty. Unless required by applicable law or agreed to in writing, Licensor provides the Work (and each Contributor provides its Contributions) on an “AS IS” BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied, including, without limitation, any warranties or conditions of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A PARTICULAR PURPOSE. You are solely responsible for determining the appropriateness of using or redistributing the Work and assume any risks associated with Your exercise of permissions under this License.
|
| 60 |
+
|
| 61 |
+
8. Limitation of Liability. In no event and under no legal theory, whether in tort (including negligence), contract, or otherwise, unless required by applicable law (such as deliberate and grossly negligent acts) or agreed to in writing, shall any Contributor be liable to You for damages, including any direct, indirect, special, incidental, or consequential damages of any character arising as a result of this License or out of the use or inability to use the Work (including but not limited to damages for loss of goodwill, work stoppage, computer failure or malfunction, or any and all other commercial damages or losses), even if such Contributor has been advised of the possibility of such damages.
|
| 62 |
+
|
| 63 |
+
9. Accepting Warranty or Additional Liability. While redistributing the Work or Derivative Works thereof, You may choose to offer, and charge a fee for, acceptance of support, warranty, indemnity, or other liability obligations and/or rights consistent with this License. However, in accepting such obligations, You may act only on Your own behalf and on Your sole responsibility, not on behalf of any other Contributor, and only if You agree to indemnify, defend, and hold each Contributor harmless for any liability incurred by, or claims asserted against, such Contributor by reason of your accepting any such warranty or additional liability.
|
| 64 |
+
|
| 65 |
+
END OF TERMS AND CONDITIONS
|
| 66 |
+
|
| 67 |
+
|
| 68 |
+
|
| 69 |
+
Other dependencies and licenses:
|
| 70 |
+
|
| 71 |
+
|
| 72 |
+
Open Source Software licensed under the Apache 2.0 license and Other Licenses of the Third-Party Components therein:
|
| 73 |
+
---------------------------------------------
|
| 74 |
+
1. basicsr
|
| 75 |
+
Copyright 2018-2020 BasicSR Authors
|
| 76 |
+
|
| 77 |
+
|
| 78 |
+
This BasicSR project is released under the Apache 2.0 license.
|
| 79 |
+
|
| 80 |
+
A copy of Apache 2.0 is included in this file.
|
| 81 |
+
|
| 82 |
+
StyleGAN2
|
| 83 |
+
The codes are modified from the repository stylegan2-pytorch. Many thanks to the author - Kim Seonghyeon 😊 for translating from the official TensorFlow codes to PyTorch ones. Here is the license of stylegan2-pytorch.
|
| 84 |
+
The official repository is https://github.com/NVlabs/stylegan2, and here is the NVIDIA license.
|
| 85 |
+
DFDNet
|
| 86 |
+
The codes are largely modified from the repository DFDNet. Their license is Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License.
|
| 87 |
+
|
| 88 |
+
Terms of the Nvidia License:
|
| 89 |
+
---------------------------------------------
|
| 90 |
+
|
| 91 |
+
1. Definitions
|
| 92 |
+
|
| 93 |
+
"Licensor" means any person or entity that distributes its Work.
|
| 94 |
+
|
| 95 |
+
"Software" means the original work of authorship made available under
|
| 96 |
+
this License.
|
| 97 |
+
|
| 98 |
+
"Work" means the Software and any additions to or derivative works of
|
| 99 |
+
the Software that are made available under this License.
|
| 100 |
+
|
| 101 |
+
"Nvidia Processors" means any central processing unit (CPU), graphics
|
| 102 |
+
processing unit (GPU), field-programmable gate array (FPGA),
|
| 103 |
+
application-specific integrated circuit (ASIC) or any combination
|
| 104 |
+
thereof designed, made, sold, or provided by Nvidia or its affiliates.
|
| 105 |
+
|
| 106 |
+
The terms "reproduce," "reproduction," "derivative works," and
|
| 107 |
+
"distribution" have the meaning as provided under U.S. copyright law;
|
| 108 |
+
provided, however, that for the purposes of this License, derivative
|
| 109 |
+
works shall not include works that remain separable from, or merely
|
| 110 |
+
link (or bind by name) to the interfaces of, the Work.
|
| 111 |
+
|
| 112 |
+
Works, including the Software, are "made available" under this License
|
| 113 |
+
by including in or with the Work either (a) a copyright notice
|
| 114 |
+
referencing the applicability of this License to the Work, or (b) a
|
| 115 |
+
copy of this License.
|
| 116 |
+
|
| 117 |
+
2. License Grants
|
| 118 |
+
|
| 119 |
+
2.1 Copyright Grant. Subject to the terms and conditions of this
|
| 120 |
+
License, each Licensor grants to you a perpetual, worldwide,
|
| 121 |
+
non-exclusive, royalty-free, copyright license to reproduce,
|
| 122 |
+
prepare derivative works of, publicly display, publicly perform,
|
| 123 |
+
sublicense and distribute its Work and any resulting derivative
|
| 124 |
+
works in any form.
|
| 125 |
+
|
| 126 |
+
3. Limitations
|
| 127 |
+
|
| 128 |
+
3.1 Redistribution. You may reproduce or distribute the Work only
|
| 129 |
+
if (a) you do so under this License, (b) you include a complete
|
| 130 |
+
copy of this License with your distribution, and (c) you retain
|
| 131 |
+
without modification any copyright, patent, trademark, or
|
| 132 |
+
attribution notices that are present in the Work.
|
| 133 |
+
|
| 134 |
+
3.2 Derivative Works. You may specify that additional or different
|
| 135 |
+
terms apply to the use, reproduction, and distribution of your
|
| 136 |
+
derivative works of the Work ("Your Terms") only if (a) Your Terms
|
| 137 |
+
provide that the use limitation in Section 3.3 applies to your
|
| 138 |
+
derivative works, and (b) you identify the specific derivative
|
| 139 |
+
works that are subject to Your Terms. Notwithstanding Your Terms,
|
| 140 |
+
this License (including the redistribution requirements in Section
|
| 141 |
+
3.1) will continue to apply to the Work itself.
|
| 142 |
+
|
| 143 |
+
3.3 Use Limitation. The Work and any derivative works thereof only
|
| 144 |
+
may be used or intended for use non-commercially. The Work or
|
| 145 |
+
derivative works thereof may be used or intended for use by Nvidia
|
| 146 |
+
or its affiliates commercially or non-commercially. As used herein,
|
| 147 |
+
"non-commercially" means for research or evaluation purposes only.
|
| 148 |
+
|
| 149 |
+
3.4 Patent Claims. If you bring or threaten to bring a patent claim
|
| 150 |
+
against any Licensor (including any claim, cross-claim or
|
| 151 |
+
counterclaim in a lawsuit) to enforce any patents that you allege
|
| 152 |
+
are infringed by any Work, then your rights under this License from
|
| 153 |
+
such Licensor (including the grants in Sections 2.1 and 2.2) will
|
| 154 |
+
terminate immediately.
|
| 155 |
+
|
| 156 |
+
3.5 Trademarks. This License does not grant any rights to use any
|
| 157 |
+
Licensor's or its affiliates' names, logos, or trademarks, except
|
| 158 |
+
as necessary to reproduce the notices described in this License.
|
| 159 |
+
|
| 160 |
+
3.6 Termination. If you violate any term of this License, then your
|
| 161 |
+
rights under this License (including the grants in Sections 2.1 and
|
| 162 |
+
2.2) will terminate immediately.
|
| 163 |
+
|
| 164 |
+
4. Disclaimer of Warranty.
|
| 165 |
+
|
| 166 |
+
THE WORK IS PROVIDED "AS IS" WITHOUT WARRANTIES OR CONDITIONS OF ANY
|
| 167 |
+
KIND, EITHER EXPRESS OR IMPLIED, INCLUDING WARRANTIES OR CONDITIONS OF
|
| 168 |
+
MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE, TITLE OR
|
| 169 |
+
NON-INFRINGEMENT. YOU BEAR THE RISK OF UNDERTAKING ANY ACTIVITIES UNDER
|
| 170 |
+
THIS LICENSE.
|
| 171 |
+
|
| 172 |
+
5. Limitation of Liability.
|
| 173 |
+
|
| 174 |
+
EXCEPT AS PROHIBITED BY APPLICABLE LAW, IN NO EVENT AND UNDER NO LEGAL
|
| 175 |
+
THEORY, WHETHER IN TORT (INCLUDING NEGLIGENCE), CONTRACT, OR OTHERWISE
|
| 176 |
+
SHALL ANY LICENSOR BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY DIRECT,
|
| 177 |
+
INDIRECT, SPECIAL, INCIDENTAL, OR CONSEQUENTIAL DAMAGES ARISING OUT OF
|
| 178 |
+
OR RELATED TO THIS LICENSE, THE USE OR INABILITY TO USE THE WORK
|
| 179 |
+
(INCLUDING BUT NOT LIMITED TO LOSS OF GOODWILL, BUSINESS INTERRUPTION,
|
| 180 |
+
LOST PROFITS OR DATA, COMPUTER FAILURE OR MALFUNCTION, OR ANY OTHER
|
| 181 |
+
COMMERCIAL DAMAGES OR LOSSES), EVEN IF THE LICENSOR HAS BEEN ADVISED OF
|
| 182 |
+
THE POSSIBILITY OF SUCH DAMAGES.
|
| 183 |
+
|
| 184 |
+
MIT License
|
| 185 |
+
|
| 186 |
+
Copyright (c) 2019 Kim Seonghyeon
|
| 187 |
+
|
| 188 |
+
Permission is hereby granted, free of charge, to any person obtaining a copy
|
| 189 |
+
of this software and associated documentation files (the "Software"), to deal
|
| 190 |
+
in the Software without restriction, including without limitation the rights
|
| 191 |
+
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
|
| 192 |
+
copies of the Software, and to permit persons to whom the Software is
|
| 193 |
+
furnished to do so, subject to the following conditions:
|
| 194 |
+
|
| 195 |
+
The above copyright notice and this permission notice shall be included in all
|
| 196 |
+
copies or substantial portions of the Software.
|
| 197 |
+
|
| 198 |
+
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
|
| 199 |
+
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
|
| 200 |
+
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
|
| 201 |
+
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
|
| 202 |
+
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
|
| 203 |
+
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
|
| 204 |
+
SOFTWARE.
|
| 205 |
+
|
| 206 |
+
|
| 207 |
+
|
| 208 |
+
Open Source Software licensed under the BSD 3-Clause license:
|
| 209 |
+
---------------------------------------------
|
| 210 |
+
1. torchvision
|
| 211 |
+
Copyright (c) Soumith Chintala 2016,
|
| 212 |
+
All rights reserved.
|
| 213 |
+
|
| 214 |
+
2. torch
|
| 215 |
+
Copyright (c) 2016- Facebook, Inc (Adam Paszke)
|
| 216 |
+
Copyright (c) 2014- Facebook, Inc (Soumith Chintala)
|
| 217 |
+
Copyright (c) 2011-2014 Idiap Research Institute (Ronan Collobert)
|
| 218 |
+
Copyright (c) 2012-2014 Deepmind Technologies (Koray Kavukcuoglu)
|
| 219 |
+
Copyright (c) 2011-2012 NEC Laboratories America (Koray Kavukcuoglu)
|
| 220 |
+
Copyright (c) 2011-2013 NYU (Clement Farabet)
|
| 221 |
+
Copyright (c) 2006-2010 NEC Laboratories America (Ronan Collobert, Leon Bottou, Iain Melvin, Jason Weston)
|
| 222 |
+
Copyright (c) 2006 Idiap Research Institute (Samy Bengio)
|
| 223 |
+
Copyright (c) 2001-2004 Idiap Research Institute (Ronan Collobert, Samy Bengio, Johnny Mariethoz)
|
| 224 |
+
|
| 225 |
+
|
| 226 |
+
Terms of the BSD 3-Clause License:
|
| 227 |
+
---------------------------------------------
|
| 228 |
+
Redistribution and use in source and binary forms, with or without modification, are permitted provided that the following conditions are met:
|
| 229 |
+
|
| 230 |
+
1. Redistributions of source code must retain the above copyright notice, this list of conditions and the following disclaimer.
|
| 231 |
+
|
| 232 |
+
2. Redistributions in binary form must reproduce the above copyright notice, this list of conditions and the following disclaimer in the documentation and/or other materials provided with the distribution.
|
| 233 |
+
|
| 234 |
+
3. Neither the name of the copyright holder nor the names of its contributors may be used to endorse or promote products derived from this software without specific prior written permission.
|
| 235 |
+
|
| 236 |
+
THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS “AS IS” AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
| 237 |
+
|
| 238 |
+
|
| 239 |
+
|
| 240 |
+
Open Source Software licensed under the BSD 3-Clause License and Other Licenses of the Third-Party Components therein:
|
| 241 |
+
---------------------------------------------
|
| 242 |
+
1. numpy
|
| 243 |
+
Copyright (c) 2005-2020, NumPy Developers.
|
| 244 |
+
All rights reserved.
|
| 245 |
+
|
| 246 |
+
A copy of BSD 3-Clause License is included in this file.
|
| 247 |
+
|
| 248 |
+
The NumPy repository and source distributions bundle several libraries that are
|
| 249 |
+
compatibly licensed. We list these here.
|
| 250 |
+
|
| 251 |
+
Name: Numpydoc
|
| 252 |
+
Files: doc/sphinxext/numpydoc/*
|
| 253 |
+
License: BSD-2-Clause
|
| 254 |
+
For details, see doc/sphinxext/LICENSE.txt
|
| 255 |
+
|
| 256 |
+
Name: scipy-sphinx-theme
|
| 257 |
+
Files: doc/scipy-sphinx-theme/*
|
| 258 |
+
License: BSD-3-Clause AND PSF-2.0 AND Apache-2.0
|
| 259 |
+
For details, see doc/scipy-sphinx-theme/LICENSE.txt
|
| 260 |
+
|
| 261 |
+
Name: lapack-lite
|
| 262 |
+
Files: numpy/linalg/lapack_lite/*
|
| 263 |
+
License: BSD-3-Clause
|
| 264 |
+
For details, see numpy/linalg/lapack_lite/LICENSE.txt
|
| 265 |
+
|
| 266 |
+
Name: tempita
|
| 267 |
+
Files: tools/npy_tempita/*
|
| 268 |
+
License: MIT
|
| 269 |
+
For details, see tools/npy_tempita/license.txt
|
| 270 |
+
|
| 271 |
+
Name: dragon4
|
| 272 |
+
Files: numpy/core/src/multiarray/dragon4.c
|
| 273 |
+
License: MIT
|
| 274 |
+
For license text, see numpy/core/src/multiarray/dragon4.c
|
| 275 |
+
|
| 276 |
+
|
| 277 |
+
|
| 278 |
+
Open Source Software licensed under the MIT license:
|
| 279 |
+
---------------------------------------------
|
| 280 |
+
1. facexlib
|
| 281 |
+
Copyright (c) 2020 Xintao Wang
|
| 282 |
+
|
| 283 |
+
2. opencv-python
|
| 284 |
+
Copyright (c) Olli-Pekka Heinisuo
|
| 285 |
+
Please note that only files in cv2 package are used.
|
| 286 |
+
|
| 287 |
+
|
| 288 |
+
Terms of the MIT License:
|
| 289 |
+
---------------------------------------------
|
| 290 |
+
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:
|
| 291 |
+
|
| 292 |
+
The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.
|
| 293 |
+
|
| 294 |
+
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.
|
| 295 |
+
|
| 296 |
+
|
| 297 |
+
|
| 298 |
+
Open Source Software licensed under the MIT license and Other Licenses of the Third-Party Components therein:
|
| 299 |
+
---------------------------------------------
|
| 300 |
+
1. tqdm
|
| 301 |
+
Copyright (c) 2013 noamraph
|
| 302 |
+
|
| 303 |
+
`tqdm` is a product of collaborative work.
|
| 304 |
+
Unless otherwise stated, all authors (see commit logs) retain copyright
|
| 305 |
+
for their respective work, and release the work under the MIT licence
|
| 306 |
+
(text below).
|
| 307 |
+
|
| 308 |
+
Exceptions or notable authors are listed below
|
| 309 |
+
in reverse chronological order:
|
| 310 |
+
|
| 311 |
+
* files: *
|
| 312 |
+
MPLv2.0 2015-2020 (c) Casper da Costa-Luis
|
| 313 |
+
[casperdcl](https://github.com/casperdcl).
|
| 314 |
+
* files: tqdm/_tqdm.py
|
| 315 |
+
MIT 2016 (c) [PR #96] on behalf of Google Inc.
|
| 316 |
+
* files: tqdm/_tqdm.py setup.py README.rst MANIFEST.in .gitignore
|
| 317 |
+
MIT 2013 (c) Noam Yorav-Raphael, original author.
|
| 318 |
+
|
| 319 |
+
[PR #96]: https://github.com/tqdm/tqdm/pull/96
|
| 320 |
+
|
| 321 |
+
|
| 322 |
+
Mozilla Public Licence (MPL) v. 2.0 - Exhibit A
|
| 323 |
+
-----------------------------------------------
|
| 324 |
+
|
| 325 |
+
This Source Code Form is subject to the terms of the
|
| 326 |
+
Mozilla Public License, v. 2.0.
|
| 327 |
+
If a copy of the MPL was not distributed with this file,
|
| 328 |
+
You can obtain one at https://mozilla.org/MPL/2.0/.
|
| 329 |
+
|
| 330 |
+
|
| 331 |
+
MIT License (MIT)
|
| 332 |
+
-----------------
|
| 333 |
+
|
| 334 |
+
Copyright (c) 2013 noamraph
|
| 335 |
+
|
| 336 |
+
Permission is hereby granted, free of charge, to any person obtaining a copy of
|
| 337 |
+
this software and associated documentation files (the "Software"), to deal in
|
| 338 |
+
the Software without restriction, including without limitation the rights to
|
| 339 |
+
use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of
|
| 340 |
+
the Software, and to permit persons to whom the Software is furnished to do so,
|
| 341 |
+
subject to the following conditions:
|
| 342 |
+
|
| 343 |
+
The above copyright notice and this permission notice shall be included in all
|
| 344 |
+
copies or substantial portions of the Software.
|
| 345 |
+
|
| 346 |
+
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
|
| 347 |
+
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS
|
| 348 |
+
FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR
|
| 349 |
+
COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER
|
| 350 |
+
IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
|
| 351 |
+
CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
|
MANIFEST.in
ADDED
|
@@ -0,0 +1,8 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
include assets/*
|
| 2 |
+
include inputs/*
|
| 3 |
+
include scripts/*.py
|
| 4 |
+
include inference_gfpgan.py
|
| 5 |
+
include VERSION
|
| 6 |
+
include LICENSE
|
| 7 |
+
include requirements.txt
|
| 8 |
+
include gfpgan/weights/README.md
|
PaperModel.md
ADDED
|
@@ -0,0 +1,76 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Installation
|
| 2 |
+
|
| 3 |
+
We now provide a *clean* version of GFPGAN, which does not require customized CUDA extensions. See [here](README.md#installation) for this easier installation.<br>
|
| 4 |
+
If you want want to use the original model in our paper, please follow the instructions below.
|
| 5 |
+
|
| 6 |
+
1. Clone repo
|
| 7 |
+
|
| 8 |
+
```bash
|
| 9 |
+
git clone https://github.com/xinntao/GFPGAN.git
|
| 10 |
+
cd GFPGAN
|
| 11 |
+
```
|
| 12 |
+
|
| 13 |
+
1. Install dependent packages
|
| 14 |
+
|
| 15 |
+
As StyleGAN2 uses customized PyTorch C++ extensions, you need to **compile them during installation** or **load them just-in-time(JIT)**.
|
| 16 |
+
You can refer to [BasicSR-INSTALL.md](https://github.com/xinntao/BasicSR/blob/master/INSTALL.md) for more details.
|
| 17 |
+
|
| 18 |
+
**Option 1: Load extensions just-in-time(JIT)** (For those just want to do simple inferences, may have less issues)
|
| 19 |
+
|
| 20 |
+
```bash
|
| 21 |
+
# Install basicsr - https://github.com/xinntao/BasicSR
|
| 22 |
+
# We use BasicSR for both training and inference
|
| 23 |
+
pip install basicsr
|
| 24 |
+
|
| 25 |
+
# Install facexlib - https://github.com/xinntao/facexlib
|
| 26 |
+
# We use face detection and face restoration helper in the facexlib package
|
| 27 |
+
pip install facexlib
|
| 28 |
+
|
| 29 |
+
pip install -r requirements.txt
|
| 30 |
+
python setup.py develop
|
| 31 |
+
|
| 32 |
+
# remember to set BASICSR_JIT=True before your running commands
|
| 33 |
+
```
|
| 34 |
+
|
| 35 |
+
**Option 2: Compile extensions during installation** (For those need to train/inference for many times)
|
| 36 |
+
|
| 37 |
+
```bash
|
| 38 |
+
# Install basicsr - https://github.com/xinntao/BasicSR
|
| 39 |
+
# We use BasicSR for both training and inference
|
| 40 |
+
# Set BASICSR_EXT=True to compile the cuda extensions in the BasicSR - It may take several minutes to compile, please be patient
|
| 41 |
+
# Add -vvv for detailed log prints
|
| 42 |
+
BASICSR_EXT=True pip install basicsr -vvv
|
| 43 |
+
|
| 44 |
+
# Install facexlib - https://github.com/xinntao/facexlib
|
| 45 |
+
# We use face detection and face restoration helper in the facexlib package
|
| 46 |
+
pip install facexlib
|
| 47 |
+
|
| 48 |
+
pip install -r requirements.txt
|
| 49 |
+
python setup.py develop
|
| 50 |
+
```
|
| 51 |
+
|
| 52 |
+
## :zap: Quick Inference
|
| 53 |
+
|
| 54 |
+
Download pre-trained models: [GFPGANv1.pth](https://github.com/TencentARC/GFPGAN/releases/download/v0.1.0/GFPGANv1.pth)
|
| 55 |
+
|
| 56 |
+
```bash
|
| 57 |
+
wget https://github.com/TencentARC/GFPGAN/releases/download/v0.1.0/GFPGANv1.pth -P experiments/pretrained_models
|
| 58 |
+
```
|
| 59 |
+
|
| 60 |
+
- Option 1: Load extensions just-in-time(JIT)
|
| 61 |
+
|
| 62 |
+
```bash
|
| 63 |
+
BASICSR_JIT=True python inference_gfpgan.py --input inputs/whole_imgs --output results --version 1
|
| 64 |
+
|
| 65 |
+
# for aligned images
|
| 66 |
+
BASICSR_JIT=True python inference_gfpgan.py --input inputs/whole_imgs --output results --version 1 --aligned
|
| 67 |
+
```
|
| 68 |
+
|
| 69 |
+
- Option 2: Have successfully compiled extensions during installation
|
| 70 |
+
|
| 71 |
+
```bash
|
| 72 |
+
python inference_gfpgan.py --input inputs/whole_imgs --output results --version 1
|
| 73 |
+
|
| 74 |
+
# for aligned images
|
| 75 |
+
python inference_gfpgan.py --input inputs/whole_imgs --output results --version 1 --aligned
|
| 76 |
+
```
|
README 2.md
ADDED
|
@@ -0,0 +1,202 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
<p align="center">
|
| 2 |
+
<img src="assets/gfpgan_logo.png" height=130>
|
| 3 |
+
</p>
|
| 4 |
+
|
| 5 |
+
## <div align="center"><b><a href="README.md">English</a> | <a href="README_CN.md">简体中文</a></b></div>
|
| 6 |
+
|
| 7 |
+
<div align="center">
|
| 8 |
+
<!-- <a href="https://twitter.com/_Xintao_" style="text-decoration:none;">
|
| 9 |
+
<img src="https://user-images.githubusercontent.com/17445847/187162058-c764ced6-952f-404b-ac85-ba95cce18e7b.png" width="4%" alt="" />
|
| 10 |
+
</a> -->
|
| 11 |
+
|
| 12 |
+
[](https://github.com/TencentARC/GFPGAN/releases)
|
| 13 |
+
[](https://pypi.org/project/gfpgan/)
|
| 14 |
+
[](https://github.com/TencentARC/GFPGAN/issues)
|
| 15 |
+
[](https://github.com/TencentARC/GFPGAN/issues)
|
| 16 |
+
[](https://github.com/TencentARC/GFPGAN/blob/master/LICENSE)
|
| 17 |
+
[](https://github.com/TencentARC/GFPGAN/blob/master/.github/workflows/pylint.yml)
|
| 18 |
+
[](https://github.com/TencentARC/GFPGAN/blob/master/.github/workflows/publish-pip.yml)
|
| 19 |
+
</div>
|
| 20 |
+
|
| 21 |
+
1. :boom: **Updated** online demo: [](https://replicate.com/tencentarc/gfpgan). Here is the [backup](https://replicate.com/xinntao/gfpgan).
|
| 22 |
+
1. :boom: **Updated** online demo: [](https://huggingface.co/spaces/Xintao/GFPGAN)
|
| 23 |
+
1. [Colab Demo](https://colab.research.google.com/drive/1sVsoBd9AjckIXThgtZhGrHRfFI6UUYOo) for GFPGAN <a href="https://colab.research.google.com/drive/1sVsoBd9AjckIXThgtZhGrHRfFI6UUYOo"><img src="https://colab.research.google.com/assets/colab-badge.svg" alt="google colab logo"></a>; (Another [Colab Demo](https://colab.research.google.com/drive/1Oa1WwKB4M4l1GmR7CtswDVgOCOeSLChA?usp=sharing) for the original paper model)
|
| 24 |
+
|
| 25 |
+
<!-- 3. Online demo: [Replicate.ai](https://replicate.com/xinntao/gfpgan) (may need to sign in, return the whole image)
|
| 26 |
+
4. Online demo: [Baseten.co](https://app.baseten.co/applications/Q04Lz0d/operator_views/8qZG6Bg) (backed by GPU, returns the whole image)
|
| 27 |
+
5. We provide a *clean* version of GFPGAN, which can run without CUDA extensions. So that it can run in **Windows** or on **CPU mode**. -->
|
| 28 |
+
|
| 29 |
+
> :rocket: **Thanks for your interest in our work. You may also want to check our new updates on the *tiny models* for *anime images and videos* in [Real-ESRGAN](https://github.com/xinntao/Real-ESRGAN/blob/master/docs/anime_video_model.md)** :blush:
|
| 30 |
+
|
| 31 |
+
GFPGAN aims at developing a **Practical Algorithm for Real-world Face Restoration**.<br>
|
| 32 |
+
It leverages rich and diverse priors encapsulated in a pretrained face GAN (*e.g.*, StyleGAN2) for blind face restoration.
|
| 33 |
+
|
| 34 |
+
:question: Frequently Asked Questions can be found in [FAQ.md](FAQ.md).
|
| 35 |
+
|
| 36 |
+
:triangular_flag_on_post: **Updates**
|
| 37 |
+
|
| 38 |
+
- :white_check_mark: Add [RestoreFormer](https://github.com/wzhouxiff/RestoreFormer) inference codes.
|
| 39 |
+
- :white_check_mark: Add [V1.4 model](https://github.com/TencentARC/GFPGAN/releases/download/v1.3.0/GFPGANv1.4.pth), which produces slightly more details and better identity than V1.3.
|
| 40 |
+
- :white_check_mark: Add **[V1.3 model](https://github.com/TencentARC/GFPGAN/releases/download/v1.3.0/GFPGANv1.3.pth)**, which produces **more natural** restoration results, and better results on *very low-quality* / *high-quality* inputs. See more in [Model zoo](#european_castle-model-zoo), [Comparisons.md](Comparisons.md)
|
| 41 |
+
- :white_check_mark: Integrated to [Huggingface Spaces](https://huggingface.co/spaces) with [Gradio](https://github.com/gradio-app/gradio). See [Gradio Web Demo](https://huggingface.co/spaces/akhaliq/GFPGAN).
|
| 42 |
+
- :white_check_mark: Support enhancing non-face regions (background) with [Real-ESRGAN](https://github.com/xinntao/Real-ESRGAN).
|
| 43 |
+
- :white_check_mark: We provide a *clean* version of GFPGAN, which does not require CUDA extensions.
|
| 44 |
+
- :white_check_mark: We provide an updated model without colorizing faces.
|
| 45 |
+
|
| 46 |
+
---
|
| 47 |
+
|
| 48 |
+
If GFPGAN is helpful in your photos/projects, please help to :star: this repo or recommend it to your friends. Thanks:blush:
|
| 49 |
+
Other recommended projects:<br>
|
| 50 |
+
:arrow_forward: [Real-ESRGAN](https://github.com/xinntao/Real-ESRGAN): A practical algorithm for general image restoration<br>
|
| 51 |
+
:arrow_forward: [BasicSR](https://github.com/xinntao/BasicSR): An open-source image and video restoration toolbox<br>
|
| 52 |
+
:arrow_forward: [facexlib](https://github.com/xinntao/facexlib): A collection that provides useful face-relation functions<br>
|
| 53 |
+
:arrow_forward: [HandyView](https://github.com/xinntao/HandyView): A PyQt5-based image viewer that is handy for view and comparison<br>
|
| 54 |
+
|
| 55 |
+
---
|
| 56 |
+
|
| 57 |
+
### :book: GFP-GAN: Towards Real-World Blind Face Restoration with Generative Facial Prior
|
| 58 |
+
|
| 59 |
+
> [[Paper](https://arxiv.org/abs/2101.04061)]   [[Project Page](https://xinntao.github.io/projects/gfpgan)]   [Demo] <br>
|
| 60 |
+
> [Xintao Wang](https://xinntao.github.io/), [Yu Li](https://yu-li.github.io/), [Honglun Zhang](https://scholar.google.com/citations?hl=en&user=KjQLROoAAAAJ), [Ying Shan](https://scholar.google.com/citations?user=4oXBp9UAAAAJ&hl=en) <br>
|
| 61 |
+
> Applied Research Center (ARC), Tencent PCG
|
| 62 |
+
|
| 63 |
+
<p align="center">
|
| 64 |
+
<img src="https://xinntao.github.io/projects/GFPGAN_src/gfpgan_teaser.jpg">
|
| 65 |
+
</p>
|
| 66 |
+
|
| 67 |
+
---
|
| 68 |
+
|
| 69 |
+
## :wrench: Dependencies and Installation
|
| 70 |
+
|
| 71 |
+
- Python >= 3.7 (Recommend to use [Anaconda](https://www.anaconda.com/download/#linux) or [Miniconda](https://docs.conda.io/en/latest/miniconda.html))
|
| 72 |
+
- [PyTorch >= 1.7](https://pytorch.org/)
|
| 73 |
+
- Option: NVIDIA GPU + [CUDA](https://developer.nvidia.com/cuda-downloads)
|
| 74 |
+
- Option: Linux
|
| 75 |
+
|
| 76 |
+
### Installation
|
| 77 |
+
|
| 78 |
+
We now provide a *clean* version of GFPGAN, which does not require customized CUDA extensions. <br>
|
| 79 |
+
If you want to use the original model in our paper, please see [PaperModel.md](PaperModel.md) for installation.
|
| 80 |
+
|
| 81 |
+
1. Clone repo
|
| 82 |
+
|
| 83 |
+
```bash
|
| 84 |
+
git clone https://github.com/TencentARC/GFPGAN.git
|
| 85 |
+
cd GFPGAN
|
| 86 |
+
```
|
| 87 |
+
|
| 88 |
+
1. Install dependent packages
|
| 89 |
+
|
| 90 |
+
```bash
|
| 91 |
+
# Install basicsr - https://github.com/xinntao/BasicSR
|
| 92 |
+
# We use BasicSR for both training and inference
|
| 93 |
+
pip install basicsr
|
| 94 |
+
|
| 95 |
+
# Install facexlib - https://github.com/xinntao/facexlib
|
| 96 |
+
# We use face detection and face restoration helper in the facexlib package
|
| 97 |
+
pip install facexlib
|
| 98 |
+
|
| 99 |
+
pip install -r requirements.txt
|
| 100 |
+
python setup.py develop
|
| 101 |
+
|
| 102 |
+
# If you want to enhance the background (non-face) regions with Real-ESRGAN,
|
| 103 |
+
# you also need to install the realesrgan package
|
| 104 |
+
pip install realesrgan
|
| 105 |
+
```
|
| 106 |
+
|
| 107 |
+
## :zap: Quick Inference
|
| 108 |
+
|
| 109 |
+
We take the v1.3 version for an example. More models can be found [here](#european_castle-model-zoo).
|
| 110 |
+
|
| 111 |
+
Download pre-trained models: [GFPGANv1.3.pth](https://github.com/TencentARC/GFPGAN/releases/download/v1.3.0/GFPGANv1.3.pth)
|
| 112 |
+
|
| 113 |
+
```bash
|
| 114 |
+
wget https://github.com/TencentARC/GFPGAN/releases/download/v1.3.0/GFPGANv1.3.pth -P experiments/pretrained_models
|
| 115 |
+
```
|
| 116 |
+
|
| 117 |
+
**Inference!**
|
| 118 |
+
|
| 119 |
+
```bash
|
| 120 |
+
python inference_gfpgan.py -i inputs/whole_imgs -o results -v 1.3 -s 2
|
| 121 |
+
```
|
| 122 |
+
|
| 123 |
+
```console
|
| 124 |
+
Usage: python inference_gfpgan.py -i inputs/whole_imgs -o results -v 1.3 -s 2 [options]...
