Mirrored_Repos/InvokeAI
1
0
Fork 0
mirror of https://github.com/invoke-ai/InvokeAI synced 2024-08-30 20:32:17 +00:00
Code Issues Packages Projects Releases Wiki Activity

fix model download and autodetection bugs

Browse Source 
- Corrected error that caused --full-precision argument to be ignored
  when models downloaded using the --yes argument.

- Improved autodetection of v1 inpainting files; no longer relies on the
  file having 'inpaint' in the name.
This commit is contained in:
Lincoln Stein
2023-02-16 21:37:50 -05:00
parent 5d617ce63d
commit f3f4c68acc
3 changed files with 31 additions and 18 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
ldm/invoke/config/model_install.py
View File

@ -368,8 +368,11 @@ def process_and_execute(app: npyscreen.NPSAppManaged):
# -------------------------------------------------------- # --------------------------------------------------------
def select_and_download_models(opt: Namespace): def select_and_download_models(opt: Namespace):
if opt.default_only: if opt.default_only:
models_to_download = default_dataset() models_to_install = default_dataset()
install_requested_models(models_to_download) install_requested_models(
install_initial_models = models_to_install,
precision = 'float32' if opt.full_precision else choose_precision(torch.device(choose_torch_device())),
)
else: else:
installApp = AddModelApplication() installApp = AddModelApplication()
installApp.opt = opt installApp.opt = opt

1
ldm/invoke/config/model_install_backend.py
View File

@ -56,7 +56,6 @@ def initial_models():
return Datasets return Datasets
return (Datasets := OmegaConf.load(Dataset_path)) return (Datasets := OmegaConf.load(Dataset_path))
def install_requested_models( def install_requested_models(
install_initial_models: List[str] = None, install_initial_models: List[str] = None,
remove_models: List[str] = None, remove_models: List[str] = None,

