working training on textures from 3D-FUTURE-model: edited dataset preparation... (1a75a7f5) · Commits · Martin Mirbauer / stylegan2-ada-pytorch

.gitignore

+1 −0

Original line number	Original line	Diff line number	Diff line
	__pycache__/		__pycache__/
	.cache/		.cache/
			training-runs/

README.md

+15 −0

Original line number	Original line	Diff line number	Diff line
			## StyleGAN2-ADA on 3D-FUTURE-model textures

			One-liner to prepare the environment (Miniconda on Ubuntu 22.04, with a 1080Ti):
			```conda create --name stylegan2 pytorch==1.8.0 torchvision==0.9.0 cudatoolkit=11.1 -c pytorch -c conda-forge && conda activate stylegan2 && pip install click requests tqdm pyspng ninja imageio-ffmpeg==0.4.3 psutil scipy```
			or create a Conda environment from `environment.yml`.

			Resize input images, save to .zip (~3GB, 16554 images):
			```python dataset_tool.py --source=/home/user/texture-mapping/2_batch_render/data/3D-FUTURE-model --dest=3D-FUTURE-model-textures-256x256.zip --width=256 --height=256 --resize-filter=lanczos```

			Start training:
			```python train.py --outdir=training-runs --data=3D-FUTURE-model-textures-256x256.zip --gpus=1```


			---

	## StyleGAN2-ADA — Official PyTorch implementation		## StyleGAN2-ADA — Official PyTorch implementation

	![Teaser image](./docs/stylegan2-ada-teaser-1024x252.png)		![Teaser image](./docs/stylegan2-ada-teaser-1024x252.png)

dataset_tool.py

+28 −2

Original line number	Original line	Diff line number	Diff line
	@@ -49,8 +49,27 @@ def is_image_ext(fname: Union[str, Path]) -> bool:

	#----------------------------------------------------------------------------		#----------------------------------------------------------------------------

			exclude = set([ # list copied from batchrender_pt3d.ipynb
			# texture.png truncated
			'6934dea0-1d66-49c4-82c6-4d54d41f9707',
			'f92e4a5d-691a-4555-a4e7-e09aa1dfd80e',
			'e9bd517b-849a-4d61-a78d-9a05285d1b32',

			# the rest may not need to be excluded
			# CUDA out of memory
			'34bd8afc-dd97-3b13-8d19-3fa28f025e51', # due to 2k texture?
			'207129c6-8b86-3fa7-bc82-54444f33d2ff', # 1k texture?

			# empty model.mtl
			'12b1ce0f-3d2a-3211-9e05-a537e5a43e20',
			'530c3dbc-ae01-3b03-a52d-7d9e61fe4f7f',

			# UserWarning: Faces have invalid indices
			'7e101ef3-7722-4af8-90d5-7c562834fabd',
			])

	def open_image_folder(source_dir, *, max_images: Optional[int]):		def open_image_folder(source_dir, *, max_images: Optional[int]):
	input_images = [str(f) for f in sorted(Path(source_dir).rglob('*')) if is_image_ext(f) and os.path.isfile(f)]		input_images = [str(f) for f in sorted(Path(source_dir).rglob('texture.png')) if is_image_ext(f) and os.path.isfile(f)]

	# Load labels.		# Load labels.
	labels = {}		labels = {}
	@@ -69,7 +88,14 @@ def open_image_folder(source_dir, *, max_images: Optional[int]):
	for idx, fname in enumerate(input_images):		for idx, fname in enumerate(input_images):
	arch_fname = os.path.relpath(fname, source_dir)		arch_fname = os.path.relpath(fname, source_dir)