|
| 125 |
+
|
| 126 |
+
-h show this help
|
| 127 |
+
-i input Input image or folder. Default: inputs/whole_imgs
|
| 128 |
+
-o output Output folder. Default: results
|
| 129 |
+
-v version GFPGAN model version. Option: 1 | 1.2 | 1.3. Default: 1.3
|
| 130 |
+
-s upscale The final upsampling scale of the image. Default: 2
|
| 131 |
+
-bg_upsampler background upsampler. Default: realesrgan
|
| 132 |
+
-bg_tile Tile size for background sampler, 0 for no tile during testing. Default: 400
|
| 133 |
+
-suffix Suffix of the restored faces
|
| 134 |
+
-only_center_face Only restore the center face
|
| 135 |
+
-aligned Input are aligned faces
|
| 136 |
+
-ext Image extension. Options: auto | jpg | png, auto means using the same extension as inputs. Default: auto
|
| 137 |
+
```
|
| 138 |
+
|
| 139 |
+
If you want to use the original model in our paper, please see [PaperModel.md](PaperModel.md) for installation and inference.
|
| 140 |
+
|
| 141 |
+
## :european_castle: Model Zoo
|
| 142 |
+
|
| 143 |
+
| Version | Model Name | Description |
|
| 144 |
+
| :---: | :---: | :---: |
|
| 145 |
+
| V1.3 | [GFPGANv1.3.pth](https://github.com/TencentARC/GFPGAN/releases/download/v1.3.0/GFPGANv1.3.pth) | Based on V1.2; **more natural** restoration results; better results on very low-quality / high-quality inputs. |
|
| 146 |
+
| V1.2 | [GFPGANCleanv1-NoCE-C2.pth](https://github.com/TencentARC/GFPGAN/releases/download/v0.2.0/GFPGANCleanv1-NoCE-C2.pth) | No colorization; no CUDA extensions are required. Trained with more data with pre-processing. |
|
| 147 |
+
| V1 | [GFPGANv1.pth](https://github.com/TencentARC/GFPGAN/releases/download/v0.1.0/GFPGANv1.pth) | The paper model, with colorization. |
|
| 148 |
+
|
| 149 |
+
The comparisons are in [Comparisons.md](Comparisons.md).
|
| 150 |
+
|
| 151 |
+
Note that V1.3 is not always better than V1.2. You may need to select different models based on your purpose and inputs.
|
| 152 |
+
|
| 153 |
+
| Version | Strengths | Weaknesses |
|
| 154 |
+
| :---: | :---: | :---: |
|
| 155 |
+
|V1.3 | ✓ natural outputs<br> ✓better results on very low-quality inputs <br> ✓ work on relatively high-quality inputs <br>✓ can have repeated (twice) restorations | ✗ not very sharp <br> ✗ have a slight change on identity |
|
| 156 |
+
|V1.2 | ✓ sharper output <br> ✓ with beauty makeup | ✗ some outputs are unnatural |
|
| 157 |
+
|
| 158 |
+
You can find **more models (such as the discriminators)** here: [[Google Drive](https://drive.google.com/drive/folders/17rLiFzcUMoQuhLnptDsKolegHWwJOnHu?usp=sharing)], OR [[Tencent Cloud 腾讯微云](https://share.weiyun.com/ShYoCCoc)]
|
| 159 |
+
|
| 160 |
+
## :computer: Training
|
| 161 |
+
|
| 162 |
+
We provide the training codes for GFPGAN (used in our paper). <br>
|
| 163 |
+
You could improve it according to your own needs.
|
| 164 |
+
|
| 165 |
+
**Tips**
|
| 166 |
+
|
| 167 |
+
1. More high quality faces can improve the restoration quality.
|
| 168 |
+
2. You may need to perform some pre-processing, such as beauty makeup.
|
| 169 |
+
|
| 170 |
+
**Procedures**
|
| 171 |
+
|
| 172 |
+
(You can try a simple version ( `options/train_gfpgan_v1_simple.yml`) that does not require face component landmarks.)
|
| 173 |
+
|
| 174 |
+
1. Dataset preparation: [FFHQ](https://github.com/NVlabs/ffhq-dataset)
|
| 175 |
+
|
| 176 |
+
1. Download pre-trained models and other data. Put them in the `experiments/pretrained_models` folder.
|
| 177 |
+
1. [Pre-trained StyleGAN2 model: StyleGAN2_512_Cmul1_FFHQ_B12G4_scratch_800k.pth](https://github.com/TencentARC/GFPGAN/releases/download/v0.1.0/StyleGAN2_512_Cmul1_FFHQ_B12G4_scratch_800k.pth)
|
| 178 |
+
1. [Component locations of FFHQ: FFHQ_eye_mouth_landmarks_512.pth](https://github.com/TencentARC/GFPGAN/releases/download/v0.1.0/FFHQ_eye_mouth_landmarks_512.pth)
|
| 179 |
+
1. [A simple ArcFace model: arcface_resnet18.pth](https://github.com/TencentARC/GFPGAN/releases/download/v0.1.0/arcface_resnet18.pth)
|
| 180 |
+
|
| 181 |
+
1. Modify the configuration file `options/train_gfpgan_v1.yml` accordingly.
|
| 182 |
+
|
| 183 |
+
1. Training
|
| 184 |
+
|
| 185 |
+
> python -m torch.distributed.launch --nproc_per_node=4 --master_port=22021 gfpgan/train.py -opt options/train_gfpgan_v1.yml --launcher pytorch
|
| 186 |
+
|
| 187 |
+
## :scroll: License and Acknowledgement
|
| 188 |
+
|
| 189 |
+
GFPGAN is released under Apache License Version 2.0.
|
| 190 |
+
|
| 191 |
+
## BibTeX
|
| 192 |
+
|
| 193 |
+
@InProceedings{wang2021gfpgan,
|
| 194 |
+
author = {Xintao Wang and Yu Li and Honglun Zhang and Ying Shan},
|
| 195 |
+
title = {Towards Real-World Blind Face Restoration with Generative Facial Prior},
|
| 196 |
+
booktitle={The IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
|
| 197 |
+
year = {2021}
|
| 198 |
+
}
|
| 199 |
+
|
| 200 |
+
## :e-mail: Contact
|
| 201 |
+
|
| 202 |
+
If you have any question, please email `xintao.wang@outlook.com` or `xintaowang@tencent.com`.
|
README_CN.md
ADDED
|
@@ -0,0 +1,7 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
<p align="center">
|
| 2 |
+
<img src="assets/gfpgan_logo.png" height=130>
|
| 3 |
+
</p>
|
| 4 |
+
|
| 5 |
+
## <div align="center"><b><a href="README.md">English</a> | <a href="README_CN.md">简体中文</a></b></div>
|
| 6 |
+
|
| 7 |
+
还未完工，欢迎贡献！
|
VERSION
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
1.3.8
|
about.py
ADDED
|
@@ -0,0 +1,33 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import streamlit as st
|
| 2 |
+
|
| 3 |
+
|
| 4 |
+
def about():
|
| 5 |
+
st.markdown("""
|
| 6 |
+
# About Me
|
| 7 |
+
|
| 8 |
+
## Social Links
|
| 9 |
+
|
| 10 |
+
- **GitHub**: [https://www.github.com/prabhukiran8790](https://www.github.com/prabhukiran8790)
|
| 11 |
+
- **LinkedIn**: [https://www.linkedin.com/in/prabhukirankonda](https://www.linkedin.com/in/prabhukirankonda)
|
| 12 |
+
- **Twitter**: [https://twitter.com/prabhukirantwt](https://twitter.com/prabhukirantwt)
|
| 13 |
+
- **Email**: [prabhukiran426@gmail.com](mailto:prabhukiran426@gmail.com)
|
| 14 |
+
|
| 15 |
+
## Website
|
| 16 |
+
|
| 17 |
+
Visit my website at [https://prabhukirankonda.vercel.app](https://prabhukirankonda.vercel.app).
|
| 18 |
+
|
| 19 |
+
# About the Project
|
| 20 |
+
|
| 21 |
+
This project is focused on GFPGAN, a remarkable creation by the ARC Lab at Tencent PCG. GFPGAN stands for **Generative Face Progressive Growing Adversarial Network**. It is an advanced face restoration model designed for various applications, with a primary focus on restoring the appearance of faces in old photographs and enhancing AI-generated faces.
|
| 22 |
+
|
| 23 |
+
## Key Features
|
| 24 |
+
|
| 25 |
+
- **Face Restoration**: GFPGAN is an innovative solution for restoring the quality and appearance of faces in vintage or damaged photographs.
|
| 26 |
+
- **AI-Generated Faces Enhancement**: It can also be used to enhance the realism and quality of AI-generated faces, making them look more natural and lifelike.
|
| 27 |
+
- **Practical Application**: This algorithm provides a practical and efficient approach to improving the visual quality of facial images, which can have numerous applications in image editing and restoration.
|
| 28 |
+
|
| 29 |
+
GFPGAN represents a significant advancement in the field of computer vision and image processing, opening up new possibilities for improving the visual quality of images containing faces.
|
| 30 |
+
|
| 31 |
+
For more details and updates about GFPGAN [github.com/TencentARC/GFPGAN](https://github.com/TencentARC/GFPGAN)
|
| 32 |
+
|
| 33 |
+
""")
|
assets/gfpgan_logo.png
ADDED
|
cog.yaml
ADDED
|
@@ -0,0 +1,22 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# This file is used for constructing replicate env
|
| 2 |
+
image: "r8.im/tencentarc/gfpgan"
|
| 3 |
+
|
| 4 |
+
build:
|
| 5 |
+
gpu: true
|
| 6 |
+
python_version: "3.8"
|
| 7 |
+
system_packages:
|
| 8 |
+
- "libgl1-mesa-glx"
|
| 9 |
+
- "libglib2.0-0"
|
| 10 |
+
python_packages:
|
| 11 |
+
- "torch==1.7.1"
|
| 12 |
+
- "torchvision==0.8.2"
|
| 13 |
+
- "numpy==1.21.1"
|
| 14 |
+
- "lmdb==1.2.1"
|
| 15 |
+
- "opencv-python==4.5.3.56"
|
| 16 |
+
- "PyYAML==5.4.1"
|
| 17 |
+
- "tqdm==4.62.2"
|
| 18 |
+
- "yapf==0.31.0"
|
| 19 |
+
- "basicsr==1.4.2"
|
| 20 |
+
- "facexlib==0.2.5"
|
| 21 |
+
|
| 22 |
+
predict: "cog_predict.py:Predictor"
|
cog_predict.py
ADDED
|
@@ -0,0 +1,161 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# flake8: noqa
|
| 2 |
+
# This file is used for deploying replicate models
|
| 3 |
+
# running: cog predict -i img=@inputs/whole_imgs/10045.png -i version='v1.4' -i scale=2
|
| 4 |
+
# push: cog push r8.im/tencentarc/gfpgan
|
| 5 |
+
# push (backup): cog push r8.im/xinntao/gfpgan
|
| 6 |
+
|
| 7 |
+
import os
|
| 8 |
+
|
| 9 |
+
os.system('python setup.py develop')
|
| 10 |
+
os.system('pip install realesrgan')
|
| 11 |
+
|
| 12 |
+
import cv2
|
| 13 |
+
import shutil
|
| 14 |
+
import tempfile
|
| 15 |
+
import torch
|
| 16 |
+
from basicsr.archs.srvgg_arch import SRVGGNetCompact
|
| 17 |
+
|
| 18 |
+
from gfpgan import GFPGANer
|
| 19 |
+
|
| 20 |
+
try:
|
| 21 |
+
from cog import BasePredictor, Input, Path
|
| 22 |
+
from realesrgan.utils import RealESRGANer
|
| 23 |
+
except Exception:
|
| 24 |
+
print('please install cog and realesrgan package')
|
| 25 |
+
|
| 26 |
+
|
| 27 |
+
class Predictor(BasePredictor):
|
| 28 |
+
|
| 29 |
+
def setup(self):
|
| 30 |
+
os.makedirs('output', exist_ok=True)
|
| 31 |
+
# download weights
|
| 32 |
+
if not os.path.exists('gfpgan/weights/realesr-general-x4v3.pth'):
|
| 33 |
+
os.system(
|
| 34 |
+
'wget https://github.com/xinntao/Real-ESRGAN/releases/download/v0.2.5.0/realesr-general-x4v3.pth -P ./gfpgan/weights'
|
| 35 |
+
)
|
| 36 |
+
if not os.path.exists('gfpgan/weights/GFPGANv1.2.pth'):
|
| 37 |
+
os.system(
|
| 38 |
+
'wget https://github.com/TencentARC/GFPGAN/releases/download/v1.3.0/GFPGANv1.2.pth -P ./gfpgan/weights')
|
| 39 |
+
if not os.path.exists('gfpgan/weights/GFPGANv1.3.pth'):
|
| 40 |
+
os.system(
|
| 41 |
+
'wget https://github.com/TencentARC/GFPGAN/releases/download/v1.3.0/GFPGANv1.3.pth -P ./gfpgan/weights')
|
| 42 |
+
if not os.path.exists('gfpgan/weights/GFPGANv1.4.pth'):
|
| 43 |
+
os.system(
|
| 44 |
+
'wget https://github.com/TencentARC/GFPGAN/releases/download/v1.3.0/GFPGANv1.4.pth -P ./gfpgan/weights')
|
| 45 |
+
if not os.path.exists('gfpgan/weights/RestoreFormer.pth'):
|
| 46 |
+
os.system(
|
| 47 |
+
'wget https://github.com/TencentARC/GFPGAN/releases/download/v1.3.4/RestoreFormer.pth -P ./gfpgan/weights'
|
| 48 |
+
)
|
| 49 |
+
|
| 50 |
+
# background enhancer with RealESRGAN
|
| 51 |
+
model = SRVGGNetCompact(num_in_ch=3, num_out_ch=3, num_feat=64, num_conv=32, upscale=4, act_type='prelu')
|
| 52 |
+
model_path = 'gfpgan/weights/realesr-general-x4v3.pth'
|
| 53 |
+
half = True if torch.cuda.is_available() else False
|
| 54 |
+
self.upsampler = RealESRGANer(
|
| 55 |
+
scale=4, model_path=model_path, model=model, tile=0, tile_pad=10, pre_pad=0, half=half)
|
| 56 |
+
|
| 57 |
+
# Use GFPGAN for face enhancement
|
| 58 |
+
self.face_enhancer = GFPGANer(
|
| 59 |
+
model_path='gfpgan/weights/GFPGANv1.4.pth',
|
| 60 |
+
upscale=2,
|
| 61 |
+
arch='clean',
|
| 62 |
+
channel_multiplier=2,
|
| 63 |
+
bg_upsampler=self.upsampler)
|
| 64 |
+
self.current_version = 'v1.4'
|
| 65 |
+
|
| 66 |
+
def predict(
|
| 67 |
+
self,
|
| 68 |
+
img: Path = Input(description='Input'),
|
| 69 |
+
version: str = Input(
|
| 70 |
+
description='GFPGAN version. v1.3: better quality. v1.4: more details and better identity.',
|
| 71 |
+
choices=['v1.2', 'v1.3', 'v1.4', 'RestoreFormer'],
|
| 72 |
+
default='v1.4'),
|
| 73 |
+
scale: float = Input(description='Rescaling factor', default=2),
|
| 74 |
+
) -> Path:
|
| 75 |
+
weight = 0.5
|
| 76 |
+
print(img, version, scale, weight)
|
| 77 |
+
try:
|
| 78 |
+
extension = os.path.splitext(os.path.basename(str(img)))[1]
|
| 79 |
+
img = cv2.imread(str(img), cv2.IMREAD_UNCHANGED)
|
| 80 |
+
if len(img.shape) == 3 and img.shape[2] == 4:
|
| 81 |
+
img_mode = 'RGBA'
|
| 82 |
+
elif len(img.shape) == 2:
|
| 83 |
+
img_mode = None
|
| 84 |
+
img = cv2.cvtColor(img, cv2.COLOR_GRAY2BGR)
|
| 85 |
+
else:
|
| 86 |
+
img_mode = None
|
| 87 |
+
|
| 88 |
+
h, w = img.shape[0:2]
|
| 89 |
+
if h < 300:
|
| 90 |
+
img = cv2.resize(img, (w * 2, h * 2), interpolation=cv2.INTER_LANCZOS4)
|
| 91 |
+
|
| 92 |
+
if self.current_version != version:
|
| 93 |
+
if version == 'v1.2':
|
| 94 |
+
self.face_enhancer = GFPGANer(
|
| 95 |
+
model_path='gfpgan/weights/GFPGANv1.2.pth',
|
| 96 |
+
upscale=2,
|
| 97 |
+
arch='clean',
|
| 98 |
+
channel_multiplier=2,
|
| 99 |
+
bg_upsampler=self.upsampler)
|
| 100 |
+
self.current_version = 'v1.2'
|
| 101 |
+
elif version == 'v1.3':
|
| 102 |
+
self.face_enhancer = GFPGANer(
|
| 103 |
+
model_path='gfpgan/weights/GFPGANv1.3.pth',
|
| 104 |
+
upscale=2,
|
| 105 |
+
arch='clean',
|
| 106 |
+
channel_multiplier=2,
|
| 107 |
+
bg_upsampler=self.upsampler)
|
| 108 |
+
self.current_version = 'v1.3'
|
| 109 |
+
elif version == 'v1.4':
|
| 110 |
+
self.face_enhancer = GFPGANer(
|
| 111 |
+
model_path='gfpgan/weights/GFPGANv1.4.pth',
|
| 112 |
+
upscale=2,
|
| 113 |
+
arch='clean',
|
| 114 |
+
channel_multiplier=2,
|
| 115 |
+
bg_upsampler=self.upsampler)
|
| 116 |
+
self.current_version = 'v1.4'
|
| 117 |
+
elif version == 'RestoreFormer':
|
| 118 |
+
self.face_enhancer = GFPGANer(
|
| 119 |
+
model_path='gfpgan/weights/RestoreFormer.pth',
|
| 120 |
+
upscale=2,
|
| 121 |
+
arch='RestoreFormer',
|
| 122 |
+
channel_multiplier=2,
|
| 123 |
+
bg_upsampler=self.upsampler)
|
| 124 |
+
|
| 125 |
+
try:
|
| 126 |
+
_, _, output = self.face_enhancer.enhance(
|
| 127 |
+
img, has_aligned=False, only_center_face=False, paste_back=True, weight=weight)
|
| 128 |
+
except RuntimeError as error:
|
| 129 |
+
print('Error', error)
|
| 130 |
+
|
| 131 |
+
try:
|
| 132 |
+
if scale != 2:
|
| 133 |
+
interpolation = cv2.INTER_AREA if scale < 2 else cv2.INTER_LANCZOS4
|
| 134 |
+
h, w = img.shape[0:2]
|
| 135 |
+
output = cv2.resize(output, (int(w * scale / 2), int(h * scale / 2)), interpolation=interpolation)
|
| 136 |
+
except Exception as error:
|
| 137 |
+
print('wrong scale input.', error)
|
| 138 |
+
|
| 139 |
+
if img_mode == 'RGBA': # RGBA images should be saved in png format
|
| 140 |
+
extension = 'png'
|
| 141 |
+
# save_path = f'output/out.{extension}'
|
| 142 |
+
# cv2.imwrite(save_path, output)
|
| 143 |
+
out_path = Path(tempfile.mkdtemp()) / f'out.{extension}'
|
| 144 |
+
cv2.imwrite(str(out_path), output)
|
| 145 |
+
except Exception as error:
|
| 146 |
+
print('global exception: ', error)
|
| 147 |
+
finally:
|
| 148 |
+
clean_folder('output')
|
| 149 |
+
return out_path
|
| 150 |
+
|
| 151 |
+
|
| 152 |
+
def clean_folder(folder):
|
| 153 |
+
for filename in os.listdir(folder):
|
| 154 |
+
file_path = os.path.join(folder, filename)
|
| 155 |
+
try:
|
| 156 |
+
if os.path.isfile(file_path) or os.path.islink(file_path):
|
| 157 |
+
os.unlink(file_path)
|
| 158 |
+
elif os.path.isdir(file_path):
|
| 159 |
+
shutil.rmtree(file_path)
|
| 160 |
+
except Exception as e:
|
| 161 |
+
print(f'Failed to delete {file_path}. Reason: {e}')
|
experiments/pretrained_models/README.md
ADDED
|
@@ -0,0 +1,7 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Pre-trained Models and Other Data
|
| 2 |
+
|
| 3 |
+
Download pre-trained models and other data. Put them in this folder.
|
| 4 |
+
|
| 5 |
+
1. [Pretrained StyleGAN2 model: StyleGAN2_512_Cmul1_FFHQ_B12G4_scratch_800k.pth](https://github.com/TencentARC/GFPGAN/releases/download/v0.1.0/StyleGAN2_512_Cmul1_FFHQ_B12G4_scratch_800k.pth)
|
| 6 |
+
1. [Component locations of FFHQ: FFHQ_eye_mouth_landmarks_512.pth](https://github.com/TencentARC/GFPGAN/releases/download/v0.1.0/FFHQ_eye_mouth_landmarks_512.pth)
|
| 7 |
+
1. [A simple ArcFace model: arcface_resnet18.pth](https://github.com/TencentARC/GFPGAN/releases/download/v0.1.0/arcface_resnet18.pth)
|
gfpgan.egg-info/PKG-INFO
ADDED
|
@@ -0,0 +1,220 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
Metadata-Version: 2.1
|
| 2 |
+
Name: gfpgan
|
| 3 |
+
Version: 1.3.8
|
| 4 |
+
Summary: GFPGAN aims at developing Practical Algorithms for Real-world Face Restoration
|
| 5 |
+
Home-page: https://github.com/TencentARC/GFPGAN
|
| 6 |
+
Author: Xintao Wang
|
| 7 |
+
Author-email: xintao.wang@outlook.com
|
| 8 |
+
License: Apache License Version 2.0
|
| 9 |
+
Keywords: computer vision,pytorch,image restoration,super-resolution,face restoration,gan,gfpgan
|
| 10 |
+
Classifier: Development Status :: 4 - Beta
|
| 11 |
+
Classifier: License :: OSI Approved :: Apache Software License
|
| 12 |
+
Classifier: Operating System :: OS Independent
|
| 13 |
+
Classifier: Programming Language :: Python :: 3
|
| 14 |
+
Classifier: Programming Language :: Python :: 3.7
|
| 15 |
+
Classifier: Programming Language :: Python :: 3.8
|
| 16 |
+
Description-Content-Type: text/markdown
|
| 17 |
+
License-File: LICENSE
|
| 18 |
+
|
| 19 |
+
<p align="center">
|
| 20 |
+
<img src="assets/gfpgan_logo.png" height=130>
|
| 21 |
+
</p>
|
| 22 |
+
|
| 23 |
+
## <div align="center"><b><a href="README.md">English</a> | <a href="README_CN.md">简体中文</a></b></div>
|
| 24 |
+
|
| 25 |
+
<div align="center">
|
| 26 |
+
<!-- <a href="https://twitter.com/_Xintao_" style="text-decoration:none;">
|
| 27 |
+
<img src="https://user-images.githubusercontent.com/17445847/187162058-c764ced6-952f-404b-ac85-ba95cce18e7b.png" width="4%" alt="" />
|
| 28 |
+
</a> -->
|
| 29 |
+
|
| 30 |
+
[](https://github.com/TencentARC/GFPGAN/releases)
|
| 31 |
+
[](https://pypi.org/project/gfpgan/)
|
| 32 |
+
[](https://github.com/TencentARC/GFPGAN/issues)
|
| 33 |
+
[](https://github.com/TencentARC/GFPGAN/issues)
|
| 34 |
+
[](https://github.com/TencentARC/GFPGAN/blob/master/LICENSE)
|
| 35 |
+
[](https://github.com/TencentARC/GFPGAN/blob/master/.github/workflows/pylint.yml)
|
| 36 |
+
[](https://github.com/TencentARC/GFPGAN/blob/master/.github/workflows/publish-pip.yml)
|
| 37 |
+
</div>
|
| 38 |
+
|
| 39 |
+
1. :boom: **Updated** online demo: [](https://replicate.com/tencentarc/gfpgan). Here is the [backup](https://replicate.com/xinntao/gfpgan).
|
| 40 |
+
1. :boom: **Updated** online demo: [](https://huggingface.co/spaces/Xintao/GFPGAN)
|
| 41 |
+
1. [Colab Demo](https://colab.research.google.com/drive/1sVsoBd9AjckIXThgtZhGrHRfFI6UUYOo) for GFPGAN <a href="https://colab.research.google.com/drive/1sVsoBd9AjckIXThgtZhGrHRfFI6UUYOo"><img src="https://colab.research.google.com/assets/colab-badge.svg" alt="google colab logo"></a>; (Another [Colab Demo](https://colab.research.google.com/drive/1Oa1WwKB4M4l1GmR7CtswDVgOCOeSLChA?usp=sharing) for the original paper model)
|
| 42 |
+
|
| 43 |
+
<!-- 3. Online demo: [Replicate.ai](https://replicate.com/xinntao/gfpgan) (may need to sign in, return the whole image)
|
| 44 |
+
4. Online demo: [Baseten.co](https://app.baseten.co/applications/Q04Lz0d/operator_views/8qZG6Bg) (backed by GPU, returns the whole image)
|
| 45 |
+
5. We provide a *clean* version of GFPGAN, which can run without CUDA extensions. So that it can run in **Windows** or on **CPU mode**. -->
|
| 46 |
+
|
| 47 |
+
> :rocket: **Thanks for your interest in our work. You may also want to check our new updates on the *tiny models* for *anime images and videos* in [Real-ESRGAN](https://github.com/xinntao/Real-ESRGAN/blob/master/docs/anime_video_model.md)** :blush:
|
| 48 |
+
|
| 49 |
+
GFPGAN aims at developing a **Practical Algorithm for Real-world Face Restoration**.<br>
|
| 50 |
+
It leverages rich and diverse priors encapsulated in a pretrained face GAN (*e.g.*, StyleGAN2) for blind face restoration.
|
| 51 |
+
|
| 52 |
+
:question: Frequently Asked Questions can be found in [FAQ.md](FAQ.md).
|
| 53 |
+
|
| 54 |
+
:triangular_flag_on_post: **Updates**
|
| 55 |
+
|
| 56 |
+
- :white_check_mark: Add [RestoreFormer](https://github.com/wzhouxiff/RestoreFormer) inference codes.
|
| 57 |
+
- :white_check_mark: Add [V1.4 model](https://github.com/TencentARC/GFPGAN/releases/download/v1.3.0/GFPGANv1.4.pth), which produces slightly more details and better identity than V1.3.
|
| 58 |
+
- :white_check_mark: Add **[V1.3 model](https://github.com/TencentARC/GFPGAN/releases/download/v1.3.0/GFPGANv1.3.pth)**, which produces **more natural** restoration results, and better results on *very low-quality* / *high-quality* inputs. See more in [Model zoo](#european_castle-model-zoo), [Comparisons.md](Comparisons.md)
|
| 59 |
+
- :white_check_mark: Integrated to [Huggingface Spaces](https://huggingface.co/spaces) with [Gradio](https://github.com/gradio-app/gradio). See [Gradio Web Demo](https://huggingface.co/spaces/akhaliq/GFPGAN).
|
| 60 |
+
- :white_check_mark: Support enhancing non-face regions (background) with [Real-ESRGAN](https://github.com/xinntao/Real-ESRGAN).
|
| 61 |
+
- :white_check_mark: We provide a *clean* version of GFPGAN, which does not require CUDA extensions.
|
| 62 |
+
- :white_check_mark: We provide an updated model without colorizing faces.
|
| 63 |
+
|
| 64 |
+
---
|
| 65 |
+
|
| 66 |
+
If GFPGAN is helpful in your photos/projects, please help to :star: this repo or recommend it to your friends. Thanks:blush:
|
| 67 |
+
Other recommended projects:<br>
|
| 68 |
+
:arrow_forward: [Real-ESRGAN](https://github.com/xinntao/Real-ESRGAN): A practical algorithm for general image restoration<br>
|
| 69 |
+
:arrow_forward: [BasicSR](https://github.com/xinntao/BasicSR): An open-source image and video restoration toolbox<br>
|
| 70 |
+
:arrow_forward: [facexlib](https://github.com/xinntao/facexlib): A collection that provides useful face-relation functions<br>
|
| 71 |
+
:arrow_forward: [HandyView](https://github.com/xinntao/HandyView): A PyQt5-based image viewer that is handy for view and comparison<br>
|
| 72 |
+
|
| 73 |
+
---
|
| 74 |
+
|
| 75 |
+
### :book: GFP-GAN: Towards Real-World Blind Face Restoration with Generative Facial Prior
|
| 76 |
+
|
| 77 |
+
> [[Paper](https://arxiv.org/abs/2101.04061)]   [[Project Page](https://xinntao.github.io/projects/gfpgan)]   [Demo] <br>
|
| 78 |
+
> [Xintao Wang](https://xinntao.github.io/), [Yu Li](https://yu-li.github.io/), [Honglun Zhang](https://scholar.google.com/citations?hl=en&user=KjQLROoAAAAJ), [Ying Shan](https://scholar.google.com/citations?user=4oXBp9UAAAAJ&hl=en) <br>
|
| 79 |
+
> Applied Research Center (ARC), Tencent PCG
|
| 80 |
+
|
| 81 |
+
<p align="center">
|
| 82 |
+
<img src="https://xinntao.github.io/projects/GFPGAN_src/gfpgan_teaser.jpg">
|
| 83 |
+
</p>
|
| 84 |
+
|
| 85 |
+
---
|
| 86 |
+
|
| 87 |
+
## :wrench: Dependencies and Installation
|
| 88 |
+
|
| 89 |
+
- Python >= 3.7 (Recommend to use [Anaconda](https://www.anaconda.com/download/#linux) or [Miniconda](https://docs.conda.io/en/latest/miniconda.html))
|
| 90 |
+
- [PyTorch >= 1.7](https://pytorch.org/)
|
| 91 |
+
- Option: NVIDIA GPU + [CUDA](https://developer.nvidia.com/cuda-downloads)
|
| 92 |
+
- Option: Linux
|
| 93 |
+
|
| 94 |
+
### Installation
|
| 95 |
+
|
| 96 |
+
We now provide a *clean* version of GFPGAN, which does not require customized CUDA extensions. <br>
|
| 97 |
+
If you want to use the original model in our paper, please see [PaperModel.md](PaperModel.md) for installation.
|
| 98 |
+
|
| 99 |
+
1. Clone repo
|
| 100 |
+
|
| 101 |
+
```bash
|
| 102 |
+
git clone https://github.com/TencentARC/GFPGAN.git
|
| 103 |
+
cd GFPGAN
|
| 104 |
+
```
|
| 105 |
+
|
| 106 |
+
1. Install dependent packages
|
| 107 |
+
|
| 108 |
+
```bash
|
| 109 |
+
# Install basicsr - https://github.com/xinntao/BasicSR
|
| 110 |
+
# We use BasicSR for both training and inference
|
| 111 |
+
pip install basicsr
|
| 112 |
+
|
| 113 |
+
# Install facexlib - https://github.com/xinntao/facexlib
|
| 114 |
+
# We use face detection and face restoration helper in the facexlib package
|
| 115 |
+
pip install facexlib
|
| 116 |
+
|
| 117 |
+
pip install -r requirements.txt
|
| 118 |
+
python setup.py develop
|
| 119 |
+
|
| 120 |
+
# If you want to enhance the background (non-face) regions with Real-ESRGAN,
|
| 121 |
+
# you also need to install the realesrgan package
|
| 122 |
+
pip install realesrgan
|
| 123 |
+
```
|
| 124 |
+
|
| 125 |
+
## :zap: Quick Inference
|
| 126 |
+
|
| 127 |
+
We take the v1.3 version for an example. More models can be found [here](#european_castle-model-zoo).
|
| 128 |
+
|
| 129 |
+
Download pre-trained models: [GFPGANv1.3.pth](https://github.com/TencentARC/GFPGAN/releases/download/v1.3.0/GFPGANv1.3.pth)
|
| 130 |
+
|
| 131 |
+
```bash
|
| 132 |
+
wget https://github.com/TencentARC/GFPGAN/releases/download/v1.3.0/GFPGANv1.3.pth -P experiments/pretrained_models
|
| 133 |
+
```
|
| 134 |
+
|
| 135 |
+
**Inference!**
|
| 136 |
+
|
| 137 |
+
```bash
|
| 138 |
+
python inference_gfpgan.py -i inputs/whole_imgs -o results -v 1.3 -s 2
|
| 139 |
+
```
|
| 140 |
+
|
| 141 |
+
```console
|
| 142 |
+
Usage: python inference_gfpgan.py -i inputs/whole_imgs -o results -v 1.3 -s 2 [options]...