41
ldm/invoke/model_manager.py
View File

@ -711,15 +711,15 @@ class ModelManager(object):
thing = path_url_or_repo # to save typing thing = path_url_or_repo # to save typing
if thing.startswith(('http:','https:','ftp:')): if thing.startswith(('http:','https:','ftp:')):
print(f'>> {thing} appears to be a URL') print(f' | {thing} appears to be a URL')
model_path = self._resolve_path(thing, 'models/ldm/stable-diffusion-v1') # _resolve_path does a download if needed model_path = self._resolve_path(thing, 'models/ldm/stable-diffusion-v1') # _resolve_path does a download if needed
elif Path(thing).is_file() and thing.endswith(('.ckpt','.safetensors')): elif Path(thing).is_file() and thing.endswith(('.ckpt','.safetensors')):
print(f'>> {thing} appears to be a checkpoint file on disk') print(f' | {thing} appears to be a checkpoint file on disk')
model_path = self._resolve_path(thing, 'models/ldm/stable-diffusion-v1') model_path = self._resolve_path(thing, 'models/ldm/stable-diffusion-v1')
elif Path(thing).is_dir() and Path(thing, 'model_index.json').exists(): elif Path(thing).is_dir() and Path(thing, 'model_index.json').exists():
print(f'>> {thing} appears to be a diffusers file on disk') print(f' | {thing} appears to be a diffusers file on disk')
model_name = self.import_diffusers_model( model_name = self.import_diffusers_model(
thing, thing,
vae=dict(repo_id='stabilityai/sd-vae-ft-mse'), vae=dict(repo_id='stabilityai/sd-vae-ft-mse'),
@ -729,17 +729,16 @@ class ModelManager(object):
elif Path(thing).is_dir(): elif Path(thing).is_dir():
print(f'>> {thing} appears to be a directory. Will scan for models to import') print(f'>> {thing} appears to be a directory. Will scan for models to import')
for m in list(Path(thing).rglob('*.ckpt')) + list(Path(thing).rglob('*.safetensors')): for m in list(Path(thing).rglob('*.ckpt')) + list(Path(thing).rglob('*.safetensors')):
print('***',m)
self.heuristic_import(str(m), convert, commit_to_conf=commit_to_conf) self.heuristic_import(str(m), convert, commit_to_conf=commit_to_conf)
return return
elif re.match(r'^[\w.+-]+/[\w.+-]+$', thing): elif re.match(r'^[\w.+-]+/[\w.+-]+$', thing):
print(f'>> {thing} appears to be a HuggingFace diffusers repo_id') print(f' | {thing} appears to be a HuggingFace diffusers repo_id')
model_name = self.import_diffuser_model(thing, commit_to_conf=commit_to_conf) model_name = self.import_diffuser_model(thing, commit_to_conf=commit_to_conf)
pipeline,_,_,_ = self._load_diffusers_model(self.config[model_name]) pipeline,_,_,_ = self._load_diffusers_model(self.config[model_name])
else: else:
print(f">> {thing}: Unknown thing. Please provide a URL, file path, directory or HuggingFace repo_id") print(f"** {thing}: Unknown thing. Please provide a URL, file path, directory or HuggingFace repo_id")
# Model_path is set in the event of a legacy checkpoint file. # Model_path is set in the event of a legacy checkpoint file.
# If not set, we're all done # If not set, we're all done
@ -747,25 +746,37 @@ class ModelManager(object):
return return
if model_path.stem in self.config: #already imported if model_path.stem in self.config: #already imported
print(f' > Already imported. Skipping')
return return
# another round of heuristics to guess the correct config file. # another round of heuristics to guess the correct config file.
model_config_file = Path(Globals.root,'configs/stable-diffusion/v1-inference.yaml') model_config_file = None
checkpoint = safetensors.torch.load_file(model_path) if model_path.suffix == '.safetensors' else torch.load(model_path) checkpoint = safetensors.torch.load_file(model_path) if model_path.suffix == '.safetensors' else torch.load(model_path)
key_name = "model.diffusion_model.input_blocks.2.1.transformer_blocks.0.attn2.to_k.weight" key_name = "model.diffusion_model.input_blocks.2.1.transformer_blocks.0.attn2.to_k.weight"
if key_name in checkpoint and checkpoint[key_name].shape[-1] == 1024: if key_name in checkpoint and checkpoint[key_name].shape[-1] == 1024:
print(f'>> {thing} appears to be an SD-v2 model; model will be converted to diffusers format') print(f' > SD-v2 model detected; model will be converted to diffusers format')
model_config_file = Path(Globals.root,'configs/stable-diffusion/v2-inference-v.yaml') model_config_file = Path(Globals.root,'configs/stable-diffusion/v2-inference-v.yaml')
convert = True convert = True
elif re.search('inpaint', str(model_path), flags=re.IGNORECASE): if not model_config_file: # still trying
print(f'>> {thing} appears to be an SD-v1 inpainting model') in_channels = None
model_config_file = Path(Globals.root,'configs/stable-diffusion/v1-inpainting-inference.yaml') try:
state_dict = checkpoint.get('state_dict') or checkpoint
in_channels = state_dict['model.diffusion_model.input_blocks.0.0.weight'].shape[1]
if in_channels == 9:
print(f' > SD-v1 inpainting model detected')
model_config_file = Path(Globals.root,'configs/stable-diffusion/v1-inpainting-inference.yaml')
elif in_channels == 4:
print(f' > SD-v1 model detected')
model_config_file = Path(Globals.root,'configs/stable-diffusion/v1-inference.yaml')
else:
print(f'** {thing} does not have an expected number of in_channels ({in_channels}). It will probably break when loaded.')
model_config_file = Path(Globals.root,'configs/stable-diffusion/v1-inference.yaml')
except KeyError:
print(f'** {thing} does not have the expected SD-v1 model fields. It will probably break when loaded.')
model_config_file = Path(Globals.root,'configs/stable-diffusion/v1-inference.yaml')
else:
print(f'>> {thing} appears to be an SD-v1 model')
if convert: if convert:
diffuser_path = Path(Globals.root, 'models',Globals.converted_ckpts_dir, model_path.stem) diffuser_path = Path(Globals.root, 'models',Globals.converted_ckpts_dir, model_path.stem)
self.convert_and_import( self.convert_and_import(