	arch_fname = arch_fname.replace('\\', '/')		arch_fname = arch_fname.replace('\\', '/')
	img = np.array(PIL.Image.open(fname))		if any(fname.find(e)>=0 for e in exclude):
			print(f'iterate_images: skipping {fname}')
			continue
			img = PIL.Image.open(fname)
			#if img.getbands() == ('R', 'G', 'B', 'A'):
			# print(f'iterate_images: {fname} is RGBA, converting to RGB')
			img = img.convert('RGB')
			img = np.array(img)
	yield dict(img=img, label=labels.get(arch_fname))		yield dict(img=img, label=labels.get(arch_fname))
	if idx >= max_idx-1:		if idx >= max_idx-1:
	break		break

environment.yml

0 → 100644

+84 −0

Original line number	Original line	Diff line number	Diff line
			name: stylegan2
			channels:
			- pytorch
			- conda-forge
			- defaults
			dependencies:
			- _libgcc_mutex=0.1=conda_forge
			- _openmp_mutex=4.5=2_gnu
			- blas=2.16=mkl
			- bzip2=1.0.8=h4bc722e_7
			- ca-certificates=2025.1.31=hbd8a1cb_1
			- cudatoolkit=11.1.1=hb139c0e_13
			- ffmpeg=4.3=hf484d3e_0
			- freetype=2.12.1=h267a509_2
			- giflib=5.2.2=hd590300_0
			- gmp=6.3.0=hac33072_2
			- gnutls=3.6.13=h85f3911_1
			- intel-openmp=2025.0.0=h06a4308_1171
			- jpeg=9b=h024ee3a_2
			- lame=3.100=h166bdaf_1003
			- lcms2=2.12=h3be6417_0
			- ld_impl_linux-64=2.43=h712a8e2_4
			- libblas=3.8.0=16_mkl
			- libcblas=3.8.0=16_mkl
			- libffi=3.4.6=h2dba641_1
			- libgcc=14.2.0=h767d61c_2
			- libgcc-ng=14.2.0=h69a702a_2
			- libgfortran-ng=7.5.0=h14aa051_20
			- libgfortran4=7.5.0=h14aa051_20
			- libgomp=14.2.0=h767d61c_2
			- libiconv=1.18=h4ce23a2_1
			- liblapack=3.8.0=16_mkl
			- liblapacke=3.8.0=16_mkl
			- liblzma=5.8.1=hb9d3cd8_0
			- liblzma-devel=5.8.1=hb9d3cd8_0
			- libnsl=2.0.1=hd590300_0
			- libpng=1.6.43=h2797004_0
			- libsqlite=3.46.0=hde9e2c9_0
			- libstdcxx=14.2.0=h8f9b012_2
			- libstdcxx-ng=14.2.0=h4852527_2
			- libtiff=4.2.0=h85742a9_0
			- libuuid=2.38.1=h0b41bf4_0
			- libwebp=1.2.0=h89dd481_0
			- libwebp-base=1.2.0=h7f98852_2
			- libxcrypt=4.4.36=hd590300_1
			- libzlib=1.2.13=h4ab18f5_6
			- lz4-c=1.9.3=h9c3ff4c_1
			- mkl=2020.2=256
			- ncurses=6.5=h2d0b736_3
			- nettle=3.6=he412f7d_0
			- numpy=1.22.4=py39hc58783e_0
			- openh264=2.1.1=h780b84a_0
			- openssl=3.5.0=h7b32b05_0
			- pillow=9.3.0=py39hace64e9_1
			- pip=25.0.1=pyh8b19718_0
			- python=3.9.19=h0755675_0_cpython
			- python_abi=3.9=7_cp39
			- pytorch=1.8.0=py3.9_cuda11.1_cudnn8.0.5_0
			- readline=8.2=h8c095d6_2
			- setuptools=79.0.1=pyhff2d567_0
			- tk=8.6.13=noxft_h4845f30_101
			- torchvision=0.9.0=py39_cu111
			- typing_extensions=4.13.2=pyh29332c3_0
			- tzdata=2025b=h78e105d_0
			- wheel=0.45.1=pyhd8ed1ab_1
			- xz=5.8.1=hbcc6ac9_0
			- xz-gpl-tools=5.8.1=hbcc6ac9_0
			- xz-tools=5.8.1=hb9d3cd8_0
			- zlib=1.2.13=h4ab18f5_6
			- zstd=1.4.9=ha95c52a_0
			- pip:
			- certifi==2025.1.31
			- charset-normalizer==3.4.1
			- click==8.1.8
			- idna==3.10
			- imageio-ffmpeg==0.4.3
			- ninja==1.11.1.4
			- psutil==7.0.0
			- pyspng==0.1.3
			- requests==2.32.3
			- scipy==1.13.1
			- tqdm==4.67.1
			- urllib3==2.4.0
			prefix: /home/user/miniconda3/envs/stylegan2