|
| 143 |
+
|
| 144 |
+
-h show this help
|
| 145 |
+
-i input Input image or folder. Default: inputs/whole_imgs
|
| 146 |
+
-o output Output folder. Default: results
|
| 147 |
+
-v version GFPGAN model version. Option: 1 | 1.2 | 1.3. Default: 1.3
|
| 148 |
+
-s upscale The final upsampling scale of the image. Default: 2
|
| 149 |
+
-bg_upsampler background upsampler. Default: realesrgan
|
| 150 |
+
-bg_tile Tile size for background sampler, 0 for no tile during testing. Default: 400
|
| 151 |
+
-suffix Suffix of the restored faces
|
| 152 |
+
-only_center_face Only restore the center face
|
| 153 |
+
-aligned Input are aligned faces
|
| 154 |
+
-ext Image extension. Options: auto | jpg | png, auto means using the same extension as inputs. Default: auto
|
| 155 |
+
```
|
| 156 |
+
|
| 157 |
+
If you want to use the original model in our paper, please see [PaperModel.md](PaperModel.md) for installation and inference.
|
| 158 |
+
|
| 159 |
+
## :european_castle: Model Zoo
|
| 160 |
+
|
| 161 |
+
| Version | Model Name | Description |
|
| 162 |
+
| :---: | :---: | :---: |
|
| 163 |
+
| V1.3 | [GFPGANv1.3.pth](https://github.com/TencentARC/GFPGAN/releases/download/v1.3.0/GFPGANv1.3.pth) | Based on V1.2; **more natural** restoration results; better results on very low-quality / high-quality inputs. |
|
| 164 |
+
| V1.2 | [GFPGANCleanv1-NoCE-C2.pth](https://github.com/TencentARC/GFPGAN/releases/download/v0.2.0/GFPGANCleanv1-NoCE-C2.pth) | No colorization; no CUDA extensions are required. Trained with more data with pre-processing. |
|
| 165 |
+
| V1 | [GFPGANv1.pth](https://github.com/TencentARC/GFPGAN/releases/download/v0.1.0/GFPGANv1.pth) | The paper model, with colorization. |
|
| 166 |
+
|
| 167 |
+
The comparisons are in [Comparisons.md](Comparisons.md).
|
| 168 |
+
|
| 169 |
+
Note that V1.3 is not always better than V1.2. You may need to select different models based on your purpose and inputs.
|
| 170 |
+
|
| 171 |
+
| Version | Strengths | Weaknesses |
|
| 172 |
+
| :---: | :---: | :---: |
|
| 173 |
+
|V1.3 | ✓ natural outputs<br> ✓better results on very low-quality inputs <br> ✓ work on relatively high-quality inputs <br>✓ can have repeated (twice) restorations | ✗ not very sharp <br> ✗ have a slight change on identity |
|
| 174 |
+
|V1.2 | ✓ sharper output <br> ✓ with beauty makeup | ✗ some outputs are unnatural |
|
| 175 |
+
|
| 176 |
+
You can find **more models (such as the discriminators)** here: [[Google Drive](https://drive.google.com/drive/folders/17rLiFzcUMoQuhLnptDsKolegHWwJOnHu?usp=sharing)], OR [[Tencent Cloud 腾讯微云](https://share.weiyun.com/ShYoCCoc)]
|
| 177 |
+
|
| 178 |
+
## :computer: Training
|
| 179 |
+
|
| 180 |
+
We provide the training codes for GFPGAN (used in our paper). <br>
|
| 181 |
+
You could improve it according to your own needs.
|
| 182 |
+
|
| 183 |
+
**Tips**
|
| 184 |
+
|
| 185 |
+
1. More high quality faces can improve the restoration quality.
|
| 186 |
+
2. You may need to perform some pre-processing, such as beauty makeup.
|
| 187 |
+
|
| 188 |
+
**Procedures**
|
| 189 |
+
|
| 190 |
+
(You can try a simple version ( `options/train_gfpgan_v1_simple.yml`) that does not require face component landmarks.)
|
| 191 |
+
|
| 192 |
+
1. Dataset preparation: [FFHQ](https://github.com/NVlabs/ffhq-dataset)
|
| 193 |
+
|
| 194 |
+
1. Download pre-trained models and other data. Put them in the `experiments/pretrained_models` folder.
|
| 195 |
+
1. [Pre-trained StyleGAN2 model: StyleGAN2_512_Cmul1_FFHQ_B12G4_scratch_800k.pth](https://github.com/TencentARC/GFPGAN/releases/download/v0.1.0/StyleGAN2_512_Cmul1_FFHQ_B12G4_scratch_800k.pth)
|
| 196 |
+
1. [Component locations of FFHQ: FFHQ_eye_mouth_landmarks_512.pth](https://github.com/TencentARC/GFPGAN/releases/download/v0.1.0/FFHQ_eye_mouth_landmarks_512.pth)
|
| 197 |
+
1. [A simple ArcFace model: arcface_resnet18.pth](https://github.com/TencentARC/GFPGAN/releases/download/v0.1.0/arcface_resnet18.pth)
|
| 198 |
+
|
| 199 |
+
1. Modify the configuration file `options/train_gfpgan_v1.yml` accordingly.
|
| 200 |
+
|
| 201 |
+
1. Training
|
| 202 |
+
|
| 203 |
+
> python -m torch.distributed.launch --nproc_per_node=4 --master_port=22021 gfpgan/train.py -opt options/train_gfpgan_v1.yml --launcher pytorch
|
| 204 |
+
|
| 205 |
+
## :scroll: License and Acknowledgement
|
| 206 |
+
|
| 207 |
+
GFPGAN is released under Apache License Version 2.0.
|
| 208 |
+
|
| 209 |
+
## BibTeX
|
| 210 |
+
|
| 211 |
+
@InProceedings{wang2021gfpgan,
|
| 212 |
+
author = {Xintao Wang and Yu Li and Honglun Zhang and Ying Shan},
|
| 213 |
+
title = {Towards Real-World Blind Face Restoration with Generative Facial Prior},
|
| 214 |
+
booktitle={The IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
|
| 215 |
+
year = {2021}
|
| 216 |
+
}
|
| 217 |
+
|
| 218 |
+
## :e-mail: Contact
|
| 219 |
+
|
| 220 |
+
If you have any question, please email `xintao.wang@outlook.com` or `xintaowang@tencent.com`.
|
gfpgan.egg-info/SOURCES.txt
ADDED
|
@@ -0,0 +1,34 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
LICENSE
|
| 2 |
+
MANIFEST.in
|
| 3 |
+
README.md
|
| 4 |
+
VERSION
|
| 5 |
+
inference_gfpgan.py
|
| 6 |
+
requirements.txt
|
| 7 |
+
setup.cfg
|
| 8 |
+
setup.py
|
| 9 |
+
assets/gfpgan_logo.png
|
| 10 |
+
gfpgan/__init__.py
|
| 11 |
+
gfpgan/train.py
|
| 12 |
+
gfpgan/utils.py
|
| 13 |
+
gfpgan/version.py
|
| 14 |
+
gfpgan.egg-info/PKG-INFO
|
| 15 |
+
gfpgan.egg-info/SOURCES.txt
|
| 16 |
+
gfpgan.egg-info/dependency_links.txt
|
| 17 |
+
gfpgan.egg-info/not-zip-safe
|
| 18 |
+
gfpgan.egg-info/requires.txt
|
| 19 |
+
gfpgan.egg-info/top_level.txt
|
| 20 |
+
gfpgan/archs/__init__.py
|
| 21 |
+
gfpgan/archs/arcface_arch.py
|
| 22 |
+
gfpgan/archs/gfpgan_bilinear_arch.py
|
| 23 |
+
gfpgan/archs/gfpganv1_arch.py
|
| 24 |
+
gfpgan/archs/gfpganv1_clean_arch.py
|
| 25 |
+
gfpgan/archs/restoreformer_arch.py
|
| 26 |
+
gfpgan/archs/stylegan2_bilinear_arch.py
|
| 27 |
+
gfpgan/archs/stylegan2_clean_arch.py
|
| 28 |
+
gfpgan/data/__init__.py
|
| 29 |
+
gfpgan/data/ffhq_degradation_dataset.py
|
| 30 |
+
gfpgan/models/__init__.py
|
| 31 |
+
gfpgan/models/gfpgan_model.py
|
| 32 |
+
gfpgan/weights/README.md
|
| 33 |
+
scripts/convert_gfpganv_to_clean.py
|
| 34 |
+
scripts/parse_landmark.py
|
gfpgan.egg-info/dependency_links.txt
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
|
gfpgan.egg-info/not-zip-safe
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
|
gfpgan.egg-info/requires.txt
ADDED
|
@@ -0,0 +1,13 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
basicsr
|
| 2 |
+
facexlib
|
| 3 |
+
realesrgan
|
| 4 |
+
torch==1.7.1
|
| 5 |
+
torchvision==0.8.2
|
| 6 |
+
numpy==1.21.1
|
| 7 |
+
lmdb==1.2.1
|
| 8 |
+
opencv-python==4.5.3.56
|
| 9 |
+
PyYAML==5.4.1
|
| 10 |
+
tqdm==4.62.2
|
| 11 |
+
yapf==0.31.0
|
| 12 |
+
basicsr==1.4.2
|
| 13 |
+
facexlib==0.2.5
|
gfpgan.egg-info/top_level.txt
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
gfpgan
|
gfpgan/__init__.py
ADDED
|
@@ -0,0 +1,7 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# flake8: noqa
|
| 2 |
+
from .archs import *
|
| 3 |
+
from .data import *
|
| 4 |
+
from .models import *
|
| 5 |
+
from .utils import *
|
| 6 |
+
|
| 7 |
+
# from .version import *
|
gfpgan/__pycache__/__init__.cpython-311.pyc
ADDED
|
Binary file (284 Bytes). View file
|
|
|
gfpgan/__pycache__/utils.cpython-311.pyc
ADDED
|
Binary file (7.05 kB). View file
|
|
|
gfpgan/archs/__init__.py
ADDED
|
@@ -0,0 +1,10 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import importlib
|
| 2 |
+
from basicsr.utils import scandir
|
| 3 |
+
from os import path as osp
|
| 4 |
+
|
| 5 |
+
# automatically scan and import arch modules for registry
|
| 6 |
+
# scan all the files that end with '_arch.py' under the archs folder
|
| 7 |
+
arch_folder = osp.dirname(osp.abspath(__file__))
|
| 8 |
+
arch_filenames = [osp.splitext(osp.basename(v))[0] for v in scandir(arch_folder) if v.endswith('_arch.py')]
|
| 9 |
+
# import all the arch modules
|
| 10 |
+
_arch_modules = [importlib.import_module(f'gfpgan.archs.{file_name}') for file_name in arch_filenames]
|
gfpgan/archs/__pycache__/__init__.cpython-311.pyc
ADDED
|
Binary file (1.16 kB). View file
|
|
|
gfpgan/archs/__pycache__/arcface_arch.cpython-311.pyc
ADDED
|
Binary file (14.6 kB). View file
|
|
|
gfpgan/archs/__pycache__/gfpgan_bilinear_arch.cpython-311.pyc
ADDED
|
Binary file (16 kB). View file
|
|
|
gfpgan/archs/__pycache__/gfpganv1_arch.cpython-311.pyc
ADDED
|
Binary file (22.6 kB). View file
|
|
|
gfpgan/archs/__pycache__/gfpganv1_clean_arch.cpython-311.pyc
ADDED
|
Binary file (17.3 kB). View file
|
|
|
gfpgan/archs/__pycache__/restoreformer_arch.cpython-311.pyc
ADDED
|
Binary file (30.7 kB). View file
|
|
|
gfpgan/archs/__pycache__/stylegan2_bilinear_arch.cpython-311.pyc
ADDED
|
Binary file (30.5 kB). View file
|
|
|
gfpgan/archs/__pycache__/stylegan2_clean_arch.cpython-311.pyc
ADDED
|
Binary file (20.3 kB). View file
|
|
|
gfpgan/archs/arcface_arch.py
ADDED
|
@@ -0,0 +1,245 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import torch.nn as nn
|
| 2 |
+
from basicsr.utils.registry import ARCH_REGISTRY
|
| 3 |
+
|
| 4 |
+
|
| 5 |
+
def conv3x3(inplanes, outplanes, stride=1):
|
| 6 |
+
"""A simple wrapper for 3x3 convolution with padding.
|
| 7 |
+
|
| 8 |
+
Args:
|
| 9 |
+
inplanes (int): Channel number of inputs.
|
| 10 |
+
outplanes (int): Channel number of outputs.
|
| 11 |
+
stride (int): Stride in convolution. Default: 1.
|
| 12 |
+
"""
|
| 13 |
+
return nn.Conv2d(inplanes, outplanes, kernel_size=3, stride=stride, padding=1, bias=False)
|
| 14 |
+
|
| 15 |
+
|
| 16 |
+
class BasicBlock(nn.Module):
|
| 17 |
+
"""Basic residual block used in the ResNetArcFace architecture.
|
| 18 |
+
|
| 19 |
+
Args:
|
| 20 |
+
inplanes (int): Channel number of inputs.
|
| 21 |
+
planes (int): Channel number of outputs.
|
| 22 |
+
stride (int): Stride in convolution. Default: 1.
|
| 23 |
+
downsample (nn.Module): The downsample module. Default: None.
|
| 24 |
+
"""
|
| 25 |
+
expansion = 1 # output channel expansion ratio
|
| 26 |
+
|
| 27 |
+
def __init__(self, inplanes, planes, stride=1, downsample=None):
|
| 28 |
+
super(BasicBlock, self).__init__()
|
| 29 |
+
self.conv1 = conv3x3(inplanes, planes, stride)
|
| 30 |
+
self.bn1 = nn.BatchNorm2d(planes)
|
| 31 |
+
self.relu = nn.ReLU(inplace=True)
|
| 32 |
+
self.conv2 = conv3x3(planes, planes)
|
| 33 |
+
self.bn2 = nn.BatchNorm2d(planes)
|
| 34 |
+
self.downsample = downsample
|
| 35 |
+
self.stride = stride
|
| 36 |
+
|
| 37 |
+
def forward(self, x):
|
| 38 |
+
residual = x
|
| 39 |
+
|
| 40 |
+
out = self.conv1(x)
|
| 41 |
+
out = self.bn1(out)
|
| 42 |
+
out = self.relu(out)
|
| 43 |
+
|
| 44 |
+
out = self.conv2(out)
|
| 45 |
+
out = self.bn2(out)
|
| 46 |
+
|
| 47 |
+
if self.downsample is not None:
|
| 48 |
+
residual = self.downsample(x)
|
| 49 |
+
|
| 50 |
+
out += residual
|
| 51 |
+
out = self.relu(out)
|
| 52 |
+
|
| 53 |
+
return out
|
| 54 |
+
|
| 55 |
+
|
| 56 |
+
class IRBlock(nn.Module):
|
| 57 |
+
"""Improved residual block (IR Block) used in the ResNetArcFace architecture.
|
| 58 |
+
|
| 59 |
+
Args:
|
| 60 |
+
inplanes (int): Channel number of inputs.
|
| 61 |
+
planes (int): Channel number of outputs.
|
| 62 |
+
stride (int): Stride in convolution. Default: 1.
|
| 63 |
+
downsample (nn.Module): The downsample module. Default: None.
|
| 64 |
+
use_se (bool): Whether use the SEBlock (squeeze and excitation block). Default: True.
|
| 65 |
+
"""
|
| 66 |
+
expansion = 1 # output channel expansion ratio
|
| 67 |
+
|
| 68 |
+
def __init__(self, inplanes, planes, stride=1, downsample=None, use_se=True):
|
| 69 |
+
super(IRBlock, self).__init__()
|
| 70 |
+
self.bn0 = nn.BatchNorm2d(inplanes)
|
| 71 |
+
self.conv1 = conv3x3(inplanes, inplanes)
|
| 72 |
+
self.bn1 = nn.BatchNorm2d(inplanes)
|
| 73 |
+
self.prelu = nn.PReLU()
|
| 74 |
+
self.conv2 = conv3x3(inplanes, planes, stride)
|
| 75 |
+
self.bn2 = nn.BatchNorm2d(planes)
|
| 76 |
+
self.downsample = downsample
|
| 77 |
+
self.stride = stride
|
| 78 |
+
self.use_se = use_se
|
| 79 |
+
if self.use_se:
|
| 80 |
+
self.se = SEBlock(planes)
|
| 81 |
+
|
| 82 |
+
def forward(self, x):
|
| 83 |
+
residual = x
|
| 84 |
+
out = self.bn0(x)
|
| 85 |
+
out = self.conv1(out)
|
| 86 |
+
out = self.bn1(out)
|
| 87 |
+
out = self.prelu(out)
|
| 88 |
+
|
| 89 |
+
out = self.conv2(out)
|
| 90 |
+
out = self.bn2(out)
|
| 91 |
+
if self.use_se:
|
| 92 |
+
out = self.se(out)
|
| 93 |
+
|
| 94 |
+
if self.downsample is not None:
|
| 95 |
+
residual = self.downsample(x)
|
| 96 |
+
|
| 97 |
+
out += residual
|
| 98 |
+
out = self.prelu(out)
|
| 99 |
+
|
| 100 |
+
return out
|
| 101 |
+
|
| 102 |
+
|
| 103 |
+
class Bottleneck(nn.Module):
|
| 104 |
+
"""Bottleneck block used in the ResNetArcFace architecture.
|
| 105 |
+
|
| 106 |
+
Args:
|
| 107 |
+
inplanes (int): Channel number of inputs.
|
| 108 |
+
planes (int): Channel number of outputs.
|
| 109 |
+
stride (int): Stride in convolution. Default: 1.
|
| 110 |
+
downsample (nn.Module): The downsample module. Default: None.
|
| 111 |
+
"""
|
| 112 |
+
expansion = 4 # output channel expansion ratio
|
| 113 |
+
|
| 114 |
+
def __init__(self, inplanes, planes, stride=1, downsample=None):
|
| 115 |
+
super(Bottleneck, self).__init__()
|
| 116 |
+
self.conv1 = nn.Conv2d(inplanes, planes, kernel_size=1, bias=False)
|
| 117 |
+
self.bn1 = nn.BatchNorm2d(planes)
|
| 118 |
+
self.conv2 = nn.Conv2d(planes, planes, kernel_size=3, stride=stride, padding=1, bias=False)
|
| 119 |
+
self.bn2 = nn.BatchNorm2d(planes)
|
| 120 |
+
self.conv3 = nn.Conv2d(planes, planes * self.expansion, kernel_size=1, bias=False)
|
| 121 |
+
self.bn3 = nn.BatchNorm2d(planes * self.expansion)
|
| 122 |
+
self.relu = nn.ReLU(inplace=True)
|
| 123 |
+
self.downsample = downsample
|
| 124 |
+
self.stride = stride
|
| 125 |
+
|
| 126 |
+
def forward(self, x):
|
| 127 |
+
residual = x
|
| 128 |
+
|
| 129 |
+
out = self.conv1(x)
|
| 130 |
+
out = self.bn1(out)
|
| 131 |
+
out = self.relu(out)
|
| 132 |
+
|
| 133 |
+
out = self.conv2(out)
|
| 134 |
+
out = self.bn2(out)
|
| 135 |
+
out = self.relu(out)
|
| 136 |
+
|
| 137 |
+
out = self.conv3(out)
|
| 138 |
+
out = self.bn3(out)
|
| 139 |
+
|
| 140 |
+
if self.downsample is not None:
|
| 141 |
+
residual = self.downsample(x)
|
| 142 |
+
|
| 143 |
+
out += residual
|
| 144 |
+
out = self.relu(out)
|
| 145 |
+
|
| 146 |
+
return out
|
| 147 |
+
|
| 148 |
+
|
| 149 |
+
class SEBlock(nn.Module):
|
| 150 |
+
"""The squeeze-and-excitation block (SEBlock) used in the IRBlock.
|
| 151 |
+
|
| 152 |
+
Args:
|
| 153 |
+
channel (int): Channel number of inputs.
|
| 154 |
+
reduction (int): Channel reduction ration. Default: 16.
|
| 155 |
+
"""
|
| 156 |
+
|
| 157 |
+
def __init__(self, channel, reduction=16):
|
| 158 |
+
super(SEBlock, self).__init__()
|
| 159 |
+
self.avg_pool = nn.AdaptiveAvgPool2d(1) # pool to 1x1 without spatial information
|
| 160 |
+
self.fc = nn.Sequential(
|
| 161 |
+
nn.Linear(channel, channel // reduction), nn.PReLU(), nn.Linear(channel // reduction, channel),
|
| 162 |
+
nn.Sigmoid())
|
| 163 |
+
|
| 164 |
+
def forward(self, x):
|
| 165 |
+
b, c, _, _ = x.size()
|
| 166 |
+
y = self.avg_pool(x).view(b, c)
|
| 167 |
+
y = self.fc(y).view(b, c, 1, 1)
|
| 168 |
+
return x * y
|
| 169 |
+
|
| 170 |
+
|
| 171 |
+
@ARCH_REGISTRY.register()
|
| 172 |
+
class ResNetArcFace(nn.Module):
|
| 173 |
+
"""ArcFace with ResNet architectures.
|
| 174 |
+
|
| 175 |
+
Ref: ArcFace: Additive Angular Margin Loss for Deep Face Recognition.
|
| 176 |
+
|
| 177 |
+
Args:
|
| 178 |
+
block (str): Block used in the ArcFace architecture.
|
| 179 |
+
layers (tuple(int)): Block numbers in each layer.
|
| 180 |
+
use_se (bool): Whether use the SEBlock (squeeze and excitation block). Default: True.
|
| 181 |
+
"""
|
| 182 |
+
|
| 183 |
+
def __init__(self, block, layers, use_se=True):
|
| 184 |
+
if block == 'IRBlock':
|
| 185 |
+
block = IRBlock
|
| 186 |
+
self.inplanes = 64
|
| 187 |
+
self.use_se = use_se
|
| 188 |
+
super(ResNetArcFace, self).__init__()
|
| 189 |
+
|
| 190 |
+
self.conv1 = nn.Conv2d(1, 64, kernel_size=3, padding=1, bias=False)
|
| 191 |
+
self.bn1 = nn.BatchNorm2d(64)
|
| 192 |
+
self.prelu = nn.PReLU()
|
| 193 |
+
self.maxpool = nn.MaxPool2d(kernel_size=2, stride=2)
|
| 194 |
+
self.layer1 = self._make_layer(block, 64, layers[0])
|
| 195 |
+
self.layer2 = self._make_layer(block, 128, layers[1], stride=2)
|
| 196 |
+
self.layer3 = self._make_layer(block, 256, layers[2], stride=2)
|
| 197 |
+
self.layer4 = self._make_layer(block, 512, layers[3], stride=2)
|
| 198 |
+
self.bn4 = nn.BatchNorm2d(512)
|
| 199 |
+
self.dropout = nn.Dropout()
|
| 200 |
+
self.fc5 = nn.Linear(512 * 8 * 8, 512)
|
| 201 |
+
self.bn5 = nn.BatchNorm1d(512)
|
| 202 |
+
|
| 203 |
+
# initialization
|
| 204 |
+
for m in self.modules():
|
| 205 |
+
if isinstance(m, nn.Conv2d):
|
| 206 |
+
nn.init.xavier_normal_(m.weight)
|
| 207 |
+
elif isinstance(m, nn.BatchNorm2d) or isinstance(m, nn.BatchNorm1d):
|
| 208 |
+
nn.init.constant_(m.weight, 1)
|
| 209 |
+
nn.init.constant_(m.bias, 0)
|
| 210 |
+
elif isinstance(m, nn.Linear):
|
| 211 |
+
nn.init.xavier_normal_(m.weight)
|
| 212 |
+
nn.init.constant_(m.bias, 0)
|
| 213 |
+
|
| 214 |
+
def _make_layer(self, block, planes, num_blocks, stride=1):
|
| 215 |
+
downsample = None
|
| 216 |
+
if stride != 1 or self.inplanes != planes * block.expansion:
|
| 217 |
+
downsample = nn.Sequential(
|
| 218 |
+
nn.Conv2d(self.inplanes, planes * block.expansion, kernel_size=1, stride=stride, bias=False),
|
| 219 |
+
nn.BatchNorm2d(planes * block.expansion),
|
| 220 |
+
)
|
| 221 |
+
layers = []
|
| 222 |
+
layers.append(block(self.inplanes, planes, stride, downsample, use_se=self.use_se))
|
| 223 |
+
self.inplanes = planes
|
| 224 |
+
for _ in range(1, num_blocks):
|
| 225 |
+
layers.append(block(self.inplanes, planes, use_se=self.use_se))
|
| 226 |
+
|
| 227 |
+
return nn.Sequential(*layers)
|
| 228 |
+
|
| 229 |
+
def forward(self, x):
|
| 230 |
+
x = self.conv1(x)
|
| 231 |
+
x = self.bn1(x)
|
| 232 |
+
x = self.prelu(x)
|
| 233 |
+
x = self.maxpool(x)
|
| 234 |
+
|
| 235 |
+
x = self.layer1(x)
|
| 236 |
+
x = self.layer2(x)
|
| 237 |
+
x = self.layer3(x)
|
| 238 |
+
x = self.layer4(x)
|
| 239 |
+
x = self.bn4(x)
|
| 240 |
+
x = self.dropout(x)
|
| 241 |
+
x = x.view(x.size(0), -1)
|
| 242 |
+
x = self.fc5(x)
|
| 243 |
+
x = self.bn5(x)
|
| 244 |
+
|
| 245 |
+
return x
|
gfpgan/archs/gfpgan_bilinear_arch.py
ADDED
|
@@ -0,0 +1,312 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import math
|
| 2 |
+
import random
|
| 3 |
+
import torch
|
| 4 |
+
from basicsr.utils.registry import ARCH_REGISTRY
|
| 5 |
+
from torch import nn
|
| 6 |
+
|
| 7 |
+
from .gfpganv1_arch import ResUpBlock
|
| 8 |
+
from .stylegan2_bilinear_arch import (ConvLayer, EqualConv2d, EqualLinear, ResBlock, ScaledLeakyReLU,
|
| 9 |
+
StyleGAN2GeneratorBilinear)
|
| 10 |
+
|
| 11 |
+
|
| 12 |
+
class StyleGAN2GeneratorBilinearSFT(StyleGAN2GeneratorBilinear):
|
| 13 |
+
"""StyleGAN2 Generator with SFT modulation (Spatial Feature Transform).
|
| 14 |
+
|
| 15 |
+
It is the bilinear version. It does not use the complicated UpFirDnSmooth function that is not friendly for
|
| 16 |
+
deployment. It can be easily converted to the clean version: StyleGAN2GeneratorCSFT.
|
| 17 |
+
|
| 18 |
+
Args:
|
| 19 |
+
out_size (int): The spatial size of outputs.
|
| 20 |
+
num_style_feat (int): Channel number of style features. Default: 512.
|
| 21 |
+
num_mlp (int): Layer number of MLP style layers. Default: 8.
|
| 22 |
+
channel_multiplier (int): Channel multiplier for large networks of StyleGAN2. Default: 2.
|
| 23 |
+
lr_mlp (float): Learning rate multiplier for mlp layers. Default: 0.01.
|
| 24 |
+
narrow (float): The narrow ratio for channels. Default: 1.
|
| 25 |
+
sft_half (bool): Whether to apply SFT on half of the input channels. Default: False.
|
| 26 |
+
"""
|
| 27 |
+
|
| 28 |
+
def __init__(self,
|
| 29 |
+
out_size,
|
| 30 |
+
num_style_feat=512,
|
| 31 |
+
num_mlp=8,
|
| 32 |
+
channel_multiplier=2,
|
| 33 |
+
lr_mlp=0.01,
|
| 34 |
+
narrow=1,
|
| 35 |
+
sft_half=False):
|
| 36 |
+
super(StyleGAN2GeneratorBilinearSFT, self).__init__(
|
| 37 |
+
out_size,
|
| 38 |
+
num_style_feat=num_style_feat,
|
| 39 |
+
num_mlp=num_mlp,
|
| 40 |
+
channel_multiplier=channel_multiplier,
|
| 41 |
+
lr_mlp=lr_mlp,
|
| 42 |
+
narrow=narrow)
|
| 43 |
+
self.sft_half = sft_half
|
| 44 |
+
|
| 45 |
+
def forward(self,
|
| 46 |
+
styles,
|
| 47 |
+
conditions,
|
| 48 |
+
input_is_latent=False,
|
| 49 |
+
noise=None,
|
| 50 |
+
randomize_noise=True,
|
| 51 |
+
truncation=1,
|
| 52 |
+
truncation_latent=None,
|
| 53 |
+
inject_index=None,
|
| 54 |
+
return_latents=False):
|
| 55 |
+
"""Forward function for StyleGAN2GeneratorBilinearSFT.
|
| 56 |
+
|
| 57 |
+
Args:
|
| 58 |
+
styles (list[Tensor]): Sample codes of styles.
|
| 59 |
+
conditions (list[Tensor]): SFT conditions to generators.
|
| 60 |
+
input_is_latent (bool): Whether input is latent style. Default: False.
|
| 61 |
+
noise (Tensor | None): Input noise or None. Default: None.
|
| 62 |
+
randomize_noise (bool): Randomize noise, used when 'noise' is False. Default: True.
|
| 63 |
+
truncation (float): The truncation ratio. Default: 1.
|
| 64 |
+
truncation_latent (Tensor | None): The truncation latent tensor. Default: None.
|
| 65 |
+
inject_index (int | None): The injection index for mixing noise. Default: None.
|
| 66 |
+
return_latents (bool): Whether to return style latents. Default: False.
|
| 67 |
+
"""
|
| 68 |
+
# style codes -> latents with Style MLP layer
|
| 69 |
+
if not input_is_latent:
|
| 70 |
+
styles = [self.style_mlp(s) for s in styles]
|
| 71 |
+
# noises
|
| 72 |
+
if noise is None:
|
| 73 |
+
if randomize_noise:
|
| 74 |
+
noise = [None] * self.num_layers # for each style conv layer
|
| 75 |
+
else: # use the stored noise
|
| 76 |
+
noise = [getattr(self.noises, f'noise{i}') for i in range(self.num_layers)]
|
| 77 |
+
# style truncation
|
| 78 |
+
if truncation < 1:
|
| 79 |
+
style_truncation = []
|
| 80 |
+
for style in styles:
|
| 81 |
+
style_truncation.append(truncation_latent + truncation * (style - truncation_latent))
|
| 82 |
+
styles = style_truncation
|
| 83 |
+
# get style latents with injection
|
| 84 |
+
if len(styles) == 1:
|
| 85 |
+
inject_index = self.num_latent
|
| 86 |
+
|
| 87 |
+
if styles[0].ndim < 3:
|
| 88 |
+
# repeat latent code for all the layers
|
| 89 |
+
latent = styles[0].unsqueeze(1).repeat(1, inject_index, 1)
|
| 90 |
+
else: # used for encoder with different latent code for each layer
|
| 91 |
+
latent = styles[0]
|
| 92 |
+
elif len(styles) == 2: # mixing noises
|
| 93 |
+
if inject_index is None:
|
| 94 |
+
inject_index = random.randint(1, self.num_latent - 1)
|
| 95 |
+
latent1 = styles[0].unsqueeze(1).repeat(1, inject_index, 1)
|
| 96 |
+
latent2 = styles[1].unsqueeze(1).repeat(1, self.num_latent - inject_index, 1)
|
| 97 |
+
latent = torch.cat([latent1, latent2], 1)
|
| 98 |
+
|
| 99 |
+
# main generation
|
| 100 |
+
out = self.constant_input(latent.shape[0])
|
| 101 |
+
out = self.style_conv1(out, latent[:, 0], noise=noise[0])
|
| 102 |
+
skip = self.to_rgb1(out, latent[:, 1])
|
| 103 |
+
|
| 104 |
+
i = 1
|
| 105 |
+
for conv1, conv2, noise1, noise2, to_rgb in zip(self.style_convs[::2], self.style_convs[1::2], noise[1::2],
|
| 106 |
+
noise[2::2], self.to_rgbs):
|
| 107 |
+
out = conv1(out, latent[:, i], noise=noise1)
|
| 108 |
+
|
| 109 |
+
# the conditions may have fewer levels
|
| 110 |
+
if i < len(conditions):
|
| 111 |
+
# SFT part to combine the conditions
|
| 112 |
+
if self.sft_half: # only apply SFT to half of the channels
|
| 113 |
+
out_same, out_sft = torch.split(out, int(out.size(1) // 2), dim=1)
|
| 114 |
+
out_sft = out_sft * conditions[i - 1] + conditions[i]
|
| 115 |
+
out = torch.cat([out_same, out_sft], dim=1)
|
| 116 |
+
else: # apply SFT to all the channels
|
| 117 |
+
out = out * conditions[i - 1] + conditions[i]
|
| 118 |
+
|
| 119 |
+
out = conv2(out, latent[:, i + 1], noise=noise2)
|
| 120 |
+
skip = to_rgb(out, latent[:, i + 2], skip) # feature back to the rgb space
|
| 121 |
+
i += 2
|
| 122 |
+
|
| 123 |
+
image = skip
|
| 124 |
+
|
| 125 |
+
if return_latents:
|
| 126 |
+
return image, latent
|
| 127 |
+
else:
|
| 128 |
+
return image, None
|
| 129 |
+
|
| 130 |
+
|
| 131 |
+
@ARCH_REGISTRY.register()
|
| 132 |
+
class GFPGANBilinear(nn.Module):
|
| 133 |
+
"""The GFPGAN architecture: Unet + StyleGAN2 decoder with SFT.
|
| 134 |
+
|
| 135 |
+
It is the bilinear version and it does not use the complicated UpFirDnSmooth function that is not friendly for
|
| 136 |
+
deployment. It can be easily converted to the clean version: GFPGANv1Clean.
|
| 137 |
+
|
| 138 |
+
|
| 139 |
+
Ref: GFP-GAN: Towards Real-World Blind Face Restoration with Generative Facial Prior.
|
| 140 |
+
|
| 141 |
+
Args:
|
| 142 |
+
out_size (int): The spatial size of outputs.
|
| 143 |
+
num_style_feat (int): Channel number of style features. Default: 512.
|
| 144 |
+
channel_multiplier (int): Channel multiplier for large networks of StyleGAN2. Default: 2.
|
| 145 |
+
decoder_load_path (str): The path to the pre-trained decoder model (usually, the StyleGAN2). Default: None.
|
| 146 |
+
fix_decoder (bool): Whether to fix the decoder. Default: True.
|
| 147 |
+
|
| 148 |
+
num_mlp (int): Layer number of MLP style layers. Default: 8.
|
| 149 |
+
lr_mlp (float): Learning rate multiplier for mlp layers. Default: 0.01.
|
| 150 |
+
input_is_latent (bool): Whether input is latent style. Default: False.
|
| 151 |
+
different_w (bool): Whether to use different latent w for different layers. Default: False.
|
| 152 |
+
narrow (float): The narrow ratio for channels. Default: 1.
|
| 153 |
+
sft_half (bool): Whether to apply SFT on half of the input channels. Default: False.
|
| 154 |
+
"""
|
| 155 |
+
|
| 156 |
+
def __init__(
|
| 157 |
+
self,
|
| 158 |
+
out_size,
|
| 159 |
+
num_style_feat=512,
|
| 160 |
+
channel_multiplier=1,
|
| 161 |
+
decoder_load_path=None,
|
| 162 |
+
fix_decoder=True,
|
| 163 |
+
# for stylegan decoder
|
| 164 |
+
num_mlp=8,
|
| 165 |
+
lr_mlp=0.01,
|
| 166 |
+
input_is_latent=False,
|
| 167 |
+
different_w=False,
|
| 168 |
+
narrow=1,
|
| 169 |
+
sft_half=False):
|
| 170 |
+
|
| 171 |
+
super(GFPGANBilinear, self).__init__()
|
| 172 |
+
self.input_is_latent = input_is_latent
|
| 173 |
+
self.different_w = different_w
|
| 174 |
+
self.num_style_feat = num_style_feat
|
| 175 |
+
|
| 176 |
+
unet_narrow = narrow * 0.5 # by default, use a half of input channels
|
| 177 |
+
channels = {
|
| 178 |
+
'4': int(512 * unet_narrow),
|
| 179 |
+
'8': int(512 * unet_narrow),
|
| 180 |
+
'16': int(512 * unet_narrow),
|
| 181 |
+
'32': int(512 * unet_narrow),
|
| 182 |
+
'64': int(256 * channel_multiplier * unet_narrow),
|
| 183 |
+
'128': int(128 * channel_multiplier * unet_narrow),
|
| 184 |
+
'256': int(64 * channel_multiplier * unet_narrow),
|
| 185 |
+
'512': int(32 * channel_multiplier * unet_narrow),
|
| 186 |
+
'1024': int(16 * channel_multiplier * unet_narrow)
|
| 187 |
+
}
|
| 188 |
+
|
| 189 |
+
self.log_size = int(math.log(out_size, 2))
|
| 190 |
+
first_out_size = 2**(int(math.log(out_size, 2)))
|
| 191 |
+
|
| 192 |
+
self.conv_body_first = ConvLayer(3, channels[f'{first_out_size}'], 1, bias=True, activate=True)
|
| 193 |
+
|
| 194 |
+
# downsample
|
| 195 |
+
in_channels = channels[f'{first_out_size}']
|
| 196 |
+
self.conv_body_down = nn.ModuleList()
|
| 197 |
+
for i in range(self.log_size, 2, -1):
|
| 198 |
+
out_channels = channels[f'{2**(i - 1)}']
|
| 199 |
+
self.conv_body_down.append(ResBlock(in_channels, out_channels))
|
| 200 |
+
in_channels = out_channels
|
| 201 |
+
|
| 202 |
+
self.final_conv = ConvLayer(in_channels, channels['4'], 3, bias=True, activate=True)
|
| 203 |
+
|
| 204 |
+
# upsample
|
| 205 |
+
in_channels = channels['4']
|
| 206 |
+
self.conv_body_up = nn.ModuleList()
|
| 207 |
+
for i in range(3, self.log_size + 1):
|
| 208 |
+
out_channels = channels[f'{2**i}']
|
| 209 |
+
self.conv_body_up.append(ResUpBlock(in_channels, out_channels))
|
| 210 |
+
in_channels = out_channels
|
| 211 |
+
|
| 212 |
+
# to RGB
|
| 213 |
+
self.toRGB = nn.ModuleList()
|
| 214 |
+
for i in range(3, self.log_size + 1):
|
| 215 |
+
self.toRGB.append(EqualConv2d(channels[f'{2**i}'], 3, 1, stride=1, padding=0, bias=True, bias_init_val=0))
|
| 216 |
+
|
| 217 |
+
if different_w:
|
| 218 |
+
linear_out_channel = (int(math.log(out_size, 2)) * 2 - 2) * num_style_feat
|
| 219 |
+
else:
|
| 220 |
+
linear_out_channel = num_style_feat
|
| 221 |
+
|
| 222 |
+
self.final_linear = EqualLinear(
|
| 223 |
+
channels['4'] * 4 * 4, linear_out_channel, bias=True, bias_init_val=0, lr_mul=1, activation=None)
|
| 224 |
+
|
| 225 |
+
# the decoder: stylegan2 generator with SFT modulations
|
| 226 |
+
self.stylegan_decoder = StyleGAN2GeneratorBilinearSFT(
|
| 227 |
+
out_size=out_size,
|
| 228 |
+
num_style_feat=num_style_feat,
|
| 229 |
+
num_mlp=num_mlp,
|
| 230 |
+
channel_multiplier=channel_multiplier,
|
| 231 |
+
lr_mlp=lr_mlp,
|
| 232 |
+
narrow=narrow,
|
| 233 |
+
sft_half=sft_half)
|
| 234 |
+
|
| 235 |
+
# load pre-trained stylegan2 model if necessary
|
| 236 |
+
if decoder_load_path:
|
| 237 |
+
self.stylegan_decoder.load_state_dict(
|
| 238 |
+
torch.load(decoder_load_path, map_location=lambda storage, loc: storage)['params_ema'])
|
| 239 |
+
# fix decoder without updating params
|
| 240 |
+
if fix_decoder:
|
| 241 |
+
for _, param in self.stylegan_decoder.named_parameters():
|
| 242 |
+
param.requires_grad = False
|
| 243 |
+
|
| 244 |
+
# for SFT modulations (scale and shift)
|
| 245 |
+
self.condition_scale = nn.ModuleList()
|
| 246 |
+
self.condition_shift = nn.ModuleList()
|
| 247 |
+
for i in range(3, self.log_size + 1):
|
| 248 |
+
out_channels = channels[f'{2**i}']
|
| 249 |
+
if sft_half:
|
| 250 |
+
sft_out_channels = out_channels
|
| 251 |
+
else:
|
| 252 |
+
sft_out_channels = out_channels * 2
|
| 253 |
+
self.condition_scale.append(
|
| 254 |
+
nn.Sequential(
|
| 255 |
+
EqualConv2d(out_channels, out_channels, 3, stride=1, padding=1, bias=True, bias_init_val=0),
|
| 256 |
+
ScaledLeakyReLU(0.2),
|
| 257 |
+
EqualConv2d(out_channels, sft_out_channels, 3, stride=1, padding=1, bias=True, bias_init_val=1)))
|
| 258 |
+
self.condition_shift.append(
|
| 259 |
+
nn.Sequential(
|
| 260 |
+
EqualConv2d(out_channels, out_channels, 3, stride=1, padding=1, bias=True, bias_init_val=0),
|
| 261 |
+
ScaledLeakyReLU(0.2),
|
| 262 |
+
EqualConv2d(out_channels, sft_out_channels, 3, stride=1, padding=1, bias=True, bias_init_val=0)))
|
| 263 |
+
|
| 264 |
+
def forward(self, x, return_latents=False, return_rgb=True, randomize_noise=True):
|
| 265 |
+
"""Forward function for GFPGANBilinear.
|
| 266 |
+
|
| 267 |
+
Args:
|
| 268 |
+
x (Tensor): Input images.
|
| 269 |
+
return_latents (bool): Whether to return style latents. Default: False.
|
| 270 |
+
return_rgb (bool): Whether return intermediate rgb images. Default: True.
|
| 271 |
+
randomize_noise (bool): Randomize noise, used when 'noise' is False. Default: True.
|
| 272 |
+
"""
|
| 273 |
+
conditions = []
|
| 274 |
+
unet_skips = []
|
| 275 |
+
out_rgbs = []
|
| 276 |
+
|
| 277 |
+
# encoder
|
| 278 |
+
feat = self.conv_body_first(x)
|
| 279 |
+
for i in range(self.log_size - 2):
|
| 280 |
+
feat = self.conv_body_down[i](feat)
|
| 281 |
+
unet_skips.insert(0, feat)
|
| 282 |
+
|
| 283 |
+
feat = self.final_conv(feat)
|
| 284 |
+
|
| 285 |
+
# style code
|
| 286 |
+
style_code = self.final_linear(feat.view(feat.size(0), -1))
|
| 287 |
+
if self.different_w:
|
| 288 |
+
style_code = style_code.view(style_code.size(0), -1, self.num_style_feat)
|
| 289 |
+
|
| 290 |
+
# decode
|
| 291 |
+
for i in range(self.log_size - 2):
|
| 292 |
+
# add unet skip
|
| 293 |
+
feat = feat + unet_skips[i]
|
| 294 |
+
# ResUpLayer
|
| 295 |
+
feat = self.conv_body_up[i](feat)
|
| 296 |
+
# generate scale and shift for SFT layers
|
| 297 |
+
scale = self.condition_scale[i](feat)
|
| 298 |
+
conditions.append(scale.clone())
|
| 299 |
+
shift = self.condition_shift[i](feat)
|
| 300 |
+
conditions.append(shift.clone())
|
| 301 |
+
# generate rgb images
|
| 302 |
+
if return_rgb:
|
| 303 |
+
out_rgbs.append(self.toRGB[i](feat))
|
| 304 |
+
|
| 305 |
+
# decoder
|
| 306 |
+
image, _ = self.stylegan_decoder([style_code],
|
| 307 |
+
conditions,
|
| 308 |
+
return_latents=return_latents,
|
| 309 |
+
input_is_latent=self.input_is_latent,
|
| 310 |
+
randomize_noise=randomize_noise)
|
| 311 |
+
|
| 312 |
+
return image, out_rgbs
|
gfpgan/archs/gfpganv1_arch.py
ADDED
|
@@ -0,0 +1,439 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import math
|
| 2 |
+
import random
|
| 3 |
+
import torch
|
| 4 |
+
from basicsr.archs.stylegan2_arch import (ConvLayer, EqualConv2d, EqualLinear, ResBlock, ScaledLeakyReLU,
|
| 5 |
+
StyleGAN2Generator)
|
| 6 |
+
from basicsr.ops.fused_act import FusedLeakyReLU
|
| 7 |
+
from basicsr.utils.registry import ARCH_REGISTRY
|
| 8 |
+
from torch import nn
|
| 9 |
+
from torch.nn import functional as F
|
| 10 |
+
|
| 11 |
+
|
| 12 |
+
class StyleGAN2GeneratorSFT(StyleGAN2Generator):
|
| 13 |
+
"""StyleGAN2 Generator with SFT modulation (Spatial Feature Transform).
|
| 14 |
+
|
| 15 |
+
Args:
|
| 16 |
+
out_size (int): The spatial size of outputs.
|
| 17 |
+
num_style_feat (int): Channel number of style features. Default: 512.
|
| 18 |
+
num_mlp (int): Layer number of MLP style layers. Default: 8.
|
| 19 |
+
channel_multiplier (int): Channel multiplier for large networks of StyleGAN2. Default: 2.
|
| 20 |
+
resample_kernel (list[int]): A list indicating the 1D resample kernel magnitude. A cross production will be
|
| 21 |
+
applied to extent 1D resample kernel to 2D resample kernel. Default: (1, 3, 3, 1).
|
| 22 |
+
lr_mlp (float): Learning rate multiplier for mlp layers. Default: 0.01.
|
| 23 |
+
narrow (float): The narrow ratio for channels. Default: 1.
|
| 24 |
+
sft_half (bool): Whether to apply SFT on half of the input channels. Default: False.
|
| 25 |
+
"""
|
| 26 |
+
|
| 27 |
+
def __init__(self,
|
| 28 |
+
out_size,
|
| 29 |
+
num_style_feat=512,
|
| 30 |
+
num_mlp=8,
|
| 31 |
+
channel_multiplier=2,
|
| 32 |
+
resample_kernel=(1, 3, 3, 1),
|
| 33 |
+
lr_mlp=0.01,
|
| 34 |
+
narrow=1,
|
| 35 |
+
sft_half=False):
|
| 36 |
+
super(StyleGAN2GeneratorSFT, self).__init__(
|
| 37 |
+
out_size,
|
| 38 |
+
num_style_feat=num_style_feat,
|
| 39 |
+
num_mlp=num_mlp,
|
| 40 |
+
channel_multiplier=channel_multiplier,
|
| 41 |
+
resample_kernel=resample_kernel,
|
| 42 |
+
lr_mlp=lr_mlp,
|
| 43 |
+
narrow=narrow)
|
| 44 |
+
self.sft_half = sft_half
|
| 45 |
+
|
| 46 |
+
def forward(self,
|
| 47 |
+
styles,
|
| 48 |
+
conditions,
|
| 49 |
+
input_is_latent=False,
|
| 50 |
+
noise=None,
|
| 51 |
+
randomize_noise=True,
|
| 52 |
+
truncation=1,
|
| 53 |
+
truncation_latent=None,
|
| 54 |
+
inject_index=None,
|
| 55 |
+
return_latents=False):
|
| 56 |
+
"""Forward function for StyleGAN2GeneratorSFT.
|
| 57 |
+
|
| 58 |
+
Args:
|
| 59 |
+
styles (list[Tensor]): Sample codes of styles.
|
| 60 |
+
conditions (list[Tensor]): SFT conditions to generators.
|
| 61 |
+
input_is_latent (bool): Whether input is latent style. Default: False.
|
| 62 |
+
noise (Tensor | None): Input noise or None. Default: None.
|
| 63 |
+
randomize_noise (bool): Randomize noise, used when 'noise' is False. Default: True.
|
| 64 |
+
truncation (float): The truncation ratio. Default: 1.
|
| 65 |
+
truncation_latent (Tensor | None): The truncation latent tensor. Default: None.
|
| 66 |
+
inject_index (int | None): The injection index for mixing noise. Default: None.
|
| 67 |
+
return_latents (bool): Whether to return style latents. Default: False.
|
| 68 |
+
"""
|
| 69 |
+
# style codes -> latents with Style MLP layer
|
| 70 |
+
if not input_is_latent:
|
| 71 |
+
styles = [self.style_mlp(s) for s in styles]
|
| 72 |
+
# noises
|
| 73 |
+
if noise is None:
|
| 74 |
+
if randomize_noise:
|
| 75 |
+
noise = [None] * self.num_layers # for each style conv layer
|
| 76 |
+
else: # use the stored noise
|
| 77 |
+
noise = [getattr(self.noises, f'noise{i}') for i in range(self.num_layers)]
|
| 78 |
+
# style truncation
|
| 79 |
+
if truncation < 1:
|
| 80 |
+
style_truncation = []
|
| 81 |
+
for style in styles:
|
| 82 |
+
style_truncation.append(truncation_latent + truncation * (style - truncation_latent))
|
| 83 |
+
styles = style_truncation
|
| 84 |
+
# get style latents with injection
|
| 85 |
+
if len(styles) == 1:
|
| 86 |
+
inject_index = self.num_latent
|
| 87 |
+
|
| 88 |
+
if styles[0].ndim < 3:
|
| 89 |
+
# repeat latent code for all the layers
|
| 90 |
+
latent = styles[0].unsqueeze(1).repeat(1, inject_index, 1)
|
| 91 |
+
else: # used for encoder with different latent code for each layer
|
| 92 |
+
latent = styles[0]
|
| 93 |
+
elif len(styles) == 2: # mixing noises
|
| 94 |
+
if inject_index is None:
|
| 95 |
+
inject_index = random.randint(1, self.num_latent - 1)
|
| 96 |
+
latent1 = styles[0].unsqueeze(1).repeat(1, inject_index, 1)
|
| 97 |
+
latent2 = styles[1].unsqueeze(1).repeat(1, self.num_latent - inject_index, 1)
|
| 98 |
+
latent = torch.cat([latent1, latent2], 1)
|
| 99 |
+
|
| 100 |
+
# main generation
|
| 101 |
+
out = self.constant_input(latent.shape[0])
|
| 102 |
+
out = self.style_conv1(out, latent[:, 0], noise=noise[0])
|
| 103 |
+
skip = self.to_rgb1(out, latent[:, 1])
|
| 104 |
+
|
| 105 |
+
i = 1
|
| 106 |
+
for conv1, conv2, noise1, noise2, to_rgb in zip(self.style_convs[::2], self.style_convs[1::2], noise[1::2],
|
| 107 |
+
noise[2::2], self.to_rgbs):
|
| 108 |
+
out = conv1(out, latent[:, i], noise=noise1)
|
| 109 |
+
|
| 110 |
+
# the conditions may have fewer levels
|
| 111 |
+
if i < len(conditions):
|
| 112 |
+
# SFT part to combine the conditions
|
| 113 |
+
if self.sft_half: # only apply SFT to half of the channels
|
| 114 |
+
out_same, out_sft = torch.split(out, int(out.size(1) // 2), dim=1)
|
| 115 |
+
out_sft = out_sft * conditions[i - 1] + conditions[i]
|
| 116 |
+
out = torch.cat([out_same, out_sft], dim=1)
|
| 117 |
+
else: # apply SFT to all the channels
|
| 118 |
+
out = out * conditions[i - 1] + conditions[i]
|
| 119 |
+
|
| 120 |
+
out = conv2(out, latent[:, i + 1], noise=noise2)
|
| 121 |
+
skip = to_rgb(out, latent[:, i + 2], skip) # feature back to the rgb space
|
| 122 |
+
i += 2
|
| 123 |
+
|
| 124 |
+
image = skip
|
| 125 |
+
|
| 126 |
+
if return_latents:
|
| 127 |
+
return image, latent
|
| 128 |
+
else:
|
| 129 |
+
return image, None
|
| 130 |
+
|
| 131 |
+
|
| 132 |
+
class ConvUpLayer(nn.Module):
|
| 133 |
+
"""Convolutional upsampling layer. It uses bilinear upsampler + Conv.
|
| 134 |
+
|
| 135 |
+
Args:
|
| 136 |
+
in_channels (int): Channel number of the input.
|
| 137 |
+
out_channels (int): Channel number of the output.
|
| 138 |
+
kernel_size (int): Size of the convolving kernel.
|
| 139 |
+
stride (int): Stride of the convolution. Default: 1
|
| 140 |
+
padding (int): Zero-padding added to both sides of the input. Default: 0.
|
| 141 |
+
bias (bool): If ``True``, adds a learnable bias to the output. Default: ``True``.
|
| 142 |
+
bias_init_val (float): Bias initialized value. Default: 0.
|
| 143 |
+
activate (bool): Whether use activateion. Default: True.
|
| 144 |
+
"""
|
| 145 |
+
|
| 146 |
+
def __init__(self,
|
| 147 |
+
in_channels,
|
| 148 |
+
out_channels,
|
| 149 |
+
kernel_size,
|
| 150 |
+
stride=1,
|
| 151 |
+
padding=0,
|
| 152 |
+
bias=True,
|
| 153 |
+
bias_init_val=0,
|
| 154 |
+
activate=True):
|
| 155 |
+
super(ConvUpLayer, self).__init__()
|
| 156 |
+
self.in_channels = in_channels
|
| 157 |
+
self.out_channels = out_channels
|
| 158 |
+
self.kernel_size = kernel_size
|
| 159 |
+
self.stride = stride
|
| 160 |
+
self.padding = padding
|
| 161 |
+
# self.scale is used to scale the convolution weights, which is related to the common initializations.
|
| 162 |
+
self.scale = 1 / math.sqrt(in_channels * kernel_size**2)
|
| 163 |
+
|
| 164 |
+
self.weight = nn.Parameter(torch.randn(out_channels, in_channels, kernel_size, kernel_size))
|
| 165 |
+
|
| 166 |
+
if bias and not activate:
|
| 167 |
+
self.bias = nn.Parameter(torch.zeros(out_channels).fill_(bias_init_val))
|
| 168 |
+
else:
|
| 169 |
+
self.register_parameter('bias', None)
|
| 170 |
+
|
| 171 |
+
# activation
|
| 172 |
+
if activate:
|
| 173 |
+
if bias:
|
| 174 |
+
self.activation = FusedLeakyReLU(out_channels)
|
| 175 |
+
else:
|
| 176 |
+
self.activation = ScaledLeakyReLU(0.2)
|
| 177 |
+
else:
|
| 178 |
+
self.activation = None
|
| 179 |
+
|
| 180 |
+
def forward(self, x):
|
| 181 |
+
# bilinear upsample
|
| 182 |
+
out = F.interpolate(x, scale_factor=2, mode='bilinear', align_corners=False)
|
| 183 |
+
# conv
|
| 184 |
+
out = F.conv2d(
|
| 185 |
+
out,
|
| 186 |
+
self.weight * self.scale,
|
| 187 |
+
bias=self.bias,
|
| 188 |
+
stride=self.stride,
|
| 189 |
+
padding=self.padding,
|
| 190 |
+
)
|
| 191 |
+
# activation
|
| 192 |
+
if self.activation is not None:
|
| 193 |
+
out = self.activation(out)
|
| 194 |
+
return out
|
| 195 |
+
|
| 196 |
+
|
| 197 |
+
class ResUpBlock(nn.Module):
|
| 198 |
+
"""Residual block with upsampling.
|
| 199 |
+
|
| 200 |
+
Args:
|
| 201 |
+
in_channels (int): Channel number of the input.
|
| 202 |
+
out_channels (int): Channel number of the output.
|
| 203 |
+
"""
|
| 204 |
+
|
| 205 |
+
def __init__(self, in_channels, out_channels):
|
| 206 |
+
super(ResUpBlock, self).__init__()
|
| 207 |
+
|
| 208 |
+
self.conv1 = ConvLayer(in_channels, in_channels, 3, bias=True, activate=True)
|
| 209 |
+
self.conv2 = ConvUpLayer(in_channels, out_channels, 3, stride=1, padding=1, bias=True, activate=True)
|
| 210 |
+
self.skip = ConvUpLayer(in_channels, out_channels, 1, bias=False, activate=False)
|
| 211 |
+
|
| 212 |
+
def forward(self, x):
|
| 213 |
+
out = self.conv1(x)
|
| 214 |
+
out = self.conv2(out)
|
| 215 |
+
skip = self.skip(x)
|
| 216 |
+
out = (out + skip) / math.sqrt(2)
|
| 217 |
+
return out
|
| 218 |
+
|
| 219 |
+
|
| 220 |
+
@ARCH_REGISTRY.register()
|
| 221 |
+
class GFPGANv1(nn.Module):
|
| 222 |
+
"""The GFPGAN architecture: Unet + StyleGAN2 decoder with SFT.
|
| 223 |
+
|
| 224 |
+
Ref: GFP-GAN: Towards Real-World Blind Face Restoration with Generative Facial Prior.
|
| 225 |
+
|
| 226 |
+
Args:
|
| 227 |
+
out_size (int): The spatial size of outputs.
|
| 228 |
+
num_style_feat (int): Channel number of style features. Default: 512.
|
| 229 |
+
channel_multiplier (int): Channel multiplier for large networks of StyleGAN2. Default: 2.
|
| 230 |
+
resample_kernel (list[int]): A list indicating the 1D resample kernel magnitude. A cross production will be
|
| 231 |
+
applied to extent 1D resample kernel to 2D resample kernel. Default: (1, 3, 3, 1).
|
| 232 |
+
decoder_load_path (str): The path to the pre-trained decoder model (usually, the StyleGAN2). Default: None.
|
| 233 |
+
fix_decoder (bool): Whether to fix the decoder. Default: True.
|
| 234 |
+
|
| 235 |
+
num_mlp (int): Layer number of MLP style layers. Default: 8.
|
| 236 |
+
lr_mlp (float): Learning rate multiplier for mlp layers. Default: 0.01.
|
| 237 |
+
input_is_latent (bool): Whether input is latent style. Default: False.
|
| 238 |
+
different_w (bool): Whether to use different latent w for different layers. Default: False.
|
| 239 |
+
narrow (float): The narrow ratio for channels. Default: 1.
|
| 240 |
+
sft_half (bool): Whether to apply SFT on half of the input channels. Default: False.
|
| 241 |
+
"""
|
| 242 |
+
|
| 243 |
+
def __init__(
|
| 244 |
+
self,
|
| 245 |
+
out_size,
|
| 246 |
+
num_style_feat=512,
|
| 247 |
+
channel_multiplier=1,
|
| 248 |
+
resample_kernel=(1, 3, 3, 1),
|
| 249 |
+
decoder_load_path=None,
|
| 250 |
+
fix_decoder=True,
|
| 251 |
+
# for stylegan decoder
|
| 252 |
+
num_mlp=8,
|
| 253 |
+
lr_mlp=0.01,
|
| 254 |
+
input_is_latent=False,
|
| 255 |
+
different_w=False,
|
| 256 |
+
narrow=1,
|
| 257 |
+
sft_half=False):
|
| 258 |
+
|
| 259 |
+
super(GFPGANv1, self).__init__()
|
| 260 |
+
self.input_is_latent = input_is_latent
|
| 261 |
+
self.different_w = different_w
|
| 262 |
+
self.num_style_feat = num_style_feat
|
| 263 |
+
|
| 264 |
+
unet_narrow = narrow * 0.5 # by default, use a half of input channels
|
| 265 |
+
channels = {
|
| 266 |
+
'4': int(512 * unet_narrow),
|
| 267 |
+
'8': int(512 * unet_narrow),
|
| 268 |
+
'16': int(512 * unet_narrow),
|
| 269 |
+
'32': int(512 * unet_narrow),
|
| 270 |
+
'64': int(256 * channel_multiplier * unet_narrow),
|
| 271 |
+
'128': int(128 * channel_multiplier * unet_narrow),
|
| 272 |
+
'256': int(64 * channel_multiplier * unet_narrow),
|
| 273 |
+
'512': int(32 * channel_multiplier * unet_narrow),
|
| 274 |
+
'1024': int(16 * channel_multiplier * unet_narrow)
|
| 275 |
+
}
|
| 276 |
+
|
| 277 |
+
self.log_size = int(math.log(out_size, 2))
|
| 278 |
+
first_out_size = 2**(int(math.log(out_size, 2)))
|
| 279 |
+
|
| 280 |
+
self.conv_body_first = ConvLayer(3, channels[f'{first_out_size}'], 1, bias=True, activate=True)
|
| 281 |
+
|
| 282 |
+
# downsample
|
| 283 |
+
in_channels = channels[f'{first_out_size}']
|
| 284 |
+
self.conv_body_down = nn.ModuleList()
|
| 285 |
+
for i in range(self.log_size, 2, -1):
|
| 286 |
+
out_channels = channels[f'{2**(i - 1)}']
|
| 287 |
+
self.conv_body_down.append(ResBlock(in_channels, out_channels, resample_kernel))
|
| 288 |
+
in_channels = out_channels
|
| 289 |
+
|
| 290 |
+
self.final_conv = ConvLayer(in_channels, channels['4'], 3, bias=True, activate=True)
|
| 291 |
+
|
| 292 |
+
# upsample
|
| 293 |
+
in_channels = channels['4']
|
| 294 |
+
self.conv_body_up = nn.ModuleList()
|
| 295 |
+
for i in range(3, self.log_size + 1):
|
| 296 |
+
out_channels = channels[f'{2**i}']
|
| 297 |
+
self.conv_body_up.append(ResUpBlock(in_channels, out_channels))
|
| 298 |
+
in_channels = out_channels
|
| 299 |
+
|
| 300 |
+
# to RGB
|
| 301 |
+
self.toRGB = nn.ModuleList()
|
| 302 |
+
for i in range(3, self.log_size + 1):
|
| 303 |
+
self.toRGB.append(EqualConv2d(channels[f'{2**i}'], 3, 1, stride=1, padding=0, bias=True, bias_init_val=0))
|
| 304 |
+
|
| 305 |
+
if different_w:
|
| 306 |
+
linear_out_channel = (int(math.log(out_size, 2)) * 2 - 2) * num_style_feat
|
| 307 |
+
else:
|
| 308 |
+
linear_out_channel = num_style_feat
|
| 309 |
+
|
| 310 |
+
self.final_linear = EqualLinear(
|
| 311 |
+
channels['4'] * 4 * 4, linear_out_channel, bias=True, bias_init_val=0, lr_mul=1, activation=None)
|
| 312 |
+
|
| 313 |
+
# the decoder: stylegan2 generator with SFT modulations
|
| 314 |
+
self.stylegan_decoder = StyleGAN2GeneratorSFT(
|
| 315 |
+
out_size=out_size,
|
| 316 |
+
num_style_feat=num_style_feat,
|
| 317 |
+
num_mlp=num_mlp,
|
| 318 |
+
channel_multiplier=channel_multiplier,
|
| 319 |
+
resample_kernel=resample_kernel,
|
| 320 |
+
lr_mlp=lr_mlp,
|
| 321 |
+
narrow=narrow,
|
| 322 |
+
sft_half=sft_half)
|
| 323 |
+
|
| 324 |
+
# load pre-trained stylegan2 model if necessary
|
| 325 |
+
if decoder_load_path:
|
| 326 |
+
self.stylegan_decoder.load_state_dict(
|
| 327 |
+
torch.load(decoder_load_path, map_location=lambda storage, loc: storage)['params_ema'])
|
| 328 |
+
# fix decoder without updating params
|
| 329 |
+
if fix_decoder:
|
| 330 |
+
for _, param in self.stylegan_decoder.named_parameters():
|
| 331 |
+
param.requires_grad = False
|
| 332 |
+
|
| 333 |
+
# for SFT modulations (scale and shift)
|
| 334 |
+
self.condition_scale = nn.ModuleList()
|
| 335 |
+
self.condition_shift = nn.ModuleList()
|
| 336 |
+
for i in range(3, self.log_size + 1):
|
| 337 |
+
out_channels = channels[f'{2**i}']
|
| 338 |
+
if sft_half:
|
| 339 |
+
sft_out_channels = out_channels
|
| 340 |
+
else:
|
| 341 |
+
sft_out_channels = out_channels * 2
|
| 342 |
+
self.condition_scale.append(
|
| 343 |
+
nn.Sequential(
|
| 344 |
+
EqualConv2d(out_channels, out_channels, 3, stride=1, padding=1, bias=True, bias_init_val=0),
|
| 345 |
+
ScaledLeakyReLU(0.2),
|
| 346 |
+
EqualConv2d(out_channels, sft_out_channels, 3, stride=1, padding=1, bias=True, bias_init_val=1)))
|
| 347 |
+
self.condition_shift.append(
|
| 348 |
+
nn.Sequential(
|
| 349 |
+
EqualConv2d(out_channels, out_channels, 3, stride=1, padding=1, bias=True, bias_init_val=0),
|
| 350 |
+
ScaledLeakyReLU(0.2),
|
| 351 |
+
EqualConv2d(out_channels, sft_out_channels, 3, stride=1, padding=1, bias=True, bias_init_val=0)))
|
| 352 |
+
|
| 353 |
+
def forward(self, x, return_latents=False, return_rgb=True, randomize_noise=True, **kwargs):
|
| 354 |
+
"""Forward function for GFPGANv1.
|
| 355 |
+
|
| 356 |
+
Args:
|
| 357 |
+
x (Tensor): Input images.
|
| 358 |
+
return_latents (bool): Whether to return style latents. Default: False.
|
| 359 |
+
return_rgb (bool): Whether return intermediate rgb images. Default: True.
|
| 360 |
+
randomize_noise (bool): Randomize noise, used when 'noise' is False. Default: True.
|
| 361 |
+
"""
|
| 362 |
+
conditions = []
|
| 363 |
+
unet_skips = []
|
| 364 |
+
out_rgbs = []
|
| 365 |
+
|
| 366 |
+
# encoder
|
| 367 |
+
feat = self.conv_body_first(x)
|
| 368 |
+
for i in range(self.log_size - 2):
|
| 369 |
+
feat = self.conv_body_down[i](feat)
|
| 370 |
+
unet_skips.insert(0, feat)
|
| 371 |
+
|
| 372 |
+
feat = self.final_conv(feat)
|
| 373 |
+
|
| 374 |
+
# style code
|
| 375 |
+
style_code = self.final_linear(feat.view(feat.size(0), -1))
|
| 376 |
+
if self.different_w:
|
| 377 |
+
style_code = style_code.view(style_code.size(0), -1, self.num_style_feat)
|
| 378 |
+
|
| 379 |
+
# decode
|
| 380 |
+
for i in range(self.log_size - 2):
|
| 381 |
+
# add unet skip
|
| 382 |
+
feat = feat + unet_skips[i]
|
| 383 |
+
# ResUpLayer
|
| 384 |
+
feat = self.conv_body_up[i](feat)
|
| 385 |
+
# generate scale and shift for SFT layers
|
| 386 |
+
scale = self.condition_scale[i](feat)
|
| 387 |
+
conditions.append(scale.clone())
|
| 388 |
+
shift = self.condition_shift[i](feat)
|
| 389 |
+
conditions.append(shift.clone())
|
| 390 |
+
# generate rgb images
|
| 391 |
+
if return_rgb:
|
| 392 |
+
out_rgbs.append(self.toRGB[i](feat))
|
| 393 |
+
|
| 394 |
+
# decoder
|
| 395 |
+
image, _ = self.stylegan_decoder([style_code],
|
| 396 |
+
conditions,
|
| 397 |
+
return_latents=return_latents,
|
| 398 |
+
input_is_latent=self.input_is_latent,
|
| 399 |
+
randomize_noise=randomize_noise)
|
| 400 |
+
|
| 401 |
+
return image, out_rgbs
|
| 402 |
+
|
| 403 |
+
|
| 404 |
+
@ARCH_REGISTRY.register()
|
| 405 |
+
class FacialComponentDiscriminator(nn.Module):
|
| 406 |
+
"""Facial component (eyes, mouth, noise) discriminator used in GFPGAN.
|
| 407 |
+
"""
|
| 408 |
+
|
| 409 |
+
def __init__(self):
|
| 410 |
+
super(FacialComponentDiscriminator, self).__init__()
|
| 411 |
+
# It now uses a VGG-style architectrue with fixed model size
|
| 412 |
+
self.conv1 = ConvLayer(3, 64, 3, downsample=False, resample_kernel=(1, 3, 3, 1), bias=True, activate=True)
|
| 413 |
+
self.conv2 = ConvLayer(64, 128, 3, downsample=True, resample_kernel=(1, 3, 3, 1), bias=True, activate=True)
|
| 414 |
+
self.conv3 = ConvLayer(128, 128, 3, downsample=False, resample_kernel=(1, 3, 3, 1), bias=True, activate=True)
|
| 415 |
+
self.conv4 = ConvLayer(128, 256, 3, downsample=True, resample_kernel=(1, 3, 3, 1), bias=True, activate=True)
|
| 416 |
+
self.conv5 = ConvLayer(256, 256, 3, downsample=False, resample_kernel=(1, 3, 3, 1), bias=True, activate=True)
|
| 417 |
+
self.final_conv = ConvLayer(256, 1, 3, bias=True, activate=False)
|
| 418 |
+
|
| 419 |
+
def forward(self, x, return_feats=False, **kwargs):
|
| 420 |
+
"""Forward function for FacialComponentDiscriminator.
|
| 421 |
+
|
| 422 |
+
Args:
|
| 423 |
+
x (Tensor): Input images.
|
| 424 |
+
return_feats (bool): Whether to return intermediate features. Default: False.
|
| 425 |
+
"""
|
| 426 |
+
feat = self.conv1(x)
|
| 427 |
+
feat = self.conv3(self.conv2(feat))
|
| 428 |
+
rlt_feats = []
|
| 429 |
+
if return_feats:
|
| 430 |
+
rlt_feats.append(feat.clone())
|
| 431 |
+
feat = self.conv5(self.conv4(feat))
|
| 432 |
+
if return_feats:
|
| 433 |
+
rlt_feats.append(feat.clone())
|
| 434 |
+
out = self.final_conv(feat)
|
| 435 |
+
|
| 436 |
+
if return_feats:
|
| 437 |
+
return out, rlt_feats
|
| 438 |
+
else:
|
| 439 |
+
return out, None
|
gfpgan/archs/gfpganv1_clean_arch.py
ADDED
|
@@ -0,0 +1,324 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import math
|
| 2 |
+
import random
|
| 3 |
+
import torch
|
| 4 |
+
from basicsr.utils.registry import ARCH_REGISTRY
|
| 5 |
+
from torch import nn
|
| 6 |
+
from torch.nn import functional as F
|
| 7 |
+
|
| 8 |
+
from .stylegan2_clean_arch import StyleGAN2GeneratorClean
|
| 9 |
+
|
| 10 |
+
|
| 11 |
+
class StyleGAN2GeneratorCSFT(StyleGAN2GeneratorClean):
|
| 12 |
+
"""StyleGAN2 Generator with SFT modulation (Spatial Feature Transform).
|
| 13 |
+
|
| 14 |
+
It is the clean version without custom compiled CUDA extensions used in StyleGAN2.
|
| 15 |
+
|
| 16 |
+
Args:
|
| 17 |
+
out_size (int): The spatial size of outputs.
|
| 18 |
+
num_style_feat (int): Channel number of style features. Default: 512.
|
| 19 |
+
num_mlp (int): Layer number of MLP style layers. Default: 8.
|
| 20 |
+
channel_multiplier (int): Channel multiplier for large networks of StyleGAN2. Default: 2.
|
| 21 |
+
narrow (float): The narrow ratio for channels. Default: 1.
|
| 22 |
+
sft_half (bool): Whether to apply SFT on half of the input channels. Default: False.
|
| 23 |
+
"""
|
| 24 |
+
|
| 25 |
+
def __init__(self, out_size, num_style_feat=512, num_mlp=8, channel_multiplier=2, narrow=1, sft_half=False):
|
| 26 |
+
super(StyleGAN2GeneratorCSFT, self).__init__(
|
| 27 |
+
out_size,
|
| 28 |
+
num_style_feat=num_style_feat,
|
| 29 |
+
num_mlp=num_mlp,
|
| 30 |
+
channel_multiplier=channel_multiplier,
|
| 31 |
+
narrow=narrow)
|
| 32 |
+
self.sft_half = sft_half
|
| 33 |
+
|
| 34 |
+
def forward(self,
|
| 35 |
+
styles,
|
| 36 |
+
conditions,
|
| 37 |
+
input_is_latent=False,
|
| 38 |
+
noise=None,
|
| 39 |
+
randomize_noise=True,
|
| 40 |
+
truncation=1,
|
| 41 |
+
truncation_latent=None,
|
| 42 |
+
inject_index=None,
|
| 43 |
+
return_latents=False):
|
| 44 |
+
"""Forward function for StyleGAN2GeneratorCSFT.
|
| 45 |
+
|
| 46 |
+
Args:
|
| 47 |
+
styles (list[Tensor]): Sample codes of styles.
|
| 48 |
+
conditions (list[Tensor]): SFT conditions to generators.
|
| 49 |
+
input_is_latent (bool): Whether input is latent style. Default: False.
|
| 50 |
+
noise (Tensor | None): Input noise or None. Default: None.
|
| 51 |
+
randomize_noise (bool): Randomize noise, used when 'noise' is False. Default: True.
|
| 52 |
+
truncation (float): The truncation ratio. Default: 1.
|
| 53 |
+
truncation_latent (Tensor | None): The truncation latent tensor. Default: None.
|
| 54 |
+
inject_index (int | None): The injection index for mixing noise. Default: None.
|
| 55 |
+
return_latents (bool): Whether to return style latents. Default: False.
|
| 56 |
+
"""
|
| 57 |
+
# style codes -> latents with Style MLP layer
|
| 58 |
+
if not input_is_latent:
|
| 59 |
+
styles = [self.style_mlp(s) for s in styles]
|
| 60 |
+
# noises
|
| 61 |
+
if noise is None:
|
| 62 |
+
if randomize_noise:
|
| 63 |
+
noise = [None] * self.num_layers # for each style conv layer
|
| 64 |
+
else: # use the stored noise
|
| 65 |
+
noise = [getattr(self.noises, f'noise{i}') for i in range(self.num_layers)]
|
| 66 |
+
# style truncation
|
| 67 |
+
if truncation < 1:
|
| 68 |
+
style_truncation = []
|
| 69 |
+
for style in styles:
|
| 70 |
+
style_truncation.append(truncation_latent + truncation * (style - truncation_latent))
|
| 71 |
+
styles = style_truncation
|
| 72 |
+
# get style latents with injection
|
| 73 |
+
if len(styles) == 1:
|
| 74 |
+
inject_index = self.num_latent
|
| 75 |
+
|
| 76 |
+
if styles[0].ndim < 3:
|
| 77 |
+
# repeat latent code for all the layers
|
| 78 |
+
latent = styles[0].unsqueeze(1).repeat(1, inject_index, 1)
|
| 79 |
+
else: # used for encoder with different latent code for each layer
|
| 80 |
+
latent = styles[0]
|
| 81 |
+
elif len(styles) == 2: # mixing noises
|
| 82 |
+
if inject_index is None:
|
| 83 |
+
inject_index = random.randint(1, self.num_latent - 1)
|
| 84 |
+
latent1 = styles[0].unsqueeze(1).repeat(1, inject_index, 1)
|
| 85 |
+
latent2 = styles[1].unsqueeze(1).repeat(1, self.num_latent - inject_index, 1)
|
| 86 |
+
latent = torch.cat([latent1, latent2], 1)
|
| 87 |
+
|
| 88 |
+
# main generation
|
| 89 |
+
out = self.constant_input(latent.shape[0])
|
| 90 |
+
out = self.style_conv1(out, latent[:, 0], noise=noise[0])
|
| 91 |
+
skip = self.to_rgb1(out, latent[:, 1])
|
| 92 |
+
|
| 93 |
+
i = 1
|
| 94 |
+
for conv1, conv2, noise1, noise2, to_rgb in zip(self.style_convs[::2], self.style_convs[1::2], noise[1::2],
|
| 95 |
+
noise[2::2], self.to_rgbs):
|
| 96 |
+
out = conv1(out, latent[:, i], noise=noise1)
|
| 97 |
+
|
| 98 |
+
# the conditions may have fewer levels
|
| 99 |
+
if i < len(conditions):
|
| 100 |
+
# SFT part to combine the conditions
|
| 101 |
+
if self.sft_half: # only apply SFT to half of the channels
|
| 102 |
+
out_same, out_sft = torch.split(out, int(out.size(1) // 2), dim=1)
|
| 103 |
+
out_sft = out_sft * conditions[i - 1] + conditions[i]
|
| 104 |
+
out = torch.cat([out_same, out_sft], dim=1)
|
| 105 |
+
else: # apply SFT to all the channels
|
| 106 |
+
out = out * conditions[i - 1] + conditions[i]
|
| 107 |
+
|
| 108 |
+
out = conv2(out, latent[:, i + 1], noise=noise2)
|
| 109 |
+
skip = to_rgb(out, latent[:, i + 2], skip) # feature back to the rgb space
|
| 110 |
+
i += 2
|
| 111 |
+
|
| 112 |
+
image = skip
|
| 113 |
+
|
| 114 |
+
if return_latents:
|
| 115 |
+
return image, latent
|
| 116 |
+
else:
|
| 117 |
+
return image, None
|
| 118 |
+
|
| 119 |
+
|
| 120 |
+
class ResBlock(nn.Module):
|
| 121 |
+
"""Residual block with bilinear upsampling/downsampling.
|
| 122 |
+
|
| 123 |
+
Args:
|
| 124 |
+
in_channels (int): Channel number of the input.
|
| 125 |
+
out_channels (int): Channel number of the output.
|
| 126 |
+
mode (str): Upsampling/downsampling mode. Options: down | up. Default: down.
|
| 127 |
+
"""
|
| 128 |
+
|
| 129 |
+
def __init__(self, in_channels, out_channels, mode='down'):
|
| 130 |
+
super(ResBlock, self).__init__()
|
| 131 |
+
|
| 132 |
+
self.conv1 = nn.Conv2d(in_channels, in_channels, 3, 1, 1)
|
| 133 |
+
self.conv2 = nn.Conv2d(in_channels, out_channels, 3, 1, 1)
|
| 134 |
+
self.skip = nn.Conv2d(in_channels, out_channels, 1, bias=False)
|
| 135 |
+
if mode == 'down':
|
| 136 |
+
self.scale_factor = 0.5
|
| 137 |
+
elif mode == 'up':
|
| 138 |
+
self.scale_factor = 2
|
| 139 |
+
|
| 140 |
+
def forward(self, x):
|
| 141 |
+
out = F.leaky_relu_(self.conv1(x), negative_slope=0.2)
|
| 142 |
+
# upsample/downsample
|
| 143 |
+
out = F.interpolate(out, scale_factor=self.scale_factor, mode='bilinear', align_corners=False)
|
| 144 |
+
out = F.leaky_relu_(self.conv2(out), negative_slope=0.2)
|
| 145 |
+
# skip
|
| 146 |
+
x = F.interpolate(x, scale_factor=self.scale_factor, mode='bilinear', align_corners=False)
|
| 147 |
+
skip = self.skip(x)
|
| 148 |
+
out = out + skip
|
| 149 |
+
return out
|
| 150 |
+
|
| 151 |
+
|
| 152 |
+
@ARCH_REGISTRY.register()
|
| 153 |
+
class GFPGANv1Clean(nn.Module):
|
| 154 |
+
"""The GFPGAN architecture: Unet + StyleGAN2 decoder with SFT.
|
| 155 |
+
|
| 156 |
+
It is the clean version without custom compiled CUDA extensions used in StyleGAN2.
|
| 157 |
+
|
| 158 |
+
Ref: GFP-GAN: Towards Real-World Blind Face Restoration with Generative Facial Prior.
|
| 159 |
+
|
| 160 |
+
Args:
|
| 161 |
+
out_size (int): The spatial size of outputs.
|
| 162 |
+
num_style_feat (int): Channel number of style features. Default: 512.
|
| 163 |
+
channel_multiplier (int): Channel multiplier for large networks of StyleGAN2. Default: 2.
|
| 164 |
+
decoder_load_path (str): The path to the pre-trained decoder model (usually, the StyleGAN2). Default: None.
|
| 165 |
+
fix_decoder (bool): Whether to fix the decoder. Default: True.
|
| 166 |
+
|
| 167 |
+
num_mlp (int): Layer number of MLP style layers. Default: 8.
|
| 168 |
+
input_is_latent (bool): Whether input is latent style. Default: False.
|
| 169 |
+
different_w (bool): Whether to use different latent w for different layers. Default: False.
|
| 170 |
+
narrow (float): The narrow ratio for channels. Default: 1.
|
| 171 |
+
sft_half (bool): Whether to apply SFT on half of the input channels. Default: False.
|
| 172 |
+
"""
|
| 173 |
+
|
| 174 |
+
def __init__(
|
| 175 |
+
self,
|
| 176 |
+
out_size,
|
| 177 |
+
num_style_feat=512,
|
| 178 |
+
channel_multiplier=1,
|
| 179 |
+
decoder_load_path=None,
|
| 180 |
+
fix_decoder=True,
|
| 181 |
+
# for stylegan decoder
|
| 182 |
+
num_mlp=8,
|
| 183 |
+
input_is_latent=False,
|
| 184 |
+
different_w=False,
|
| 185 |
+
narrow=1,
|
| 186 |
+
sft_half=False):
|
| 187 |
+
|
| 188 |
+
super(GFPGANv1Clean, self).__init__()
|
| 189 |
+
self.input_is_latent = input_is_latent
|
| 190 |
+
self.different_w = different_w
|
| 191 |
+
self.num_style_feat = num_style_feat
|
| 192 |
+
|
| 193 |
+
unet_narrow = narrow * 0.5 # by default, use a half of input channels
|
| 194 |
+
channels = {
|
| 195 |
+
'4': int(512 * unet_narrow),
|
| 196 |
+
'8': int(512 * unet_narrow),
|
| 197 |
+
'16': int(512 * unet_narrow),
|
| 198 |
+
'32': int(512 * unet_narrow),
|
| 199 |
+
'64': int(256 * channel_multiplier * unet_narrow),
|
| 200 |
+
'128': int(128 * channel_multiplier * unet_narrow),
|
| 201 |
+
'256': int(64 * channel_multiplier * unet_narrow),
|
| 202 |
+
'512': int(32 * channel_multiplier * unet_narrow),
|
| 203 |
+
'1024': int(16 * channel_multiplier * unet_narrow)
|
| 204 |
+
}
|
| 205 |
+
|
| 206 |
+
self.log_size = int(math.log(out_size, 2))
|
| 207 |
+
first_out_size = 2**(int(math.log(out_size, 2)))
|
| 208 |
+
|
| 209 |
+
self.conv_body_first = nn.Conv2d(3, channels[f'{first_out_size}'], 1)
|
| 210 |
+
|
| 211 |
+
# downsample
|
| 212 |
+
in_channels = channels[f'{first_out_size}']
|
| 213 |
+
self.conv_body_down = nn.ModuleList()
|
| 214 |
+
for i in range(self.log_size, 2, -1):
|
| 215 |
+
out_channels = channels[f'{2**(i - 1)}']
|
| 216 |
+
self.conv_body_down.append(ResBlock(in_channels, out_channels, mode='down'))
|
| 217 |
+
in_channels = out_channels
|
| 218 |
+
|
| 219 |
+
self.final_conv = nn.Conv2d(in_channels, channels['4'], 3, 1, 1)
|
| 220 |
+
|
| 221 |
+
# upsample
|
| 222 |
+
in_channels = channels['4']
|
| 223 |
+
self.conv_body_up = nn.ModuleList()
|
| 224 |
+
for i in range(3, self.log_size + 1):
|
| 225 |
+
out_channels = channels[f'{2**i}']
|
| 226 |
+
self.conv_body_up.append(ResBlock(in_channels, out_channels, mode='up'))
|
| 227 |
+
in_channels = out_channels
|
| 228 |
+
|
| 229 |
+
# to RGB
|
| 230 |
+
self.toRGB = nn.ModuleList()
|
| 231 |
+
for i in range(3, self.log_size + 1):
|
| 232 |
+
self.toRGB.append(nn.Conv2d(channels[f'{2**i}'], 3, 1))
|
| 233 |
+
|
| 234 |
+
if different_w:
|
| 235 |
+
linear_out_channel = (int(math.log(out_size, 2)) * 2 - 2) * num_style_feat
|
| 236 |
+
else:
|
| 237 |
+
linear_out_channel = num_style_feat
|
| 238 |
+
|
| 239 |
+
self.final_linear = nn.Linear(channels['4'] * 4 * 4, linear_out_channel)
|
| 240 |
+
|
| 241 |
+
# the decoder: stylegan2 generator with SFT modulations
|
| 242 |
+
self.stylegan_decoder = StyleGAN2GeneratorCSFT(
|
| 243 |
+
out_size=out_size,
|
| 244 |
+
num_style_feat=num_style_feat,
|
| 245 |
+
num_mlp=num_mlp,
|
| 246 |
+
channel_multiplier=channel_multiplier,
|
| 247 |
+
narrow=narrow,
|
| 248 |
+
sft_half=sft_half)
|
| 249 |
+
|
| 250 |
+
# load pre-trained stylegan2 model if necessary
|
| 251 |
+
if decoder_load_path:
|
| 252 |
+
self.stylegan_decoder.load_state_dict(
|
| 253 |
+
torch.load(decoder_load_path, map_location=lambda storage, loc: storage)['params_ema'])
|
| 254 |
+
# fix decoder without updating params
|
| 255 |
+
if fix_decoder:
|
| 256 |
+
for _, param in self.stylegan_decoder.named_parameters():
|
| 257 |
+
param.requires_grad = False
|
| 258 |
+
|
| 259 |
+
# for SFT modulations (scale and shift)
|
| 260 |
+
self.condition_scale = nn.ModuleList()
|
| 261 |
+
self.condition_shift = nn.ModuleList()
|
| 262 |
+
for i in range(3, self.log_size + 1):
|
| 263 |
+
out_channels = channels[f'{2**i}']
|
| 264 |
+
if sft_half:
|
| 265 |
+
sft_out_channels = out_channels
|
| 266 |
+
else:
|
| 267 |
+
sft_out_channels = out_channels * 2
|
| 268 |
+
self.condition_scale.append(
|
| 269 |
+
nn.Sequential(
|
| 270 |
+
nn.Conv2d(out_channels, out_channels, 3, 1, 1), nn.LeakyReLU(0.2, True),
|
| 271 |
+
nn.Conv2d(out_channels, sft_out_channels, 3, 1, 1)))
|
| 272 |
+
self.condition_shift.append(
|
| 273 |
+
nn.Sequential(
|
| 274 |
+
nn.Conv2d(out_channels, out_channels, 3, 1, 1), nn.LeakyReLU(0.2, True),
|
| 275 |
+
nn.Conv2d(out_channels, sft_out_channels, 3, 1, 1)))
|
| 276 |
+
|
| 277 |
+
def forward(self, x, return_latents=False, return_rgb=True, randomize_noise=True, **kwargs):
|
| 278 |
+
"""Forward function for GFPGANv1Clean.
|
| 279 |
+
|
| 280 |
+
Args:
|
| 281 |
+
x (Tensor): Input images.
|
| 282 |
+
return_latents (bool): Whether to return style latents. Default: False.
|
| 283 |
+
return_rgb (bool): Whether return intermediate rgb images. Default: True.
|
| 284 |
+
randomize_noise (bool): Randomize noise, used when 'noise' is False. Default: True.
|
| 285 |
+
"""
|
| 286 |
+
conditions = []
|
| 287 |
+
unet_skips = []
|
| 288 |
+
out_rgbs = []
|
| 289 |
+
|
| 290 |
+
# encoder
|
| 291 |
+
feat = F.leaky_relu_(self.conv_body_first(x), negative_slope=0.2)
|
| 292 |
+
for i in range(self.log_size - 2):
|
| 293 |
+
feat = self.conv_body_down[i](feat)
|
| 294 |
+
unet_skips.insert(0, feat)
|
| 295 |
+
feat = F.leaky_relu_(self.final_conv(feat), negative_slope=0.2)
|
| 296 |
+
|
| 297 |
+
# style code
|
| 298 |
+
style_code = self.final_linear(feat.view(feat.size(0), -1))
|
| 299 |
+
if self.different_w:
|
| 300 |
+
style_code = style_code.view(style_code.size(0), -1, self.num_style_feat)
|
| 301 |
+
|
| 302 |
+
# decode
|
| 303 |
+
for i in range(self.log_size - 2):
|
| 304 |
+
# add unet skip
|
| 305 |
+
feat = feat + unet_skips[i]
|
| 306 |
+
# ResUpLayer
|
| 307 |
+
feat = self.conv_body_up[i](feat)
|
| 308 |
+
# generate scale and shift for SFT layers
|
| 309 |
+
scale = self.condition_scale[i](feat)
|
| 310 |
+
conditions.append(scale.clone())
|
| 311 |
+
shift = self.condition_shift[i](feat)
|
| 312 |
+
conditions.append(shift.clone())
|
| 313 |
+
# generate rgb images
|
| 314 |
+
if return_rgb:
|
| 315 |
+
out_rgbs.append(self.toRGB[i](feat))
|
| 316 |
+
|
| 317 |
+
# decoder
|
| 318 |
+
image, _ = self.stylegan_decoder([style_code],
|
| 319 |
+
conditions,
|
| 320 |
+
return_latents=return_latents,
|
| 321 |
+
input_is_latent=self.input_is_latent,
|
| 322 |
+
randomize_noise=randomize_noise)
|
| 323 |
+
|
| 324 |
+
return image, out_rgbs
|
gfpgan/archs/restoreformer_arch.py
ADDED
|
@@ -0,0 +1,658 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""Modified from https://github.com/wzhouxiff/RestoreFormer
|
| 2 |
+
"""
|
| 3 |
+
import numpy as np
|
| 4 |
+
import torch
|
| 5 |
+
import torch.nn as nn
|
| 6 |
+
import torch.nn.functional as F
|
| 7 |
+
|
| 8 |
+
|
| 9 |
+
class VectorQuantizer(nn.Module):
|
| 10 |
+
"""
|
| 11 |
+
see https://github.com/MishaLaskin/vqvae/blob/d761a999e2267766400dc646d82d3ac3657771d4/models/quantizer.py
|
| 12 |
+
____________________________________________
|
| 13 |
+
Discretization bottleneck part of the VQ-VAE.
|
| 14 |
+
Inputs:
|
| 15 |
+
- n_e : number of embeddings
|
| 16 |
+
- e_dim : dimension of embedding
|
| 17 |
+
- beta : commitment cost used in loss term, beta * ||z_e(x)-sg[e]||^2
|
| 18 |
+
_____________________________________________
|
| 19 |
+
"""
|
| 20 |
+
|
| 21 |
+
def __init__(self, n_e, e_dim, beta):
|
| 22 |
+
super(VectorQuantizer, self).__init__()
|
| 23 |
+
self.n_e = n_e
|
| 24 |
+
self.e_dim = e_dim
|
| 25 |
+
self.beta = beta
|
| 26 |
+
|
| 27 |
+
self.embedding = nn.Embedding(self.n_e, self.e_dim)
|
| 28 |
+
self.embedding.weight.data.uniform_(-1.0 / self.n_e, 1.0 / self.n_e)
|
| 29 |
+
|
| 30 |
+
def forward(self, z):
|
| 31 |
+
"""
|
| 32 |
+
Inputs the output of the encoder network z and maps it to a discrete
|
| 33 |
+
one-hot vector that is the index of the closest embedding vector e_j
|
| 34 |
+
z (continuous) -> z_q (discrete)
|
| 35 |
+
z.shape = (batch, channel, height, width)
|
| 36 |
+
quantization pipeline:
|
| 37 |
+
1. get encoder input (B,C,H,W)
|
| 38 |
+
2. flatten input to (B*H*W,C)
|
| 39 |
+
"""
|
| 40 |
+
# reshape z -> (batch, height, width, channel) and flatten
|
| 41 |
+
z = z.permute(0, 2, 3, 1).contiguous()
|
| 42 |
+
z_flattened = z.view(-1, self.e_dim)
|
| 43 |
+
# distances from z to embeddings e_j (z - e)^2 = z^2 + e^2 - 2 e * z
|
| 44 |
+
|
| 45 |
+
d = torch.sum(z_flattened ** 2, dim=1, keepdim=True) + \
|
| 46 |
+
torch.sum(self.embedding.weight**2, dim=1) - 2 * \
|
| 47 |
+
torch.matmul(z_flattened, self.embedding.weight.t())
|
| 48 |
+
|
| 49 |
+
# could possible replace this here
|
| 50 |
+
# #\start...
|
| 51 |
+
# find closest encodings
|
| 52 |
+
|
| 53 |
+
min_value, min_encoding_indices = torch.min(d, dim=1)
|
| 54 |
+
|
| 55 |
+
min_encoding_indices = min_encoding_indices.unsqueeze(1)
|
| 56 |
+
|
| 57 |
+
min_encodings = torch.zeros(min_encoding_indices.shape[0], self.n_e).to(z)
|
| 58 |
+
min_encodings.scatter_(1, min_encoding_indices, 1)
|
| 59 |
+
|
| 60 |
+
# dtype min encodings: torch.float32
|
| 61 |
+
# min_encodings shape: torch.Size([2048, 512])
|
| 62 |
+
# min_encoding_indices.shape: torch.Size([2048, 1])
|
| 63 |
+
|
| 64 |
+
# get quantized latent vectors
|
| 65 |
+
z_q = torch.matmul(min_encodings, self.embedding.weight).view(z.shape)
|
| 66 |
+
# .........\end
|
| 67 |
+
|
| 68 |
+
# with:
|
| 69 |
+
# .........\start
|
| 70 |
+
# min_encoding_indices = torch.argmin(d, dim=1)
|
| 71 |
+
# z_q = self.embedding(min_encoding_indices)
|
| 72 |
+
# ......\end......... (TODO)
|
| 73 |
+
|
| 74 |
+
# compute loss for embedding
|
| 75 |
+
loss = torch.mean((z_q.detach() - z)**2) + self.beta * torch.mean((z_q - z.detach())**2)
|
| 76 |
+
|
| 77 |
+
# preserve gradients
|
| 78 |
+
z_q = z + (z_q - z).detach()
|
| 79 |
+
|
| 80 |
+
# perplexity
|
| 81 |
+
|
| 82 |
+
e_mean = torch.mean(min_encodings, dim=0)
|
| 83 |
+
perplexity = torch.exp(-torch.sum(e_mean * torch.log(e_mean + 1e-10)))
|
| 84 |
+
|
| 85 |
+
# reshape back to match original input shape
|
| 86 |
+
z_q = z_q.permute(0, 3, 1, 2).contiguous()
|
| 87 |
+
|
| 88 |
+
return z_q, loss, (perplexity, min_encodings, min_encoding_indices, d)
|
| 89 |
+
|
| 90 |
+
def get_codebook_entry(self, indices, shape):
|
| 91 |
+
# shape specifying (batch, height, width, channel)
|
| 92 |
+
# TODO: check for more easy handling with nn.Embedding
|
| 93 |
+
min_encodings = torch.zeros(indices.shape[0], self.n_e).to(indices)
|
| 94 |
+
min_encodings.scatter_(1, indices[:, None], 1)
|
| 95 |
+
|
| 96 |
+
# get quantized latent vectors
|
| 97 |
+
z_q = torch.matmul(min_encodings.float(), self.embedding.weight)
|
| 98 |
+
|
| 99 |
+
if shape is not None:
|
| 100 |
+
z_q = z_q.view(shape)
|
| 101 |
+
|
| 102 |
+
# reshape back to match original input shape
|
| 103 |
+
z_q = z_q.permute(0, 3, 1, 2).contiguous()
|
| 104 |
+
|
| 105 |
+
return z_q
|
| 106 |
+
|
| 107 |
+
|
| 108 |
+
# pytorch_diffusion + derived encoder decoder
|
| 109 |
+
def nonlinearity(x):
|
| 110 |
+
# swish
|
| 111 |
+
return x * torch.sigmoid(x)
|
| 112 |
+
|
| 113 |
+
|
| 114 |
+
def Normalize(in_channels):
|
| 115 |
+
return torch.nn.GroupNorm(num_groups=32, num_channels=in_channels, eps=1e-6, affine=True)
|
| 116 |
+
|
| 117 |
+
|
| 118 |
+
class Upsample(nn.Module):
|
| 119 |
+
|
| 120 |
+
def __init__(self, in_channels, with_conv):
|
| 121 |
+
super().__init__()
|
| 122 |
+
self.with_conv = with_conv
|
| 123 |
+
if self.with_conv:
|
| 124 |
+
self.conv = torch.nn.Conv2d(in_channels, in_channels, kernel_size=3, stride=1, padding=1)
|
| 125 |
+
|
| 126 |
+
def forward(self, x):
|
| 127 |
+
x = torch.nn.functional.interpolate(x, scale_factor=2.0, mode='nearest')
|
| 128 |
+
if self.with_conv:
|
| 129 |
+
x = self.conv(x)
|
| 130 |
+
return x
|
| 131 |
+
|
| 132 |
+
|
| 133 |
+
class Downsample(nn.Module):
|
| 134 |
+
|
| 135 |
+
def __init__(self, in_channels, with_conv):
|
| 136 |
+
super().__init__()
|
| 137 |
+
self.with_conv = with_conv
|
| 138 |
+
if self.with_conv:
|
| 139 |
+
# no asymmetric padding in torch conv, must do it ourselves
|
| 140 |
+
self.conv = torch.nn.Conv2d(in_channels, in_channels, kernel_size=3, stride=2, padding=0)
|
| 141 |
+
|
| 142 |
+
def forward(self, x):
|
| 143 |
+
if self.with_conv:
|
| 144 |
+
pad = (0, 1, 0, 1)
|
| 145 |
+
x = torch.nn.functional.pad(x, pad, mode='constant', value=0)
|
| 146 |
+
x = self.conv(x)
|
| 147 |
+
else:
|
| 148 |
+
x = torch.nn.functional.avg_pool2d(x, kernel_size=2, stride=2)
|
| 149 |
+
return x
|
| 150 |
+
|
| 151 |
+
|
| 152 |
+
class ResnetBlock(nn.Module):
|
| 153 |
+
|
| 154 |
+
def __init__(self, *, in_channels, out_channels=None, conv_shortcut=False, dropout, temb_channels=512):
|
| 155 |
+
super().__init__()
|
| 156 |
+
self.in_channels = in_channels
|
| 157 |
+
out_channels = in_channels if out_channels is None else out_channels
|
| 158 |
+
self.out_channels = out_channels
|
| 159 |
+
self.use_conv_shortcut = conv_shortcut
|
| 160 |
+
|
| 161 |
+
self.norm1 = Normalize(in_channels)
|
| 162 |
+
self.conv1 = torch.nn.Conv2d(in_channels, out_channels, kernel_size=3, stride=1, padding=1)
|
| 163 |
+
if temb_channels > 0:
|
| 164 |
+
self.temb_proj = torch.nn.Linear(temb_channels, out_channels)
|
| 165 |
+
self.norm2 = Normalize(out_channels)
|
| 166 |
+
self.dropout = torch.nn.Dropout(dropout)
|
| 167 |
+
self.conv2 = torch.nn.Conv2d(out_channels, out_channels, kernel_size=3, stride=1, padding=1)
|
| 168 |
+
if self.in_channels != self.out_channels:
|
| 169 |
+
if self.use_conv_shortcut:
|
| 170 |
+
self.conv_shortcut = torch.nn.Conv2d(in_channels, out_channels, kernel_size=3, stride=1, padding=1)
|
| 171 |
+
else:
|
| 172 |
+
self.nin_shortcut = torch.nn.Conv2d(in_channels, out_channels, kernel_size=1, stride=1, padding=0)
|
| 173 |
+
|
| 174 |
+
def forward(self, x, temb):
|
| 175 |
+
h = x
|
| 176 |
+
h = self.norm1(h)
|
| 177 |
+
h = nonlinearity(h)
|
| 178 |
+
h = self.conv1(h)
|
| 179 |
+
|
| 180 |
+
if temb is not None:
|
| 181 |
+
h = h + self.temb_proj(nonlinearity(temb))[:, :, None, None]
|
| 182 |
+
|
| 183 |
+
h = self.norm2(h)
|
| 184 |
+
h = nonlinearity(h)
|
| 185 |
+
h = self.dropout(h)
|
| 186 |
+
h = self.conv2(h)
|
| 187 |
+
|
| 188 |
+
if self.in_channels != self.out_channels:
|
| 189 |
+
if self.use_conv_shortcut:
|
| 190 |
+
x = self.conv_shortcut(x)
|
| 191 |
+
else:
|
| 192 |
+
x = self.nin_shortcut(x)
|
| 193 |
+
|
| 194 |
+
return x + h
|
| 195 |
+
|
| 196 |
+
|
| 197 |
+
class MultiHeadAttnBlock(nn.Module):
|
| 198 |
+
|
| 199 |
+
def __init__(self, in_channels, head_size=1):
|
| 200 |
+
super().__init__()
|
| 201 |
+
self.in_channels = in_channels
|
| 202 |
+
self.head_size = head_size
|
| 203 |
+
self.att_size = in_channels // head_size
|
| 204 |
+
assert (in_channels % head_size == 0), 'The size of head should be divided by the number of channels.'
|
| 205 |
+
|
| 206 |
+
self.norm1 = Normalize(in_channels)
|
| 207 |
+
self.norm2 = Normalize(in_channels)
|
| 208 |
+
|
| 209 |
+
self.q = torch.nn.Conv2d(in_channels, in_channels, kernel_size=1, stride=1, padding=0)
|
| 210 |
+
self.k = torch.nn.Conv2d(in_channels, in_channels, kernel_size=1, stride=1, padding=0)
|
| 211 |
+
self.v = torch.nn.Conv2d(in_channels, in_channels, kernel_size=1, stride=1, padding=0)
|
| 212 |
+
self.proj_out = torch.nn.Conv2d(in_channels, in_channels, kernel_size=1, stride=1, padding=0)
|
| 213 |
+
self.num = 0
|
| 214 |
+
|
| 215 |
+
def forward(self, x, y=None):
|
| 216 |
+
h_ = x
|
| 217 |
+
h_ = self.norm1(h_)
|
| 218 |
+
if y is None:
|
| 219 |
+
y = h_
|
| 220 |
+
else:
|
| 221 |
+
y = self.norm2(y)
|
| 222 |
+
|
| 223 |
+
q = self.q(y)
|
| 224 |
+
k = self.k(h_)
|
| 225 |
+
v = self.v(h_)
|
| 226 |
+
|
| 227 |
+
# compute attention
|
| 228 |
+
b, c, h, w = q.shape
|
| 229 |
+
q = q.reshape(b, self.head_size, self.att_size, h * w)
|
| 230 |
+
q = q.permute(0, 3, 1, 2) # b, hw, head, att
|
| 231 |
+
|
| 232 |
+
k = k.reshape(b, self.head_size, self.att_size, h * w)
|
| 233 |
+
k = k.permute(0, 3, 1, 2)
|
| 234 |
+
|
| 235 |
+
v = v.reshape(b, self.head_size, self.att_size, h * w)
|
| 236 |
+
v = v.permute(0, 3, 1, 2)
|
| 237 |
+
|
| 238 |
+
q = q.transpose(1, 2)
|
| 239 |
+
v = v.transpose(1, 2)
|
| 240 |
+
k = k.transpose(1, 2).transpose(2, 3)
|
| 241 |
+
|
| 242 |
+
scale = int(self.att_size)**(-0.5)
|
| 243 |
+
q.mul_(scale)
|
| 244 |
+
w_ = torch.matmul(q, k)
|
| 245 |
+
w_ = F.softmax(w_, dim=3)
|
| 246 |
+
|
| 247 |
+
w_ = w_.matmul(v)
|
| 248 |
+
|
| 249 |
+
w_ = w_.transpose(1, 2).contiguous() # [b, h*w, head, att]
|
| 250 |
+
w_ = w_.view(b, h, w, -1)
|
| 251 |
+
w_ = w_.permute(0, 3, 1, 2)
|
| 252 |
+
|
| 253 |
+
w_ = self.proj_out(w_)
|
| 254 |
+
|
| 255 |
+
return x + w_
|
| 256 |
+
|
| 257 |
+
|
| 258 |
+
class MultiHeadEncoder(nn.Module):
|
| 259 |
+
|
| 260 |
+
def __init__(self,
|
| 261 |
+
ch,
|
| 262 |
+
out_ch,
|
| 263 |
+
ch_mult=(1, 2, 4, 8),
|
| 264 |
+
num_res_blocks=2,
|
| 265 |
+
attn_resolutions=(16, ),
|
| 266 |
+
dropout=0.0,
|
| 267 |
+
resamp_with_conv=True,
|
| 268 |
+
in_channels=3,
|
| 269 |
+
resolution=512,
|
| 270 |
+
z_channels=256,
|
| 271 |
+
double_z=True,
|
| 272 |
+
enable_mid=True,
|
| 273 |
+
head_size=1,
|
| 274 |
+
**ignore_kwargs):
|
| 275 |
+
super().__init__()
|
| 276 |
+
self.ch = ch
|
| 277 |
+
self.temb_ch = 0
|
| 278 |
+
self.num_resolutions = len(ch_mult)
|
| 279 |
+
self.num_res_blocks = num_res_blocks
|
| 280 |
+
self.resolution = resolution
|
| 281 |
+
self.in_channels = in_channels
|
| 282 |
+
self.enable_mid = enable_mid
|
| 283 |
+
|
| 284 |
+
# downsampling
|
| 285 |
+
self.conv_in = torch.nn.Conv2d(in_channels, self.ch, kernel_size=3, stride=1, padding=1)
|
| 286 |
+
|
| 287 |
+
curr_res = resolution
|
| 288 |
+
in_ch_mult = (1, ) + tuple(ch_mult)
|
| 289 |
+
self.down = nn.ModuleList()
|
| 290 |
+
for i_level in range(self.num_resolutions):
|
| 291 |
+
block = nn.ModuleList()
|
| 292 |
+
attn = nn.ModuleList()
|
| 293 |
+
block_in = ch * in_ch_mult[i_level]
|
| 294 |
+
block_out = ch * ch_mult[i_level]
|
| 295 |
+
for i_block in range(self.num_res_blocks):
|
| 296 |
+
block.append(
|
| 297 |
+
ResnetBlock(
|
| 298 |
+
in_channels=block_in, out_channels=block_out, temb_channels=self.temb_ch, dropout=dropout))
|
| 299 |
+
block_in = block_out
|
| 300 |
+
if curr_res in attn_resolutions:
|
| 301 |
+
attn.append(MultiHeadAttnBlock(block_in, head_size))
|
| 302 |
+
down = nn.Module()
|
| 303 |
+
down.block = block
|
| 304 |
+
down.attn = attn
|
| 305 |
+
if i_level != self.num_resolutions - 1:
|
| 306 |
+
down.downsample = Downsample(block_in, resamp_with_conv)
|
| 307 |
+
curr_res = curr_res // 2
|
| 308 |
+
self.down.append(down)
|
| 309 |
+
|
| 310 |
+
# middle
|
| 311 |
+
if self.enable_mid:
|
| 312 |
+
self.mid = nn.Module()
|
| 313 |
+
self.mid.block_1 = ResnetBlock(
|
| 314 |
+
in_channels=block_in, out_channels=block_in, temb_channels=self.temb_ch, dropout=dropout)
|
| 315 |
+
self.mid.attn_1 = MultiHeadAttnBlock(block_in, head_size)
|
| 316 |
+
self.mid.block_2 = ResnetBlock(
|
| 317 |
+
in_channels=block_in, out_channels=block_in, temb_channels=self.temb_ch, dropout=dropout)
|
| 318 |
+
|
| 319 |
+
# end
|
| 320 |
+
self.norm_out = Normalize(block_in)
|
| 321 |
+
self.conv_out = torch.nn.Conv2d(
|
| 322 |
+
block_in, 2 * z_channels if double_z else z_channels, kernel_size=3, stride=1, padding=1)
|
| 323 |
+
|
| 324 |
+
def forward(self, x):
|
| 325 |
+
hs = {}
|
| 326 |
+
# timestep embedding
|
| 327 |
+
temb = None
|
| 328 |
+
|
| 329 |
+
# downsampling
|
| 330 |
+
h = self.conv_in(x)
|
| 331 |
+
hs['in'] = h
|
| 332 |
+
for i_level in range(self.num_resolutions):
|
| 333 |
+
for i_block in range(self.num_res_blocks):
|
| 334 |
+
h = self.down[i_level].block[i_block](h, temb)
|
| 335 |
+
if len(self.down[i_level].attn) > 0:
|
| 336 |
+
h = self.down[i_level].attn[i_block](h)
|
| 337 |
+
|
| 338 |
+
if i_level != self.num_resolutions - 1:
|
| 339 |
+
# hs.append(h)
|
| 340 |
+
hs['block_' + str(i_level)] = h
|
| 341 |
+
h = self.down[i_level].downsample(h)
|
| 342 |
+
|
| 343 |
+
# middle
|
| 344 |
+
# h = hs[-1]
|
| 345 |
+
if self.enable_mid:
|
| 346 |
+
h = self.mid.block_1(h, temb)
|
| 347 |
+
hs['block_' + str(i_level) + '_atten'] = h
|
| 348 |
+
h = self.mid.attn_1(h)
|
| 349 |
+
h = self.mid.block_2(h, temb)
|
| 350 |
+
hs['mid_atten'] = h
|
| 351 |
+
|
| 352 |
+
# end
|
| 353 |
+
h = self.norm_out(h)
|
| 354 |
+
h = nonlinearity(h)
|
| 355 |
+
h = self.conv_out(h)
|
| 356 |
+
# hs.append(h)
|
| 357 |
+
hs['out'] = h
|
| 358 |
+
|
| 359 |
+
return hs
|
| 360 |
+
|
| 361 |
+
|
| 362 |
+
class MultiHeadDecoder(nn.Module):
|
| 363 |
+
|
| 364 |
+
def __init__(self,
|
| 365 |
+
ch,
|
| 366 |
+
out_ch,
|
| 367 |
+
ch_mult=(1, 2, 4, 8),
|
| 368 |
+
num_res_blocks=2,
|
| 369 |
+
attn_resolutions=(16, ),
|
| 370 |
+
dropout=0.0,
|
| 371 |
+
resamp_with_conv=True,
|
| 372 |
+
in_channels=3,
|
| 373 |
+
resolution=512,
|
| 374 |
+
z_channels=256,
|
| 375 |
+
give_pre_end=False,
|
| 376 |
+
enable_mid=True,
|
| 377 |
+
head_size=1,
|
| 378 |
+
**ignorekwargs):
|
| 379 |
+
super().__init__()
|
| 380 |
+
self.ch = ch
|
| 381 |
+
self.temb_ch = 0
|
| 382 |
+
self.num_resolutions = len(ch_mult)
|
| 383 |
+
self.num_res_blocks = num_res_blocks
|
| 384 |
+
self.resolution = resolution
|
| 385 |
+
self.in_channels = in_channels
|
| 386 |
+
self.give_pre_end = give_pre_end
|
| 387 |
+
self.enable_mid = enable_mid
|
| 388 |
+
|
| 389 |
+
# compute in_ch_mult, block_in and curr_res at lowest res
|
| 390 |
+
block_in = ch * ch_mult[self.num_resolutions - 1]
|
| 391 |
+
curr_res = resolution // 2**(self.num_resolutions - 1)
|
| 392 |
+
self.z_shape = (1, z_channels, curr_res, curr_res)
|
| 393 |
+
print('Working with z of shape {} = {} dimensions.'.format(self.z_shape, np.prod(self.z_shape)))
|
| 394 |
+
|
| 395 |
+
# z to block_in
|
| 396 |
+
self.conv_in = torch.nn.Conv2d(z_channels, block_in, kernel_size=3, stride=1, padding=1)
|
| 397 |
+
|
| 398 |
+
# middle
|
| 399 |
+
if self.enable_mid:
|
| 400 |
+
self.mid = nn.Module()
|
| 401 |
+
self.mid.block_1 = ResnetBlock(
|
| 402 |
+
in_channels=block_in, out_channels=block_in, temb_channels=self.temb_ch, dropout=dropout)
|
| 403 |
+
self.mid.attn_1 = MultiHeadAttnBlock(block_in, head_size)
|
| 404 |
+
self.mid.block_2 = ResnetBlock(
|
| 405 |
+
in_channels=block_in, out_channels=block_in, temb_channels=self.temb_ch, dropout=dropout)
|
| 406 |
+
|
| 407 |
+
# upsampling
|
| 408 |
+
self.up = nn.ModuleList()
|
| 409 |
+
for i_level in reversed(range(self.num_resolutions)):
|
| 410 |
+
block = nn.ModuleList()
|
| 411 |
+
attn = nn.ModuleList()
|
| 412 |
+
block_out = ch * ch_mult[i_level]
|
| 413 |
+
for i_block in range(self.num_res_blocks + 1):
|
| 414 |
+
block.append(
|
| 415 |
+
ResnetBlock(
|
| 416 |
+
in_channels=block_in, out_channels=block_out, temb_channels=self.temb_ch, dropout=dropout))
|
| 417 |
+
block_in = block_out
|
| 418 |
+
if curr_res in attn_resolutions:
|
| 419 |
+
attn.append(MultiHeadAttnBlock(block_in, head_size))
|
| 420 |
+
up = nn.Module()
|
| 421 |
+
up.block = block
|
| 422 |
+
up.attn = attn
|
| 423 |
+
if i_level != 0:
|
| 424 |
+
up.upsample = Upsample(block_in, resamp_with_conv)
|
| 425 |
+
curr_res = curr_res * 2
|
| 426 |
+
self.up.insert(0, up) # prepend to get consistent order
|
| 427 |
+
|
| 428 |
+
# end
|
| 429 |
+
self.norm_out = Normalize(block_in)
|
| 430 |
+
self.conv_out = torch.nn.Conv2d(block_in, out_ch, kernel_size=3, stride=1, padding=1)
|
| 431 |
+
|
| 432 |
+
def forward(self, z):
|
| 433 |
+
# assert z.shape[1:] == self.z_shape[1:]
|
| 434 |
+
self.last_z_shape = z.shape
|
| 435 |
+
|
| 436 |
+
# timestep embedding
|
| 437 |
+
temb = None
|
| 438 |
+
|
| 439 |
+
# z to block_in
|
| 440 |
+
h = self.conv_in(z)
|
| 441 |
+
|
| 442 |
+
# middle
|
| 443 |
+
if self.enable_mid:
|
| 444 |
+
h = self.mid.block_1(h, temb)
|
| 445 |
+
h = self.mid.attn_1(h)
|
| 446 |
+
h = self.mid.block_2(h, temb)
|
| 447 |
+
|
| 448 |
+
# upsampling
|
| 449 |
+
for i_level in reversed(range(self.num_resolutions)):
|
| 450 |
+
for i_block in range(self.num_res_blocks + 1):
|
| 451 |
+
h = self.up[i_level].block[i_block](h, temb)
|
| 452 |
+
if len(self.up[i_level].attn) > 0:
|
| 453 |
+
h = self.up[i_level].attn[i_block](h)
|
| 454 |
+
if i_level != 0:
|
| 455 |
+
h = self.up[i_level].upsample(h)
|
| 456 |
+
|
| 457 |
+
# end
|
| 458 |
+
if self.give_pre_end:
|
| 459 |
+
return h
|
| 460 |
+
|
| 461 |
+
h = self.norm_out(h)
|
| 462 |
+
h = nonlinearity(h)
|
| 463 |
+
h = self.conv_out(h)
|
| 464 |
+
return h
|
| 465 |
+
|
| 466 |
+
|
| 467 |
+
class MultiHeadDecoderTransformer(nn.Module):
|
| 468 |
+
|
| 469 |
+
def __init__(self,
|
| 470 |
+
ch,
|
| 471 |
+
out_ch,
|
| 472 |
+
ch_mult=(1, 2, 4, 8),
|
| 473 |
+
num_res_blocks=2,
|
| 474 |
+
attn_resolutions=(16, ),
|
| 475 |
+
dropout=0.0,
|
| 476 |
+
resamp_with_conv=True,
|
| 477 |
+
in_channels=3,
|
| 478 |
+
resolution=512,
|
| 479 |
+
z_channels=256,
|
| 480 |
+
give_pre_end=False,
|
| 481 |
+
enable_mid=True,
|
| 482 |
+
head_size=1,
|
| 483 |
+
**ignorekwargs):
|
| 484 |
+
super().__init__()
|
| 485 |
+
self.ch = ch
|
| 486 |
+
self.temb_ch = 0
|
| 487 |
+
self.num_resolutions = len(ch_mult)
|
| 488 |
+
self.num_res_blocks = num_res_blocks
|
| 489 |
+
self.resolution = resolution
|
| 490 |
+
self.in_channels = in_channels
|
| 491 |
+
self.give_pre_end = give_pre_end
|
| 492 |
+
self.enable_mid = enable_mid
|
| 493 |
+
|
| 494 |
+
# compute in_ch_mult, block_in and curr_res at lowest res
|
| 495 |
+
block_in = ch * ch_mult[self.num_resolutions - 1]
|
| 496 |
+
curr_res = resolution // 2**(self.num_resolutions - 1)
|
| 497 |
+
self.z_shape = (1, z_channels, curr_res, curr_res)
|
| 498 |
+
print('Working with z of shape {} = {} dimensions.'.format(self.z_shape, np.prod(self.z_shape)))
|
| 499 |
+
|
| 500 |
+
# z to block_in
|
| 501 |
+
self.conv_in = torch.nn.Conv2d(z_channels, block_in, kernel_size=3, stride=1, padding=1)
|
| 502 |
+
|
| 503 |
+
# middle
|
| 504 |
+
if self.enable_mid:
|
| 505 |
+
self.mid = nn.Module()
|
| 506 |
+
self.mid.block_1 = ResnetBlock(
|
| 507 |
+
in_channels=block_in, out_channels=block_in, temb_channels=self.temb_ch, dropout=dropout)
|
| 508 |
+
self.mid.attn_1 = MultiHeadAttnBlock(block_in, head_size)
|
| 509 |
+
self.mid.block_2 = ResnetBlock(
|
| 510 |
+
in_channels=block_in, out_channels=block_in, temb_channels=self.temb_ch, dropout=dropout)
|
| 511 |
+
|
| 512 |
+
# upsampling
|
| 513 |
+
self.up = nn.ModuleList()
|
| 514 |
+
for i_level in reversed(range(self.num_resolutions)):
|
| 515 |
+
block = nn.ModuleList()
|
| 516 |
+
attn = nn.ModuleList()
|
| 517 |
+
block_out = ch * ch_mult[i_level]
|
| 518 |
+
for i_block in range(self.num_res_blocks + 1):
|
| 519 |
+
block.append(
|
| 520 |
+
ResnetBlock(
|
| 521 |
+
in_channels=block_in, out_channels=block_out, temb_channels=self.temb_ch, dropout=dropout))
|
| 522 |
+
block_in = block_out
|
| 523 |
+
if curr_res in attn_resolutions:
|
| 524 |
+
attn.append(MultiHeadAttnBlock(block_in, head_size))
|
| 525 |
+
up = nn.Module()
|
| 526 |
+
up.block = block
|
| 527 |
+
up.attn = attn
|
| 528 |
+
if i_level != 0:
|
| 529 |
+
up.upsample = Upsample(block_in, resamp_with_conv)
|
| 530 |
+
curr_res = curr_res * 2
|
| 531 |
+
self.up.insert(0, up) # prepend to get consistent order
|
| 532 |
+
|
| 533 |
+
# end
|
| 534 |
+
self.norm_out = Normalize(block_in)
|
| 535 |
+
self.conv_out = torch.nn.Conv2d(block_in, out_ch, kernel_size=3, stride=1, padding=1)
|
| 536 |
+
|
| 537 |
+
def forward(self, z, hs):
|
| 538 |
+
# assert z.shape[1:] == self.z_shape[1:]
|
| 539 |
+
# self.last_z_shape = z.shape
|
| 540 |
+
|
| 541 |
+
# timestep embedding
|
| 542 |
+
temb = None
|
| 543 |
+
|
| 544 |
+
# z to block_in
|
| 545 |
+
h = self.conv_in(z)
|
| 546 |
+
|
| 547 |
+
# middle
|
| 548 |
+
if self.enable_mid:
|
| 549 |
+
h = self.mid.block_1(h, temb)
|
| 550 |
+
h = self.mid.attn_1(h, hs['mid_atten'])
|
| 551 |
+
h = self.mid.block_2(h, temb)
|
| 552 |
+
|
| 553 |
+
# upsampling
|
| 554 |
+
for i_level in reversed(range(self.num_resolutions)):
|
| 555 |
+
for i_block in range(self.num_res_blocks + 1):
|
| 556 |
+
h = self.up[i_level].block[i_block](h, temb)
|
| 557 |
+
if len(self.up[i_level].attn) > 0:
|
| 558 |
+
h = self.up[i_level].attn[i_block](h, hs['block_' + str(i_level) + '_atten'])
|
| 559 |
+
# hfeature = h.clone()
|
| 560 |
+
if i_level != 0:
|
| 561 |
+
h = self.up[i_level].upsample(h)
|
| 562 |
+
|
| 563 |
+
# end
|
| 564 |
+
if self.give_pre_end:
|
| 565 |
+
return h
|
| 566 |
+
|
| 567 |
+
h = self.norm_out(h)
|
| 568 |
+
h = nonlinearity(h)
|
| 569 |
+
h = self.conv_out(h)
|
| 570 |
+
return h
|
| 571 |
+
|
| 572 |
+
|
| 573 |
+
class RestoreFormer(nn.Module):
|
| 574 |
+
|
| 575 |
+
def __init__(self,
|
| 576 |
+
n_embed=1024,
|
| 577 |
+
embed_dim=256,
|
| 578 |
+
ch=64,
|
| 579 |
+
out_ch=3,
|
| 580 |
+
ch_mult=(1, 2, 2, 4, 4, 8),
|
| 581 |
+
num_res_blocks=2,
|
| 582 |
+
attn_resolutions=(16, ),
|
| 583 |
+
dropout=0.0,
|
| 584 |
+
in_channels=3,
|
| 585 |
+
resolution=512,
|
| 586 |
+
z_channels=256,
|
| 587 |
+
double_z=False,
|
| 588 |
+
enable_mid=True,
|
| 589 |
+
fix_decoder=False,
|
| 590 |
+
fix_codebook=True,
|
| 591 |
+
fix_encoder=False,
|
| 592 |
+
head_size=8):
|
| 593 |
+
super(RestoreFormer, self).__init__()
|
| 594 |
+
|
| 595 |
+
self.encoder = MultiHeadEncoder(
|
| 596 |
+
ch=ch,
|
| 597 |
+
out_ch=out_ch,
|
| 598 |
+
ch_mult=ch_mult,
|
| 599 |
+
num_res_blocks=num_res_blocks,
|
| 600 |
+
attn_resolutions=attn_resolutions,
|
| 601 |
+
dropout=dropout,
|
| 602 |
+
in_channels=in_channels,
|
| 603 |
+
resolution=resolution,
|
| 604 |
+
z_channels=z_channels,
|
| 605 |
+
double_z=double_z,
|
| 606 |
+
enable_mid=enable_mid,
|
| 607 |
+
head_size=head_size)
|
| 608 |
+
self.decoder = MultiHeadDecoderTransformer(
|
| 609 |
+
ch=ch,
|
| 610 |
+
out_ch=out_ch,
|
| 611 |
+
ch_mult=ch_mult,
|
| 612 |
+
num_res_blocks=num_res_blocks,
|
| 613 |
+
attn_resolutions=attn_resolutions,
|
| 614 |
+
dropout=dropout,
|
| 615 |
+
in_channels=in_channels,
|
| 616 |
+
resolution=resolution,
|
| 617 |
+
z_channels=z_channels,
|
| 618 |
+
enable_mid=enable_mid,
|
| 619 |
+
head_size=head_size)
|
| 620 |
+
|
| 621 |
+
self.quantize = VectorQuantizer(n_embed, embed_dim, beta=0.25)
|
| 622 |
+
|
| 623 |
+
self.quant_conv = torch.nn.Conv2d(z_channels, embed_dim, 1)
|
| 624 |
+
self.post_quant_conv = torch.nn.Conv2d(embed_dim, z_channels, 1)
|
| 625 |
+
|
| 626 |
+
if fix_decoder:
|
| 627 |
+
for _, param in self.decoder.named_parameters():
|
| 628 |
+
param.requires_grad = False
|
| 629 |
+
for _, param in self.post_quant_conv.named_parameters():
|
| 630 |
+
param.requires_grad = False
|
| 631 |
+
for _, param in self.quantize.named_parameters():
|
| 632 |
+
param.requires_grad = False
|
| 633 |
+
elif fix_codebook:
|
| 634 |
+
for _, param in self.quantize.named_parameters():
|
| 635 |
+
param.requires_grad = False
|
| 636 |
+
|
| 637 |
+
if fix_encoder:
|
| 638 |
+
for _, param in self.encoder.named_parameters():
|
| 639 |
+
param.requires_grad = False
|
| 640 |
+
|
| 641 |
+
def encode(self, x):
|
| 642 |
+
|
| 643 |
+
hs = self.encoder(x)
|
| 644 |
+
h = self.quant_conv(hs['out'])
|
| 645 |
+
quant, emb_loss, info = self.quantize(h)
|
| 646 |
+
return quant, emb_loss, info, hs
|
| 647 |
+
|
| 648 |
+
def decode(self, quant, hs):
|
| 649 |
+
quant = self.post_quant_conv(quant)
|
| 650 |
+
dec = self.decoder(quant, hs)
|
| 651 |
+
|
| 652 |
+
return dec
|
| 653 |
+
|
| 654 |
+
def forward(self, input, **kwargs):
|
| 655 |
+
quant, diff, info, hs = self.encode(input)
|
| 656 |
+
dec = self.decode(quant, hs)
|
| 657 |
+
|
| 658 |
+
return dec, None
|
gfpgan/archs/stylegan2_bilinear_arch.py
ADDED
|
@@ -0,0 +1,613 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import math
|
| 2 |
+
import random
|
| 3 |
+
import torch
|
| 4 |
+
from basicsr.ops.fused_act import FusedLeakyReLU, fused_leaky_relu
|
| 5 |
+
from basicsr.utils.registry import ARCH_REGISTRY
|
| 6 |
+
from torch import nn
|
| 7 |
+
from torch.nn import functional as F
|
| 8 |
+
|
| 9 |
+
|
| 10 |
+
class NormStyleCode(nn.Module):
|
| 11 |
+
|
| 12 |
+
def forward(self, x):
|
| 13 |
+
"""Normalize the style codes.
|
| 14 |
+
|
| 15 |
+
Args:
|
| 16 |
+
x (Tensor): Style codes with shape (b, c).
|
| 17 |
+
|
| 18 |
+
Returns:
|
| 19 |
+
Tensor: Normalized tensor.
|
| 20 |
+
"""
|
| 21 |
+
return x * torch.rsqrt(torch.mean(x**2, dim=1, keepdim=True) + 1e-8)
|
| 22 |
+
|
| 23 |
+
|
| 24 |
+
class EqualLinear(nn.Module):
|
| 25 |
+
"""Equalized Linear as StyleGAN2.
|
| 26 |
+
|
| 27 |
+
Args:
|
| 28 |
+
in_channels (int): Size of each sample.
|
| 29 |
+
out_channels (int): Size of each output sample.
|
| 30 |
+
bias (bool): If set to ``False``, the layer will not learn an additive
|
| 31 |
+
bias. Default: ``True``.
|
| 32 |
+
bias_init_val (float): Bias initialized value. Default: 0.
|
| 33 |
+
lr_mul (float): Learning rate multiplier. Default: 1.
|
| 34 |
+
activation (None | str): The activation after ``linear`` operation.
|
| 35 |
+
Supported: 'fused_lrelu', None. Default: None.
|
| 36 |
+
"""
|
| 37 |
+
|
| 38 |
+
def __init__(self, in_channels, out_channels, bias=True, bias_init_val=0, lr_mul=1, activation=None):
|
| 39 |
+
super(EqualLinear, self).__init__()
|
| 40 |
+
self.in_channels = in_channels
|
| 41 |
+
self.out_channels = out_channels
|
| 42 |
+
self.lr_mul = lr_mul
|
| 43 |
+
self.activation = activation
|
| 44 |
+
if self.activation not in ['fused_lrelu', None]:
|
| 45 |
+
raise ValueError(f'Wrong activation value in EqualLinear: {activation}'
|
| 46 |
+
"Supported ones are: ['fused_lrelu', None].")
|
| 47 |
+
self.scale = (1 / math.sqrt(in_channels)) * lr_mul
|
| 48 |
+
|
| 49 |
+
self.weight = nn.Parameter(torch.randn(out_channels, in_channels).div_(lr_mul))
|
| 50 |
+
if bias:
|
| 51 |
+
self.bias = nn.Parameter(torch.zeros(out_channels).fill_(bias_init_val))
|
| 52 |
+
else:
|
| 53 |
+
self.register_parameter('bias', None)
|
| 54 |
+
|
| 55 |
+
def forward(self, x):
|
| 56 |
+
if self.bias is None:
|
| 57 |
+
bias = None
|
| 58 |
+
else:
|
| 59 |
+
bias = self.bias * self.lr_mul
|
| 60 |
+
if self.activation == 'fused_lrelu':
|
| 61 |
+
out = F.linear(x, self.weight * self.scale)
|
| 62 |
+
out = fused_leaky_relu(out, bias)
|
| 63 |
+
else:
|
| 64 |
+
out = F.linear(x, self.weight * self.scale, bias=bias)
|
| 65 |
+
return out
|
| 66 |
+
|
| 67 |
+
def __repr__(self):
|
| 68 |
+
return (f'{self.__class__.__name__}(in_channels={self.in_channels}, '
|
| 69 |
+
f'out_channels={self.out_channels}, bias={self.bias is not None})')
|
| 70 |
+
|
| 71 |
+
|
| 72 |
+
class ModulatedConv2d(nn.Module):
|
| 73 |
+
"""Modulated Conv2d used in StyleGAN2.
|
| 74 |
+
|
| 75 |
+
There is no bias in ModulatedConv2d.
|
| 76 |
+
|
| 77 |
+
Args:
|
| 78 |
+
in_channels (int): Channel number of the input.
|
| 79 |
+
out_channels (int): Channel number of the output.
|
| 80 |
+
kernel_size (int): Size of the convolving kernel.
|
| 81 |
+
num_style_feat (int): Channel number of style features.
|
| 82 |
+
demodulate (bool): Whether to demodulate in the conv layer.
|
| 83 |
+
Default: True.
|
| 84 |
+
sample_mode (str | None): Indicating 'upsample', 'downsample' or None.
|
| 85 |
+
Default: None.
|
| 86 |
+
eps (float): A value added to the denominator for numerical stability.
|
| 87 |
+
Default: 1e-8.
|
| 88 |
+
"""
|
| 89 |
+
|
| 90 |
+
def __init__(self,
|
| 91 |
+
in_channels,
|
| 92 |
+
out_channels,
|
| 93 |
+
kernel_size,
|
| 94 |
+
num_style_feat,
|
| 95 |
+
demodulate=True,
|
| 96 |
+
sample_mode=None,
|
| 97 |
+
eps=1e-8,
|
| 98 |
+
interpolation_mode='bilinear'):
|
| 99 |
+
super(ModulatedConv2d, self).__init__()
|
| 100 |
+
self.in_channels = in_channels
|
| 101 |
+
self.out_channels = out_channels
|
| 102 |
+
self.kernel_size = kernel_size
|
| 103 |
+
self.demodulate = demodulate
|
| 104 |
+
self.sample_mode = sample_mode
|
| 105 |
+
self.eps = eps
|
| 106 |
+
self.interpolation_mode = interpolation_mode
|
| 107 |
+
if self.interpolation_mode == 'nearest':
|
| 108 |
+
self.align_corners = None
|
| 109 |
+
else:
|
| 110 |
+
self.align_corners = False
|
| 111 |
+
|
| 112 |
+
self.scale = 1 / math.sqrt(in_channels * kernel_size**2)
|
| 113 |
+
# modulation inside each modulated conv
|
| 114 |
+
self.modulation = EqualLinear(
|
| 115 |
+
num_style_feat, in_channels, bias=True, bias_init_val=1, lr_mul=1, activation=None)
|
| 116 |
+
|
| 117 |
+
self.weight = nn.Parameter(torch.randn(1, out_channels, in_channels, kernel_size, kernel_size))
|
| 118 |
+
self.padding = kernel_size // 2
|
| 119 |
+
|
| 120 |
+
def forward(self, x, style):
|
| 121 |
+
"""Forward function.
|
| 122 |
+
|
| 123 |
+
Args:
|
| 124 |
+
x (Tensor): Tensor with shape (b, c, h, w).
|
| 125 |
+
style (Tensor): Tensor with shape (b, num_style_feat).
|
| 126 |
+
|
| 127 |
+
Returns:
|
| 128 |
+
Tensor: Modulated tensor after convolution.
|
| 129 |
+
"""
|
| 130 |
+
b, c, h, w = x.shape # c = c_in
|
| 131 |
+
# weight modulation
|
| 132 |
+
style = self.modulation(style).view(b, 1, c, 1, 1)
|
| 133 |
+
# self.weight: (1, c_out, c_in, k, k); style: (b, 1, c, 1, 1)
|
| 134 |
+
weight = self.scale * self.weight * style # (b, c_out, c_in, k, k)
|
| 135 |
+
|
| 136 |
+
if self.demodulate:
|
| 137 |
+
demod = torch.rsqrt(weight.pow(2).sum([2, 3, 4]) + self.eps)
|
| 138 |
+
weight = weight * demod.view(b, self.out_channels, 1, 1, 1)
|
| 139 |
+
|
| 140 |
+
weight = weight.view(b * self.out_channels, c, self.kernel_size, self.kernel_size)
|
| 141 |
+
|
| 142 |
+
if self.sample_mode == 'upsample':
|
| 143 |
+
x = F.interpolate(x, scale_factor=2, mode=self.interpolation_mode, align_corners=self.align_corners)
|
| 144 |
+
elif self.sample_mode == 'downsample':
|
| 145 |
+
x = F.interpolate(x, scale_factor=0.5, mode=self.interpolation_mode, align_corners=self.align_corners)
|
| 146 |
+
|
| 147 |
+
b, c, h, w = x.shape
|
| 148 |
+
x = x.view(1, b * c, h, w)
|
| 149 |
+
# weight: (b*c_out, c_in, k, k), groups=b
|
| 150 |
+
out = F.conv2d(x, weight, padding=self.padding, groups=b)
|
| 151 |
+
out = out.view(b, self.out_channels, *out.shape[2:4])
|
| 152 |
+
|
| 153 |
+
return out
|
| 154 |
+
|
| 155 |
+
def __repr__(self):
|
| 156 |
+
return (f'{self.__class__.__name__}(in_channels={self.in_channels}, '
|
| 157 |
+
f'out_channels={self.out_channels}, '
|
| 158 |
+
f'kernel_size={self.kernel_size}, '
|
| 159 |
+
f'demodulate={self.demodulate}, sample_mode={self.sample_mode})')
|
| 160 |
+
|
| 161 |
+
|
| 162 |
+
class StyleConv(nn.Module):
|
| 163 |
+
"""Style conv.
|
| 164 |
+
|
| 165 |
+
Args:
|
| 166 |
+
in_channels (int): Channel number of the input.
|
| 167 |
+
out_channels (int): Channel number of the output.
|
| 168 |
+
kernel_size (int): Size of the convolving kernel.
|
| 169 |
+
num_style_feat (int): Channel number of style features.
|
| 170 |
+
demodulate (bool): Whether demodulate in the conv layer. Default: True.
|
| 171 |
+
sample_mode (str | None): Indicating 'upsample', 'downsample' or None.
|
| 172 |
+
Default: None.
|
| 173 |
+
"""
|
| 174 |
+
|
| 175 |
+
def __init__(self,
|
| 176 |
+
in_channels,
|
| 177 |
+
out_channels,
|
| 178 |
+
kernel_size,
|
| 179 |
+
num_style_feat,
|
| 180 |
+
demodulate=True,
|
| 181 |
+
sample_mode=None,
|
| 182 |
+
interpolation_mode='bilinear'):
|
| 183 |
+
super(StyleConv, self).__init__()
|
| 184 |
+
self.modulated_conv = ModulatedConv2d(
|
| 185 |
+
in_channels,
|
| 186 |
+
out_channels,
|
| 187 |
+
kernel_size,
|
| 188 |
+
num_style_feat,
|
| 189 |
+
demodulate=demodulate,
|
| 190 |
+
sample_mode=sample_mode,
|
| 191 |
+
interpolation_mode=interpolation_mode)
|
| 192 |
+
self.weight = nn.Parameter(torch.zeros(1)) # for noise injection
|
| 193 |
+
self.activate = FusedLeakyReLU(out_channels)
|
| 194 |
+
|
| 195 |
+
def forward(self, x, style, noise=None):
|
| 196 |
+
# modulate
|
| 197 |
+
out = self.modulated_conv(x, style)
|
| 198 |
+
# noise injection
|
| 199 |
+
if noise is None:
|
| 200 |
+
b, _, h, w = out.shape
|
| 201 |
+
noise = out.new_empty(b, 1, h, w).normal_()
|
| 202 |
+
out = out + self.weight * noise
|
| 203 |
+
# activation (with bias)
|
| 204 |
+
out = self.activate(out)
|
| 205 |
+
return out
|
| 206 |
+
|
| 207 |
+
|
| 208 |
+
class ToRGB(nn.Module):
|
| 209 |
+
"""To RGB from features.
|
| 210 |
+
|
| 211 |
+
Args:
|
| 212 |
+
in_channels (int): Channel number of input.
|
| 213 |
+
num_style_feat (int): Channel number of style features.
|
| 214 |
+
upsample (bool): Whether to upsample. Default: True.
|
| 215 |
+
"""
|
| 216 |
+
|
| 217 |
+
def __init__(self, in_channels, num_style_feat, upsample=True, interpolation_mode='bilinear'):
|
| 218 |
+
super(ToRGB, self).__init__()
|
| 219 |
+
self.upsample = upsample
|
| 220 |
+
self.interpolation_mode = interpolation_mode
|
| 221 |
+
if self.interpolation_mode == 'nearest':
|
| 222 |
+
self.align_corners = None
|
| 223 |
+
else:
|
| 224 |
+
self.align_corners = False
|
| 225 |
+
self.modulated_conv = ModulatedConv2d(
|
| 226 |
+
in_channels,
|
| 227 |
+
3,
|
| 228 |
+
kernel_size=1,
|
| 229 |
+
num_style_feat=num_style_feat,
|
| 230 |
+
demodulate=False,
|
| 231 |
+
sample_mode=None,
|
| 232 |
+
interpolation_mode=interpolation_mode)
|
| 233 |
+
self.bias = nn.Parameter(torch.zeros(1, 3, 1, 1))
|
| 234 |
+
|
| 235 |
+
def forward(self, x, style, skip=None):
|
| 236 |
+
"""Forward function.
|
| 237 |
+
|
| 238 |
+
Args:
|
| 239 |
+
x (Tensor): Feature tensor with shape (b, c, h, w).
|
| 240 |
+
style (Tensor): Tensor with shape (b, num_style_feat).
|
| 241 |
+
skip (Tensor): Base/skip tensor. Default: None.
|
| 242 |
+
|
| 243 |
+
Returns:
|
| 244 |
+
Tensor: RGB images.
|
| 245 |
+
"""
|
| 246 |
+
out = self.modulated_conv(x, style)
|
| 247 |
+
out = out + self.bias
|
| 248 |
+
if skip is not None:
|
| 249 |
+
if self.upsample:
|
| 250 |
+
skip = F.interpolate(
|
| 251 |
+
skip, scale_factor=2, mode=self.interpolation_mode, align_corners=self.align_corners)
|
| 252 |
+
out = out + skip
|
| 253 |
+
return out
|
| 254 |
+
|
| 255 |
+
|
| 256 |
+
class ConstantInput(nn.Module):
|
| 257 |
+
"""Constant input.
|
| 258 |
+
|
| 259 |
+
Args:
|
| 260 |
+
num_channel (int): Channel number of constant input.
|
| 261 |
+
size (int): Spatial size of constant input.
|
| 262 |
+
"""
|
| 263 |
+
|
| 264 |
+
def __init__(self, num_channel, size):
|
| 265 |
+
super(ConstantInput, self).__init__()
|
| 266 |
+
self.weight = nn.Parameter(torch.randn(1, num_channel, size, size))
|
| 267 |
+
|
| 268 |
+
def forward(self, batch):
|
| 269 |
+
out = self.weight.repeat(batch, 1, 1, 1)
|
| 270 |
+
return out
|
| 271 |
+
|
| 272 |
+
|
| 273 |
+
@ARCH_REGISTRY.register()
|
| 274 |
+
class StyleGAN2GeneratorBilinear(nn.Module):
|
| 275 |
+
"""StyleGAN2 Generator.
|
| 276 |
+
|
| 277 |
+
Args:
|
| 278 |
+
out_size (int): The spatial size of outputs.
|
| 279 |
+
num_style_feat (int): Channel number of style features. Default: 512.
|
| 280 |
+
num_mlp (int): Layer number of MLP style layers. Default: 8.
|
| 281 |
+
channel_multiplier (int): Channel multiplier for large networks of
|
| 282 |
+
StyleGAN2. Default: 2.
|
| 283 |
+
lr_mlp (float): Learning rate multiplier for mlp layers. Default: 0.01.
|
| 284 |
+
narrow (float): Narrow ratio for channels. Default: 1.0.
|
| 285 |
+
"""
|
| 286 |
+
|
| 287 |
+
def __init__(self,
|
| 288 |
+
out_size,
|
| 289 |
+
num_style_feat=512,
|
| 290 |
+
num_mlp=8,
|
| 291 |
+
channel_multiplier=2,
|
| 292 |
+
lr_mlp=0.01,
|
| 293 |
+
narrow=1,
|
| 294 |
+
interpolation_mode='bilinear'):
|
| 295 |
+
super(StyleGAN2GeneratorBilinear, self).__init__()
|
| 296 |
+
# Style MLP layers
|
| 297 |
+
self.num_style_feat = num_style_feat
|
| 298 |
+
style_mlp_layers = [NormStyleCode()]
|
| 299 |
+
for i in range(num_mlp):
|
| 300 |
+
style_mlp_layers.append(
|
| 301 |
+
EqualLinear(
|
| 302 |
+
num_style_feat, num_style_feat, bias=True, bias_init_val=0, lr_mul=lr_mlp,
|
| 303 |
+
activation='fused_lrelu'))
|
| 304 |
+
self.style_mlp = nn.Sequential(*style_mlp_layers)
|
| 305 |
+
|
| 306 |
+
channels = {
|
| 307 |
+
'4': int(512 * narrow),
|
| 308 |
+
'8': int(512 * narrow),
|
| 309 |
+
'16': int(512 * narrow),
|
| 310 |
+
'32': int(512 * narrow),
|
| 311 |
+
'64': int(256 * channel_multiplier * narrow),
|
| 312 |
+
'128': int(128 * channel_multiplier * narrow),
|
| 313 |
+
'256': int(64 * channel_multiplier * narrow),
|
| 314 |
+
'512': int(32 * channel_multiplier * narrow),
|
| 315 |
+
'1024': int(16 * channel_multiplier * narrow)
|
| 316 |
+
}
|
| 317 |
+
self.channels = channels
|
| 318 |
+
|
| 319 |
+
self.constant_input = ConstantInput(channels['4'], size=4)
|
| 320 |
+
self.style_conv1 = StyleConv(
|
| 321 |
+
channels['4'],
|
| 322 |
+
channels['4'],
|
| 323 |
+
kernel_size=3,
|
| 324 |
+
num_style_feat=num_style_feat,
|
| 325 |
+
demodulate=True,
|
| 326 |
+
sample_mode=None,
|
| 327 |
+
interpolation_mode=interpolation_mode)
|
| 328 |
+
self.to_rgb1 = ToRGB(channels['4'], num_style_feat, upsample=False, interpolation_mode=interpolation_mode)
|
| 329 |
+
|
| 330 |
+
self.log_size = int(math.log(out_size, 2))
|
| 331 |
+
self.num_layers = (self.log_size - 2) * 2 + 1
|
| 332 |
+
self.num_latent = self.log_size * 2 - 2
|
| 333 |
+
|
| 334 |
+
self.style_convs = nn.ModuleList()
|
| 335 |
+
self.to_rgbs = nn.ModuleList()
|
| 336 |
+
self.noises = nn.Module()
|
| 337 |
+
|
| 338 |
+
in_channels = channels['4']
|
| 339 |
+
# noise
|
| 340 |
+
for layer_idx in range(self.num_layers):
|
| 341 |
+
resolution = 2**((layer_idx + 5) // 2)
|
| 342 |
+
shape = [1, 1, resolution, resolution]
|
| 343 |
+
self.noises.register_buffer(f'noise{layer_idx}', torch.randn(*shape))
|
| 344 |
+
# style convs and to_rgbs
|
| 345 |
+
for i in range(3, self.log_size + 1):
|
| 346 |
+
out_channels = channels[f'{2**i}']
|
| 347 |
+
self.style_convs.append(
|
| 348 |
+
StyleConv(
|
| 349 |
+
in_channels,
|
| 350 |
+
out_channels,
|
| 351 |
+
kernel_size=3,
|
| 352 |
+
num_style_feat=num_style_feat,
|
| 353 |
+
demodulate=True,
|
| 354 |
+
sample_mode='upsample',
|
| 355 |
+
interpolation_mode=interpolation_mode))
|
| 356 |
+
self.style_convs.append(
|
| 357 |
+
StyleConv(
|
| 358 |
+
out_channels,
|
| 359 |
+
out_channels,
|
| 360 |
+
kernel_size=3,
|
| 361 |
+
num_style_feat=num_style_feat,
|
| 362 |
+
demodulate=True,
|
| 363 |
+
sample_mode=None,
|
| 364 |
+
interpolation_mode=interpolation_mode))
|
| 365 |
+
self.to_rgbs.append(
|
| 366 |
+
ToRGB(out_channels, num_style_feat, upsample=True, interpolation_mode=interpolation_mode))
|
| 367 |
+
in_channels = out_channels
|
| 368 |
+
|
| 369 |
+
def make_noise(self):
|
| 370 |
+
"""Make noise for noise injection."""
|
| 371 |
+
device = self.constant_input.weight.device
|
| 372 |
+
noises = [torch.randn(1, 1, 4, 4, device=device)]
|
| 373 |
+
|
| 374 |
+
for i in range(3, self.log_size + 1):
|
| 375 |
+
for _ in range(2):
|
| 376 |
+
noises.append(torch.randn(1, 1, 2**i, 2**i, device=device))
|
| 377 |
+
|
| 378 |
+
return noises
|
| 379 |
+
|
| 380 |
+
def get_latent(self, x):
|
| 381 |
+
return self.style_mlp(x)
|
| 382 |
+
|
| 383 |
+
def mean_latent(self, num_latent):
|
| 384 |
+
latent_in = torch.randn(num_latent, self.num_style_feat, device=self.constant_input.weight.device)
|
| 385 |
+
latent = self.style_mlp(latent_in).mean(0, keepdim=True)
|
| 386 |
+
return latent
|
| 387 |
+
|
| 388 |
+
def forward(self,
|
| 389 |
+
styles,
|
| 390 |
+
input_is_latent=False,
|
| 391 |
+
noise=None,
|
| 392 |
+
randomize_noise=True,
|
| 393 |
+
truncation=1,
|
| 394 |
+
truncation_latent=None,
|
| 395 |
+
inject_index=None,
|
| 396 |
+
return_latents=False):
|
| 397 |
+
"""Forward function for StyleGAN2Generator.
|
| 398 |
+
|
| 399 |
+
Args:
|
| 400 |
+
styles (list[Tensor]): Sample codes of styles.
|
| 401 |
+
input_is_latent (bool): Whether input is latent style.
|
| 402 |
+
Default: False.
|
| 403 |
+
noise (Tensor | None): Input noise or None. Default: None.
|
| 404 |
+
randomize_noise (bool): Randomize noise, used when 'noise' is
|
| 405 |
+
False. Default: True.
|
| 406 |
+
truncation (float): TODO. Default: 1.
|
| 407 |
+
truncation_latent (Tensor | None): TODO. Default: None.
|
| 408 |
+
inject_index (int | None): The injection index for mixing noise.
|
| 409 |
+
Default: None.
|
| 410 |
+
return_latents (bool): Whether to return style latents.
|
| 411 |
+
Default: False.
|
| 412 |
+
"""
|
| 413 |
+
# style codes -> latents with Style MLP layer
|
| 414 |
+
if not input_is_latent:
|
| 415 |
+
styles = [self.style_mlp(s) for s in styles]
|
| 416 |
+
# noises
|
| 417 |
+
if noise is None:
|
| 418 |
+
if randomize_noise:
|
| 419 |
+
noise = [None] * self.num_layers # for each style conv layer
|
| 420 |
+
else: # use the stored noise
|
| 421 |
+
noise = [getattr(self.noises, f'noise{i}') for i in range(self.num_layers)]
|
| 422 |
+
# style truncation
|
| 423 |
+
if truncation < 1:
|
| 424 |
+
style_truncation = []
|
| 425 |
+
for style in styles:
|
| 426 |
+
style_truncation.append(truncation_latent + truncation * (style - truncation_latent))
|
| 427 |
+
styles = style_truncation
|
| 428 |
+
# get style latent with injection
|
| 429 |
+
if len(styles) == 1:
|
| 430 |
+
inject_index = self.num_latent
|
| 431 |
+
|
| 432 |
+
if styles[0].ndim < 3:
|
| 433 |
+
# repeat latent code for all the layers
|
| 434 |
+
latent = styles[0].unsqueeze(1).repeat(1, inject_index, 1)
|
| 435 |
+
else: # used for encoder with different latent code for each layer
|
| 436 |
+
latent = styles[0]
|
| 437 |
+
elif len(styles) == 2: # mixing noises
|
| 438 |
+
if inject_index is None:
|
| 439 |
+
inject_index = random.randint(1, self.num_latent - 1)
|
| 440 |
+
latent1 = styles[0].unsqueeze(1).repeat(1, inject_index, 1)
|
| 441 |
+
latent2 = styles[1].unsqueeze(1).repeat(1, self.num_latent - inject_index, 1)
|
| 442 |
+
latent = torch.cat([latent1, latent2], 1)
|
| 443 |
+
|
| 444 |
+
# main generation
|
| 445 |
+
out = self.constant_input(latent.shape[0])
|
| 446 |
+
out = self.style_conv1(out, latent[:, 0], noise=noise[0])
|
| 447 |
+
skip = self.to_rgb1(out, latent[:, 1])
|
| 448 |
+
|
| 449 |
+
i = 1
|
| 450 |
+
for conv1, conv2, noise1, noise2, to_rgb in zip(self.style_convs[::2], self.style_convs[1::2], noise[1::2],
|
| 451 |
+
noise[2::2], self.to_rgbs):
|
| 452 |
+
out = conv1(out, latent[:, i], noise=noise1)
|
| 453 |
+
out = conv2(out, latent[:, i + 1], noise=noise2)
|
| 454 |
+
skip = to_rgb(out, latent[:, i + 2], skip)
|
| 455 |
+
i += 2
|
| 456 |
+
|
| 457 |
+
image = skip
|
| 458 |
+
|
| 459 |
+
if return_latents:
|
| 460 |
+
return image, latent
|
| 461 |
+
else:
|
| 462 |
+
return image, None
|
| 463 |
+
|
| 464 |
+
|
| 465 |
+
class ScaledLeakyReLU(nn.Module):
|
| 466 |
+
"""Scaled LeakyReLU.
|
| 467 |
+
|
| 468 |
+
Args:
|
| 469 |
+
negative_slope (float): Negative slope. Default: 0.2.
|
| 470 |
+
"""
|
| 471 |
+
|
| 472 |
+
def __init__(self, negative_slope=0.2):
|
| 473 |
+
super(ScaledLeakyReLU, self).__init__()
|
| 474 |
+
self.negative_slope = negative_slope
|
| 475 |
+
|
| 476 |
+
def forward(self, x):
|
| 477 |
+
out = F.leaky_relu(x, negative_slope=self.negative_slope)
|
| 478 |
+
return out * math.sqrt(2)
|
| 479 |
+
|
| 480 |
+
|
| 481 |
+
class EqualConv2d(nn.Module):
|
| 482 |
+
"""Equalized Linear as StyleGAN2.
|
| 483 |
+
|
| 484 |
+
Args:
|
| 485 |
+
in_channels (int): Channel number of the input.
|
| 486 |
+
out_channels (int): Channel number of the output.
|
| 487 |
+
kernel_size (int): Size of the convolving kernel.
|
| 488 |
+
stride (int): Stride of the convolution. Default: 1
|
| 489 |
+
padding (int): Zero-padding added to both sides of the input.
|
| 490 |
+
Default: 0.
|
| 491 |
+
bias (bool): If ``True``, adds a learnable bias to the output.
|
| 492 |
+
Default: ``True``.
|
| 493 |
+
bias_init_val (float): Bias initialized value. Default: 0.
|
| 494 |
+
"""
|
| 495 |
+
|
| 496 |
+
def __init__(self, in_channels, out_channels, kernel_size, stride=1, padding=0, bias=True, bias_init_val=0):
|
| 497 |
+
super(EqualConv2d, self).__init__()
|
| 498 |
+
self.in_channels = in_channels
|
| 499 |
+
self.out_channels = out_channels
|
| 500 |
+
self.kernel_size = kernel_size
|
| 501 |
+
self.stride = stride
|
| 502 |
+
self.padding = padding
|
| 503 |
+
self.scale = 1 / math.sqrt(in_channels * kernel_size**2)
|
| 504 |
+
|
| 505 |
+
self.weight = nn.Parameter(torch.randn(out_channels, in_channels, kernel_size, kernel_size))
|
| 506 |
+
if bias:
|
| 507 |
+
self.bias = nn.Parameter(torch.zeros(out_channels).fill_(bias_init_val))
|
| 508 |
+
else:
|
| 509 |
+
self.register_parameter('bias', None)
|
| 510 |
+
|
| 511 |
+
def forward(self, x):
|
| 512 |
+
out = F.conv2d(
|
| 513 |
+
x,
|
| 514 |
+
self.weight * self.scale,
|
| 515 |
+
bias=self.bias,
|
| 516 |
+
stride=self.stride,
|
| 517 |
+
padding=self.padding,
|
| 518 |
+
)
|
| 519 |
+
|
| 520 |
+
return out
|
| 521 |
+
|
| 522 |
+
def __repr__(self):
|
| 523 |
+
return (f'{self.__class__.__name__}(in_channels={self.in_channels}, '
|
| 524 |
+
f'out_channels={self.out_channels}, '
|
| 525 |
+
f'kernel_size={self.kernel_size},'
|
| 526 |
+
f' stride={self.stride}, padding={self.padding}, '
|
| 527 |
+
f'bias={self.bias is not None})')
|
| 528 |
+
|
| 529 |
+
|
| 530 |
+
class ConvLayer(nn.Sequential):
|
| 531 |
+
"""Conv Layer used in StyleGAN2 Discriminator.
|
| 532 |
+
|
| 533 |
+
Args:
|
| 534 |
+
in_channels (int): Channel number of the input.
|
| 535 |
+
out_channels (int): Channel number of the output.
|
| 536 |
+
kernel_size (int): Kernel size.
|
| 537 |
+
downsample (bool): Whether downsample by a factor of 2.
|
| 538 |
+
Default: False.
|
| 539 |
+
bias (bool): Whether with bias. Default: True.
|
| 540 |
+
activate (bool): Whether use activateion. Default: True.
|
| 541 |
+
"""
|
| 542 |
+
|
| 543 |
+
def __init__(self,
|
| 544 |
+
in_channels,
|
| 545 |
+
out_channels,
|
| 546 |
+
kernel_size,
|
| 547 |
+
downsample=False,
|
| 548 |
+
bias=True,
|
| 549 |
+
activate=True,
|
| 550 |
+
interpolation_mode='bilinear'):
|
| 551 |
+
layers = []
|
| 552 |
+
self.interpolation_mode = interpolation_mode
|
| 553 |
+
# downsample
|
| 554 |
+
if downsample:
|
| 555 |
+
if self.interpolation_mode == 'nearest':
|
| 556 |
+
self.align_corners = None
|
| 557 |
+
else:
|
| 558 |
+
self.align_corners = False
|
| 559 |
+
|
| 560 |
+
layers.append(
|
| 561 |
+
torch.nn.Upsample(scale_factor=0.5, mode=interpolation_mode, align_corners=self.align_corners))
|
| 562 |
+
stride = 1
|
| 563 |
+
self.padding = kernel_size // 2
|
| 564 |
+
# conv
|
| 565 |
+
layers.append(
|
| 566 |
+
EqualConv2d(
|
| 567 |
+
in_channels, out_channels, kernel_size, stride=stride, padding=self.padding, bias=bias
|
| 568 |
+
and not activate))
|
| 569 |
+
# activation
|
| 570 |
+
if activate:
|
| 571 |
+
if bias:
|
| 572 |
+
layers.append(FusedLeakyReLU(out_channels))
|
| 573 |
+
else:
|
| 574 |
+
layers.append(ScaledLeakyReLU(0.2))
|
| 575 |
+
|
| 576 |
+
super(ConvLayer, self).__init__(*layers)
|
| 577 |
+
|
| 578 |
+
|
| 579 |
+
class ResBlock(nn.Module):
|
| 580 |
+
"""Residual block used in StyleGAN2 Discriminator.
|
| 581 |
+
|
| 582 |
+
Args:
|
| 583 |
+
in_channels (int): Channel number of the input.
|
| 584 |
+
out_channels (int): Channel number of the output.
|
| 585 |
+
"""
|
| 586 |
+
|
| 587 |
+
def __init__(self, in_channels, out_channels, interpolation_mode='bilinear'):
|
| 588 |
+
super(ResBlock, self).__init__()
|
| 589 |
+
|
| 590 |
+
self.conv1 = ConvLayer(in_channels, in_channels, 3, bias=True, activate=True)
|
| 591 |
+
self.conv2 = ConvLayer(
|
| 592 |
+
in_channels,
|
| 593 |
+
out_channels,
|
| 594 |
+
3,
|
| 595 |
+
downsample=True,
|
| 596 |
+
interpolation_mode=interpolation_mode,
|
| 597 |
+
bias=True,
|
| 598 |
+
activate=True)
|
| 599 |
+
self.skip = ConvLayer(
|
| 600 |
+
in_channels,
|
| 601 |
+
out_channels,
|
| 602 |
+
1,
|
| 603 |
+
downsample=True,
|
| 604 |
+
interpolation_mode=interpolation_mode,
|
| 605 |
+
bias=False,
|
| 606 |
+
activate=False)
|
| 607 |
+
|
| 608 |
+
def forward(self, x):
|
| 609 |
+
out = self.conv1(x)
|
| 610 |
+
out = self.conv2(out)
|
| 611 |
+
skip = self.skip(x)
|
| 612 |
+
out = (out + skip) / math.sqrt(2)
|
| 613 |
+
return out
|
gfpgan/archs/stylegan2_clean_arch.py
ADDED
|
@@ -0,0 +1,368 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import math
|
| 2 |
+
import random
|
| 3 |
+
import torch
|
| 4 |
+
from basicsr.archs.arch_util import default_init_weights
|
| 5 |
+
from basicsr.utils.registry import ARCH_REGISTRY
|
| 6 |
+
from torch import nn
|
| 7 |
+
from torch.nn import functional as F
|
| 8 |
+
|
| 9 |
+
|
| 10 |
+
class NormStyleCode(nn.Module):
|
| 11 |
+
|
| 12 |
+
def forward(self, x):
|
| 13 |
+
"""Normalize the style codes.
|
| 14 |
+
|
| 15 |
+
Args:
|
| 16 |
+
x (Tensor): Style codes with shape (b, c).
|
| 17 |
+
|
| 18 |
+
Returns:
|
| 19 |
+
Tensor: Normalized tensor.
|
| 20 |
+
"""
|
| 21 |
+
return x * torch.rsqrt(torch.mean(x**2, dim=1, keepdim=True) + 1e-8)
|
| 22 |
+
|
| 23 |
+
|
| 24 |
+
class ModulatedConv2d(nn.Module):
|
| 25 |
+
"""Modulated Conv2d used in StyleGAN2.
|
| 26 |
+
|
| 27 |
+
There is no bias in ModulatedConv2d.
|
| 28 |
+
|
| 29 |
+
Args:
|
| 30 |
+
in_channels (int): Channel number of the input.
|
| 31 |
+
out_channels (int): Channel number of the output.
|
| 32 |
+
kernel_size (int): Size of the convolving kernel.
|
| 33 |
+
num_style_feat (int): Channel number of style features.
|
| 34 |
+
demodulate (bool): Whether to demodulate in the conv layer. Default: True.
|
| 35 |
+
sample_mode (str | None): Indicating 'upsample', 'downsample' or None. Default: None.
|
| 36 |
+
eps (float): A value added to the denominator for numerical stability. Default: 1e-8.
|
| 37 |
+
"""
|
| 38 |
+
|
| 39 |
+
def __init__(self,
|
| 40 |
+
in_channels,
|
| 41 |
+
out_channels,
|
| 42 |
+
kernel_size,
|
| 43 |
+
num_style_feat,
|
| 44 |
+
demodulate=True,
|
| 45 |
+
sample_mode=None,
|
| 46 |
+
eps=1e-8):
|
| 47 |
+
super(ModulatedConv2d, self).__init__()
|
| 48 |
+
self.in_channels = in_channels
|
| 49 |
+
self.out_channels = out_channels
|
| 50 |
+
self.kernel_size = kernel_size
|
| 51 |
+
self.demodulate = demodulate
|
| 52 |
+
self.sample_mode = sample_mode
|
| 53 |
+
self.eps = eps
|
| 54 |
+
|
| 55 |
+
# modulation inside each modulated conv
|
| 56 |
+
self.modulation = nn.Linear(num_style_feat, in_channels, bias=True)
|
| 57 |
+
# initialization
|
| 58 |
+
default_init_weights(self.modulation, scale=1, bias_fill=1, a=0, mode='fan_in', nonlinearity='linear')
|
| 59 |
+
|
| 60 |
+
self.weight = nn.Parameter(
|
| 61 |
+
torch.randn(1, out_channels, in_channels, kernel_size, kernel_size) /
|
| 62 |
+
math.sqrt(in_channels * kernel_size**2))
|
| 63 |
+
self.padding = kernel_size // 2
|
| 64 |
+
|
| 65 |
+
def forward(self, x, style):
|
| 66 |
+
"""Forward function.
|
| 67 |
+
|
| 68 |
+
Args:
|
| 69 |
+
x (Tensor): Tensor with shape (b, c, h, w).
|
| 70 |
+
style (Tensor): Tensor with shape (b, num_style_feat).
|
| 71 |
+
|
| 72 |
+
Returns:
|
| 73 |
+
Tensor: Modulated tensor after convolution.
|
| 74 |
+
"""
|
| 75 |
+
b, c, h, w = x.shape # c = c_in
|
| 76 |
+
# weight modulation
|
| 77 |
+
style = self.modulation(style).view(b, 1, c, 1, 1)
|
| 78 |
+
# self.weight: (1, c_out, c_in, k, k); style: (b, 1, c, 1, 1)
|
| 79 |
+
weight = self.weight * style # (b, c_out, c_in, k, k)
|
| 80 |
+
|
| 81 |
+
if self.demodulate:
|
| 82 |
+
demod = torch.rsqrt(weight.pow(2).sum([2, 3, 4]) + self.eps)
|
| 83 |
+
weight = weight * demod.view(b, self.out_channels, 1, 1, 1)
|
| 84 |
+
|
| 85 |
+
weight = weight.view(b * self.out_channels, c, self.kernel_size, self.kernel_size)
|
| 86 |
+
|
| 87 |
+
# upsample or downsample if necessary
|
| 88 |
+
if self.sample_mode == 'upsample':
|
| 89 |
+
x = F.interpolate(x, scale_factor=2, mode='bilinear', align_corners=False)
|
| 90 |
+
elif self.sample_mode == 'downsample':
|
| 91 |
+
x = F.interpolate(x, scale_factor=0.5, mode='bilinear', align_corners=False)
|
| 92 |
+
|
| 93 |
+
b, c, h, w = x.shape
|
| 94 |
+
x = x.view(1, b * c, h, w)
|
| 95 |
+
# weight: (b*c_out, c_in, k, k), groups=b
|
| 96 |
+
out = F.conv2d(x, weight, padding=self.padding, groups=b)
|
| 97 |
+
out = out.view(b, self.out_channels, *out.shape[2:4])
|
| 98 |
+
|
| 99 |
+
return out
|
| 100 |
+
|
| 101 |
+
def __repr__(self):
|
| 102 |
+
return (f'{self.__class__.__name__}(in_channels={self.in_channels}, out_channels={self.out_channels}, '
|
| 103 |
+
f'kernel_size={self.kernel_size}, demodulate={self.demodulate}, sample_mode={self.sample_mode})')
|
| 104 |
+
|
| 105 |
+
|
| 106 |
+
class StyleConv(nn.Module):
|
| 107 |
+
"""Style conv used in StyleGAN2.
|
| 108 |
+
|
| 109 |
+
Args:
|
| 110 |
+
in_channels (int): Channel number of the input.
|
| 111 |
+
out_channels (int): Channel number of the output.
|
| 112 |
+
kernel_size (int): Size of the convolving kernel.
|
| 113 |
+
num_style_feat (int): Channel number of style features.
|
| 114 |
+
demodulate (bool): Whether demodulate in the conv layer. Default: True.
|
| 115 |
+
sample_mode (str | None): Indicating 'upsample', 'downsample' or None. Default: None.
|
| 116 |
+
"""
|
| 117 |
+
|
| 118 |
+
def __init__(self, in_channels, out_channels, kernel_size, num_style_feat, demodulate=True, sample_mode=None):
|
| 119 |
+
super(StyleConv, self).__init__()
|
| 120 |
+
self.modulated_conv = ModulatedConv2d(
|
| 121 |
+
in_channels, out_channels, kernel_size, num_style_feat, demodulate=demodulate, sample_mode=sample_mode)
|
| 122 |
+
self.weight = nn.Parameter(torch.zeros(1)) # for noise injection
|
| 123 |
+
self.bias = nn.Parameter(torch.zeros(1, out_channels, 1, 1))
|
| 124 |
+
self.activate = nn.LeakyReLU(negative_slope=0.2, inplace=True)
|
| 125 |
+
|
| 126 |
+
def forward(self, x, style, noise=None):
|
| 127 |
+
# modulate
|
| 128 |
+
out = self.modulated_conv(x, style) * 2**0.5 # for conversion
|
| 129 |
+
# noise injection
|
| 130 |
+
if noise is None:
|
| 131 |
+
b, _, h, w = out.shape
|
| 132 |
+
noise = out.new_empty(b, 1, h, w).normal_()
|
| 133 |
+
out = out + self.weight * noise
|
| 134 |
+
# add bias
|
| 135 |
+
out = out + self.bias
|
| 136 |
+
# activation
|
| 137 |
+
out = self.activate(out)
|
| 138 |
+
return out
|
| 139 |
+
|
| 140 |
+
|
| 141 |
+
class ToRGB(nn.Module):
|
| 142 |
+
"""To RGB (image space) from features.
|
| 143 |
+
|
| 144 |
+
Args:
|
| 145 |
+
in_channels (int): Channel number of input.
|
| 146 |
+
num_style_feat (int): Channel number of style features.
|
| 147 |
+
upsample (bool): Whether to upsample. Default: True.
|
| 148 |
+
"""
|
| 149 |
+
|
| 150 |
+
def __init__(self, in_channels, num_style_feat, upsample=True):
|
| 151 |
+
super(ToRGB, self).__init__()
|
| 152 |
+
self.upsample = upsample
|
| 153 |
+
self.modulated_conv = ModulatedConv2d(
|
| 154 |
+
in_channels, 3, kernel_size=1, num_style_feat=num_style_feat, demodulate=False, sample_mode=None)
|
| 155 |
+
self.bias = nn.Parameter(torch.zeros(1, 3, 1, 1))
|
| 156 |
+
|
| 157 |
+
def forward(self, x, style, skip=None):
|
| 158 |
+
"""Forward function.
|
| 159 |
+
|
| 160 |
+
Args:
|
| 161 |
+
x (Tensor): Feature tensor with shape (b, c, h, w).
|
| 162 |
+
style (Tensor): Tensor with shape (b, num_style_feat).
|
| 163 |
+
skip (Tensor): Base/skip tensor. Default: None.
|
| 164 |
+
|
| 165 |
+
Returns:
|
| 166 |
+
Tensor: RGB images.
|
| 167 |
+
"""
|
| 168 |
+
out = self.modulated_conv(x, style)
|
| 169 |
+
out = out + self.bias
|
| 170 |
+
if skip is not None:
|
| 171 |
+
if self.upsample:
|
| 172 |
+
skip = F.interpolate(skip, scale_factor=2, mode='bilinear', align_corners=False)
|
| 173 |
+
out = out + skip
|
| 174 |
+
return out
|
| 175 |
+
|
| 176 |
+
|
| 177 |
+
class ConstantInput(nn.Module):
|
| 178 |
+
"""Constant input.
|
| 179 |
+
|
| 180 |
+
Args:
|
| 181 |
+
num_channel (int): Channel number of constant input.
|
| 182 |
+
size (int): Spatial size of constant input.
|
| 183 |
+
"""
|
| 184 |
+
|
| 185 |
+
def __init__(self, num_channel, size):
|
| 186 |
+
super(ConstantInput, self).__init__()
|
| 187 |
+
self.weight = nn.Parameter(torch.randn(1, num_channel, size, size))
|
| 188 |
+
|
| 189 |
+
def forward(self, batch):
|
| 190 |
+
out = self.weight.repeat(batch, 1, 1, 1)
|
| 191 |
+
return out
|
| 192 |
+
|
| 193 |
+
|
| 194 |
+
@ARCH_REGISTRY.register()
|
| 195 |
+
class StyleGAN2GeneratorClean(nn.Module):
|
| 196 |
+
"""Clean version of StyleGAN2 Generator.
|
| 197 |
+
|
| 198 |
+
Args:
|
| 199 |
+
out_size (int): The spatial size of outputs.
|
| 200 |
+
num_style_feat (int): Channel number of style features. Default: 512.
|
| 201 |
+
num_mlp (int): Layer number of MLP style layers. Default: 8.
|
| 202 |
+
channel_multiplier (int): Channel multiplier for large networks of StyleGAN2. Default: 2.
|
| 203 |
+
narrow (float): Narrow ratio for channels. Default: 1.0.
|
| 204 |
+
"""
|
| 205 |
+
|
| 206 |
+
def __init__(self, out_size, num_style_feat=512, num_mlp=8, channel_multiplier=2, narrow=1):
|
| 207 |
+
super(StyleGAN2GeneratorClean, self).__init__()
|
| 208 |
+
# Style MLP layers
|
| 209 |
+
self.num_style_feat = num_style_feat
|
| 210 |
+
style_mlp_layers = [NormStyleCode()]
|
| 211 |
+
for i in range(num_mlp):
|
| 212 |
+
style_mlp_layers.extend(
|
| 213 |
+
[nn.Linear(num_style_feat, num_style_feat, bias=True),
|
| 214 |
+
nn.LeakyReLU(negative_slope=0.2, inplace=True)])
|
| 215 |
+
self.style_mlp = nn.Sequential(*style_mlp_layers)
|
| 216 |
+
# initialization
|
| 217 |
+
default_init_weights(self.style_mlp, scale=1, bias_fill=0, a=0.2, mode='fan_in', nonlinearity='leaky_relu')
|
| 218 |
+
|
| 219 |
+
# channel list
|
| 220 |
+
channels = {
|
| 221 |
+
'4': int(512 * narrow),
|
| 222 |
+
'8': int(512 * narrow),
|
| 223 |
+
'16': int(512 * narrow),
|
| 224 |
+
'32': int(512 * narrow),
|
| 225 |
+
'64': int(256 * channel_multiplier * narrow),
|
| 226 |
+
'128': int(128 * channel_multiplier * narrow),
|
| 227 |
+
'256': int(64 * channel_multiplier * narrow),
|
| 228 |
+
'512': int(32 * channel_multiplier * narrow),
|
| 229 |
+
'1024': int(16 * channel_multiplier * narrow)
|
| 230 |
+
}
|
| 231 |
+
self.channels = channels
|
| 232 |
+
|
| 233 |
+
self.constant_input = ConstantInput(channels['4'], size=4)
|
| 234 |
+
self.style_conv1 = StyleConv(
|
| 235 |
+
channels['4'],
|
| 236 |
+
channels['4'],
|
| 237 |
+
kernel_size=3,
|
| 238 |
+
num_style_feat=num_style_feat,
|
| 239 |
+
demodulate=True,
|
| 240 |
+
sample_mode=None)
|
| 241 |
+
self.to_rgb1 = ToRGB(channels['4'], num_style_feat, upsample=False)
|
| 242 |
+
|
| 243 |
+
self.log_size = int(math.log(out_size, 2))
|
| 244 |
+
self.num_layers = (self.log_size - 2) * 2 + 1
|
| 245 |
+
self.num_latent = self.log_size * 2 - 2
|
| 246 |
+
|
| 247 |
+
self.style_convs = nn.ModuleList()
|
| 248 |
+
self.to_rgbs = nn.ModuleList()
|
| 249 |
+
self.noises = nn.Module()
|
| 250 |
+
|
| 251 |
+
in_channels = channels['4']
|
| 252 |
+
# noise
|
| 253 |
+
for layer_idx in range(self.num_layers):
|
| 254 |
+
resolution = 2**((layer_idx + 5) // 2)
|
| 255 |
+
shape = [1, 1, resolution, resolution]
|
| 256 |
+
self.noises.register_buffer(f'noise{layer_idx}', torch.randn(*shape))
|
| 257 |
+
# style convs and to_rgbs
|
| 258 |
+
for i in range(3, self.log_size + 1):
|
| 259 |
+
out_channels = channels[f'{2**i}']
|
| 260 |
+
self.style_convs.append(
|
| 261 |
+
StyleConv(
|
| 262 |
+
in_channels,
|
| 263 |
+
out_channels,
|
| 264 |
+
kernel_size=3,
|
| 265 |
+
num_style_feat=num_style_feat,
|
| 266 |
+
demodulate=True,
|
| 267 |
+
sample_mode='upsample'))
|
| 268 |
+
self.style_convs.append(
|
| 269 |
+
StyleConv(
|
| 270 |
+
out_channels,
|
| 271 |
+
out_channels,
|
| 272 |
+
kernel_size=3,
|
| 273 |
+
num_style_feat=num_style_feat,
|
| 274 |
+
demodulate=True,
|
| 275 |
+
sample_mode=None))
|
| 276 |
+
self.to_rgbs.append(ToRGB(out_channels, num_style_feat, upsample=True))
|
| 277 |
+
in_channels = out_channels
|
| 278 |
+
|
| 279 |
+
def make_noise(self):
|
| 280 |
+
"""Make noise for noise injection."""
|
| 281 |
+
device = self.constant_input.weight.device
|
| 282 |
+
noises = [torch.randn(1, 1, 4, 4, device=device)]
|
| 283 |
+
|
| 284 |
+
for i in range(3, self.log_size + 1):
|
| 285 |
+
for _ in range(2):
|
| 286 |
+
noises.append(torch.randn(1, 1, 2**i, 2**i, device=device))
|
| 287 |
+
|
| 288 |
+
return noises
|
| 289 |
+
|
| 290 |
+
def get_latent(self, x):
|
| 291 |
+
return self.style_mlp(x)
|
| 292 |
+
|
| 293 |
+
def mean_latent(self, num_latent):
|
| 294 |
+
latent_in = torch.randn(num_latent, self.num_style_feat, device=self.constant_input.weight.device)
|
| 295 |
+
latent = self.style_mlp(latent_in).mean(0, keepdim=True)
|
| 296 |
+
return latent
|
| 297 |
+
|
| 298 |
+
def forward(self,
|
| 299 |
+
styles,
|
| 300 |
+
input_is_latent=False,
|
| 301 |
+
noise=None,
|
| 302 |
+
randomize_noise=True,
|
| 303 |
+
truncation=1,
|
| 304 |
+
truncation_latent=None,
|
| 305 |
+
inject_index=None,
|
| 306 |
+
return_latents=False):
|
| 307 |
+
"""Forward function for StyleGAN2GeneratorClean.
|
| 308 |
+
|
| 309 |
+
Args:
|
| 310 |
+
styles (list[Tensor]): Sample codes of styles.
|
| 311 |
+
input_is_latent (bool): Whether input is latent style. Default: False.
|
| 312 |
+
noise (Tensor | None): Input noise or None. Default: None.
|
| 313 |
+
randomize_noise (bool): Randomize noise, used when 'noise' is False. Default: True.
|
| 314 |
+
truncation (float): The truncation ratio. Default: 1.
|
| 315 |
+
truncation_latent (Tensor | None): The truncation latent tensor. Default: None.
|
| 316 |
+
inject_index (int | None): The injection index for mixing noise. Default: None.
|
| 317 |
+
return_latents (bool): Whether to return style latents. Default: False.
|
| 318 |
+
"""
|
| 319 |
+
# style codes -> latents with Style MLP layer
|
| 320 |
+
if not input_is_latent:
|
| 321 |
+
styles = [self.style_mlp(s) for s in styles]
|
| 322 |
+
# noises
|
| 323 |
+
if noise is None:
|
| 324 |
+
if randomize_noise:
|
| 325 |
+
noise = [None] * self.num_layers # for each style conv layer
|
| 326 |
+
else: # use the stored noise
|
| 327 |
+
noise = [getattr(self.noises, f'noise{i}') for i in range(self.num_layers)]
|
| 328 |
+
# style truncation
|
| 329 |
+
if truncation < 1:
|
| 330 |
+
style_truncation = []
|
| 331 |
+
for style in styles:
|
| 332 |
+
style_truncation.append(truncation_latent + truncation * (style - truncation_latent))
|
| 333 |
+
styles = style_truncation
|
| 334 |
+
# get style latents with injection
|
| 335 |
+
if len(styles) == 1:
|
| 336 |
+
inject_index = self.num_latent
|
| 337 |
+
|
| 338 |
+
if styles[0].ndim < 3:
|
| 339 |
+
# repeat latent code for all the layers
|
| 340 |
+
latent = styles[0].unsqueeze(1).repeat(1, inject_index, 1)
|
| 341 |
+
else: # used for encoder with different latent code for each layer
|
| 342 |
+
latent = styles[0]
|
| 343 |
+
elif len(styles) == 2: # mixing noises
|
| 344 |
+
if inject_index is None:
|
| 345 |
+
inject_index = random.randint(1, self.num_latent - 1)
|
| 346 |
+
latent1 = styles[0].unsqueeze(1).repeat(1, inject_index, 1)
|
| 347 |
+
latent2 = styles[1].unsqueeze(1).repeat(1, self.num_latent - inject_index, 1)
|
| 348 |
+
latent = torch.cat([latent1, latent2], 1)
|
| 349 |
+
|
| 350 |
+
# main generation
|
| 351 |
+
out = self.constant_input(latent.shape[0])
|
| 352 |
+
out = self.style_conv1(out, latent[:, 0], noise=noise[0])
|
| 353 |
+
skip = self.to_rgb1(out, latent[:, 1])
|
| 354 |
+
|
| 355 |
+
i = 1
|
| 356 |
+
for conv1, conv2, noise1, noise2, to_rgb in zip(self.style_convs[::2], self.style_convs[1::2], noise[1::2],
|
| 357 |
+
noise[2::2], self.to_rgbs):
|
| 358 |
+
out = conv1(out, latent[:, i], noise=noise1)
|
| 359 |
+
out = conv2(out, latent[:, i + 1], noise=noise2)
|
| 360 |
+
skip = to_rgb(out, latent[:, i + 2], skip) # feature back to the rgb space
|
| 361 |
+
i += 2
|
| 362 |
+
|
| 363 |
+
image = skip
|
| 364 |
+
|
| 365 |
+
if return_latents:
|
| 366 |
+
return image, latent
|
| 367 |
+
else:
|
| 368 |
+
return image, None
|
gfpgan/data/__init__.py
ADDED
|
@@ -0,0 +1,10 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import importlib
|
| 2 |
+
from basicsr.utils import scandir
|
| 3 |
+
from os import path as osp
|
| 4 |
+
|
| 5 |
+
# automatically scan and import dataset modules for registry
|
| 6 |
+
# scan all the files that end with '_dataset.py' under the data folder
|
| 7 |
+
data_folder = osp.dirname(osp.abspath(__file__))
|
| 8 |
+
dataset_filenames = [osp.splitext(osp.basename(v))[0] for v in scandir(data_folder) if v.endswith('_dataset.py')]
|
| 9 |
+
# import all the dataset modules
|
| 10 |
+
_dataset_modules = [importlib.import_module(f'gfpgan.data.{file_name}') for file_name in dataset_filenames]
|
gfpgan/data/__pycache__/__init__.cpython-311.pyc
ADDED
|
Binary file (1.18 kB). View file
|
|
|
gfpgan/data/__pycache__/ffhq_degradation_dataset.cpython-311.pyc
ADDED
|
Binary file (14.4 kB). View file
|
|
|
gfpgan/data/ffhq_degradation_dataset.py
ADDED
|
@@ -0,0 +1,230 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import cv2
|
| 2 |
+
import math
|
| 3 |
+
import numpy as np
|
| 4 |
+
import os.path as osp
|
| 5 |
+
import torch
|
| 6 |
+
import torch.utils.data as data
|
| 7 |
+
from basicsr.data import degradations as degradations
|
| 8 |
+
from basicsr.data.data_util import paths_from_folder
|
| 9 |
+
from basicsr.data.transforms import augment
|
| 10 |
+
from basicsr.utils import FileClient, get_root_logger, imfrombytes, img2tensor
|
| 11 |
+
from basicsr.utils.registry import DATASET_REGISTRY
|
| 12 |
+
from torchvision.transforms.functional import (adjust_brightness, adjust_contrast, adjust_hue, adjust_saturation,
|
| 13 |
+
normalize)
|
| 14 |
+
|
| 15 |
+
|
| 16 |
+
@DATASET_REGISTRY.register()
|
| 17 |
+
class FFHQDegradationDataset(data.Dataset):
|
| 18 |
+
"""FFHQ dataset for GFPGAN.
|
| 19 |
+
|
| 20 |
+
It reads high resolution images, and then generate low-quality (LQ) images on-the-fly.
|
| 21 |
+
|
| 22 |
+
Args:
|
| 23 |
+
opt (dict): Config for train datasets. It contains the following keys:
|
| 24 |
+
dataroot_gt (str): Data root path for gt.
|
| 25 |
+
io_backend (dict): IO backend type and other kwarg.
|
| 26 |
+
mean (list | tuple): Image mean.
|
| 27 |
+
std (list | tuple): Image std.
|
| 28 |
+
use_hflip (bool): Whether to horizontally flip.
|
| 29 |
+
Please see more options in the codes.
|
| 30 |
+
"""
|
| 31 |
+
|
| 32 |
+
def __init__(self, opt):
|
| 33 |
+
super(FFHQDegradationDataset, self).__init__()
|
| 34 |
+
self.opt = opt
|
| 35 |
+
# file client (io backend)
|
| 36 |
+
self.file_client = None
|
| 37 |
+
self.io_backend_opt = opt['io_backend']
|
| 38 |
+
|
| 39 |
+
self.gt_folder = opt['dataroot_gt']
|
| 40 |
+
self.mean = opt['mean']
|
| 41 |
+
self.std = opt['std']
|
| 42 |
+
self.out_size = opt['out_size']
|
| 43 |
+
|
| 44 |
+
self.crop_components = opt.get('crop_components', False) # facial components
|
| 45 |
+
self.eye_enlarge_ratio = opt.get('eye_enlarge_ratio', 1) # whether enlarge eye regions
|
| 46 |
+
|
| 47 |
+
if self.crop_components:
|
| 48 |
+
# load component list from a pre-process pth files
|
| 49 |
+
self.components_list = torch.load(opt.get('component_path'))
|
| 50 |
+
|
| 51 |
+
# file client (lmdb io backend)
|
| 52 |
+
if self.io_backend_opt['type'] == 'lmdb':
|
| 53 |
+
self.io_backend_opt['db_paths'] = self.gt_folder
|
| 54 |
+
if not self.gt_folder.endswith('.lmdb'):
|
| 55 |
+
raise ValueError(f"'dataroot_gt' should end with '.lmdb', but received {self.gt_folder}")
|
| 56 |
+
with open(osp.join(self.gt_folder, 'meta_info.txt')) as fin:
|
| 57 |
+
self.paths = [line.split('.')[0] for line in fin]
|
| 58 |
+
else:
|
| 59 |
+
# disk backend: scan file list from a folder
|
| 60 |
+
self.paths = paths_from_folder(self.gt_folder)
|
| 61 |
+
|
| 62 |
+
# degradation configurations
|
| 63 |
+
self.blur_kernel_size = opt['blur_kernel_size']
|
| 64 |
+
self.kernel_list = opt['kernel_list']
|
| 65 |
+
self.kernel_prob = opt['kernel_prob']
|
| 66 |
+
self.blur_sigma = opt['blur_sigma']
|
| 67 |
+
self.downsample_range = opt['downsample_range']
|
| 68 |
+
self.noise_range = opt['noise_range']
|
| 69 |
+
self.jpeg_range = opt['jpeg_range']
|
| 70 |
+
|
| 71 |
+
# color jitter
|
| 72 |
+
self.color_jitter_prob = opt.get('color_jitter_prob')
|
| 73 |
+
self.color_jitter_pt_prob = opt.get('color_jitter_pt_prob')
|
| 74 |
+
self.color_jitter_shift = opt.get('color_jitter_shift', 20)
|
| 75 |
+
# to gray
|
| 76 |
+
self.gray_prob = opt.get('gray_prob')
|
| 77 |
+
|
| 78 |
+
logger = get_root_logger()
|
| 79 |
+
logger.info(f'Blur: blur_kernel_size {self.blur_kernel_size}, sigma: [{", ".join(map(str, self.blur_sigma))}]')
|
| 80 |
+
logger.info(f'Downsample: downsample_range [{", ".join(map(str, self.downsample_range))}]')
|
| 81 |
+
logger.info(f'Noise: [{", ".join(map(str, self.noise_range))}]')
|
| 82 |
+
logger.info(f'JPEG compression: [{", ".join(map(str, self.jpeg_range))}]')
|
| 83 |
+
|
| 84 |
+
if self.color_jitter_prob is not None:
|
| 85 |
+
logger.info(f'Use random color jitter. Prob: {self.color_jitter_prob}, shift: {self.color_jitter_shift}')
|
| 86 |
+
if self.gray_prob is not None:
|
| 87 |
+
logger.info(f'Use random gray. Prob: {self.gray_prob}')
|
| 88 |
+
self.color_jitter_shift /= 255.
|
| 89 |
+
|
| 90 |
+
@staticmethod
|
| 91 |
+
def color_jitter(img, shift):
|
| 92 |
+
"""jitter color: randomly jitter the RGB values, in numpy formats"""
|
| 93 |
+
jitter_val = np.random.uniform(-shift, shift, 3).astype(np.float32)
|
| 94 |
+
img = img + jitter_val
|
| 95 |
+
img = np.clip(img, 0, 1)
|
| 96 |
+
return img
|
| 97 |
+
|
| 98 |
+
@staticmethod
|
| 99 |
+
def color_jitter_pt(img, brightness, contrast, saturation, hue):
|
| 100 |
+
"""jitter color: randomly jitter the brightness, contrast, saturation, and hue, in torch Tensor formats"""
|
| 101 |
+
fn_idx = torch.randperm(4)
|
| 102 |
+
for fn_id in fn_idx:
|
| 103 |
+
if fn_id == 0 and brightness is not None:
|
| 104 |
+
brightness_factor = torch.tensor(1.0).uniform_(brightness[0], brightness[1]).item()
|
| 105 |
+
img = adjust_brightness(img, brightness_factor)
|
| 106 |
+
|
| 107 |
+
if fn_id == 1 and contrast is not None:
|
| 108 |
+
contrast_factor = torch.tensor(1.0).uniform_(contrast[0], contrast[1]).item()
|
| 109 |
+
img = adjust_contrast(img, contrast_factor)
|
| 110 |
+
|
| 111 |
+
if fn_id == 2 and saturation is not None:
|
| 112 |
+
saturation_factor = torch.tensor(1.0).uniform_(saturation[0], saturation[1]).item()
|
| 113 |
+
img = adjust_saturation(img, saturation_factor)
|
| 114 |
+
|
| 115 |
+
if fn_id == 3 and hue is not None:
|
| 116 |
+
hue_factor = torch.tensor(1.0).uniform_(hue[0], hue[1]).item()
|
| 117 |
+
img = adjust_hue(img, hue_factor)
|
| 118 |
+
return img
|
| 119 |
+
|
| 120 |
+
def get_component_coordinates(self, index, status):
|
| 121 |
+
"""Get facial component (left_eye, right_eye, mouth) coordinates from a pre-loaded pth file"""
|
| 122 |
+
components_bbox = self.components_list[f'{index:08d}']
|
| 123 |
+
if status[0]: # hflip
|
| 124 |
+
# exchange right and left eye
|
| 125 |
+
tmp = components_bbox['left_eye']
|
| 126 |
+
components_bbox['left_eye'] = components_bbox['right_eye']
|
| 127 |
+
components_bbox['right_eye'] = tmp
|
| 128 |
+
# modify the width coordinate
|
| 129 |
+
components_bbox['left_eye'][0] = self.out_size - components_bbox['left_eye'][0]
|
| 130 |
+
components_bbox['right_eye'][0] = self.out_size - components_bbox['right_eye'][0]
|
| 131 |
+
components_bbox['mouth'][0] = self.out_size - components_bbox['mouth'][0]
|
| 132 |
+
|
| 133 |
+
# get coordinates
|
| 134 |
+
locations = []
|
| 135 |
+
for part in ['left_eye', 'right_eye', 'mouth']:
|
| 136 |
+
mean = components_bbox[part][0:2]
|
| 137 |
+
half_len = components_bbox[part][2]
|
| 138 |
+
if 'eye' in part:
|
| 139 |
+
half_len *= self.eye_enlarge_ratio
|
| 140 |
+
loc = np.hstack((mean - half_len + 1, mean + half_len))
|
| 141 |
+
loc = torch.from_numpy(loc).float()
|
| 142 |
+
locations.append(loc)
|
| 143 |
+
return locations
|
| 144 |
+
|
| 145 |
+
def __getitem__(self, index):
|
| 146 |
+
if self.file_client is None:
|
| 147 |
+
self.file_client = FileClient(self.io_backend_opt.pop('type'), **self.io_backend_opt)
|
| 148 |
+
|
| 149 |
+
# load gt image
|
| 150 |
+
# Shape: (h, w, c); channel order: BGR; image range: [0, 1], float32.
|
| 151 |
+
gt_path = self.paths[index]
|
| 152 |
+
img_bytes = self.file_client.get(gt_path)
|
| 153 |
+
img_gt = imfrombytes(img_bytes, float32=True)
|
| 154 |
+
|
| 155 |
+
# random horizontal flip
|
| 156 |
+
img_gt, status = augment(img_gt, hflip=self.opt['use_hflip'], rotation=False, return_status=True)
|
| 157 |
+
h, w, _ = img_gt.shape
|
| 158 |
+
|
| 159 |
+
# get facial component coordinates
|
| 160 |
+
if self.crop_components:
|
| 161 |
+
locations = self.get_component_coordinates(index, status)
|
| 162 |
+
loc_left_eye, loc_right_eye, loc_mouth = locations
|
| 163 |
+
|
| 164 |
+
# ------------------------ generate lq image ------------------------ #
|
| 165 |
+
# blur
|
| 166 |
+
kernel = degradations.random_mixed_kernels(
|
| 167 |
+
self.kernel_list,
|
| 168 |
+
self.kernel_prob,
|
| 169 |
+
self.blur_kernel_size,
|
| 170 |
+
self.blur_sigma,
|
| 171 |
+
self.blur_sigma, [-math.pi, math.pi],
|
| 172 |
+
noise_range=None)
|
| 173 |
+
img_lq = cv2.filter2D(img_gt, -1, kernel)
|
| 174 |
+
# downsample
|
| 175 |
+
scale = np.random.uniform(self.downsample_range[0], self.downsample_range[1])
|
| 176 |
+
img_lq = cv2.resize(img_lq, (int(w // scale), int(h // scale)), interpolation=cv2.INTER_LINEAR)
|
| 177 |
+
# noise
|
| 178 |
+
if self.noise_range is not None:
|
| 179 |
+
img_lq = degradations.random_add_gaussian_noise(img_lq, self.noise_range)
|
| 180 |
+
# jpeg compression
|
| 181 |
+
if self.jpeg_range is not None:
|
| 182 |
+
img_lq = degradations.random_add_jpg_compression(img_lq, self.jpeg_range)
|
| 183 |
+
|
| 184 |
+
# resize to original size
|
| 185 |
+
img_lq = cv2.resize(img_lq, (w, h), interpolation=cv2.INTER_LINEAR)
|
| 186 |
+
|
| 187 |
+
# random color jitter (only for lq)
|
| 188 |
+
if self.color_jitter_prob is not None and (np.random.uniform() < self.color_jitter_prob):
|
| 189 |
+
img_lq = self.color_jitter(img_lq, self.color_jitter_shift)
|
| 190 |
+
# random to gray (only for lq)
|
| 191 |
+
if self.gray_prob and np.random.uniform() < self.gray_prob:
|
| 192 |
+
img_lq = cv2.cvtColor(img_lq, cv2.COLOR_BGR2GRAY)
|
| 193 |
+
img_lq = np.tile(img_lq[:, :, None], [1, 1, 3])
|
| 194 |
+
if self.opt.get('gt_gray'): # whether convert GT to gray images
|
| 195 |
+
img_gt = cv2.cvtColor(img_gt, cv2.COLOR_BGR2GRAY)
|
| 196 |
+
img_gt = np.tile(img_gt[:, :, None], [1, 1, 3]) # repeat the color channels
|
| 197 |
+
|
| 198 |
+
# BGR to RGB, HWC to CHW, numpy to tensor
|
| 199 |
+
img_gt, img_lq = img2tensor([img_gt, img_lq], bgr2rgb=True, float32=True)
|
| 200 |
+
|
| 201 |
+
# random color jitter (pytorch version) (only for lq)
|
| 202 |
+
if self.color_jitter_pt_prob is not None and (np.random.uniform() < self.color_jitter_pt_prob):
|
| 203 |
+
brightness = self.opt.get('brightness', (0.5, 1.5))
|
| 204 |
+
contrast = self.opt.get('contrast', (0.5, 1.5))
|
| 205 |
+
saturation = self.opt.get('saturation', (0, 1.5))
|
| 206 |
+
hue = self.opt.get('hue', (-0.1, 0.1))
|
| 207 |
+
img_lq = self.color_jitter_pt(img_lq, brightness, contrast, saturation, hue)
|
| 208 |
+
|
| 209 |
+
# round and clip
|
| 210 |
+
img_lq = torch.clamp((img_lq * 255.0).round(), 0, 255) / 255.
|
| 211 |
+
|
| 212 |
+
# normalize
|
| 213 |
+
normalize(img_gt, self.mean, self.std, inplace=True)
|
| 214 |
+
normalize(img_lq, self.mean, self.std, inplace=True)
|
| 215 |
+
|
| 216 |
+
if self.crop_components:
|
| 217 |
+
return_dict = {
|
| 218 |
+
'lq': img_lq,
|
| 219 |
+
'gt': img_gt,
|
| 220 |
+
'gt_path': gt_path,
|
| 221 |
+
'loc_left_eye': loc_left_eye,
|
| 222 |
+
'loc_right_eye': loc_right_eye,
|
| 223 |
+
'loc_mouth': loc_mouth
|
| 224 |
+
}
|
| 225 |
+
return return_dict
|
| 226 |
+
else:
|
| 227 |
+
return {'lq': img_lq, 'gt': img_gt, 'gt_path': gt_path}
|
| 228 |
+
|
| 229 |
+
def __len__(self):
|
| 230 |
+
return len(self.paths)
|
gfpgan/models/__init__.py
ADDED
|
@@ -0,0 +1,10 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import importlib
|
| 2 |
+
from basicsr.utils import scandir
|
| 3 |
+
from os import path as osp
|
| 4 |
+
|
| 5 |
+
# automatically scan and import model modules for registry
|
| 6 |
+
# scan all the files that end with '_model.py' under the model folder
|
| 7 |
+
model_folder = osp.dirname(osp.abspath(__file__))
|
| 8 |
+
model_filenames = [osp.splitext(osp.basename(v))[0] for v in scandir(model_folder) if v.endswith('_model.py')]
|
| 9 |
+
# import all the model modules
|
| 10 |
+
_model_modules = [importlib.import_module(f'gfpgan.models.{file_name}') for file_name in model_filenames]
|