resolved merge conflicts

This commit is contained in:
Lincoln Stein 2022-08-24 11:50:48 -04:00
commit c24a16ccb0
8 changed files with 138 additions and 126 deletions

15
.gitignore vendored
View File

@ -1,3 +1,7 @@
# ignore default image save location and model symbolic link
outputs/
models/ldm/stable-diffusion-v1/model.ckpt
# Byte-compiled / optimized / DLL files # Byte-compiled / optimized / DLL files
__pycache__/ __pycache__/
*.py[cod] *.py[cod]
@ -6,6 +10,10 @@ __pycache__/
# C extensions # C extensions
*.so *.so
# emacs autosave files
*~
#*
# Distribution / packaging # Distribution / packaging
.Python .Python
build/ build/
@ -20,6 +28,7 @@ parts/
sdist/ sdist/
var/ var/
wheels/ wheels/
pip-wheel-metadata/
share/python-wheels/ share/python-wheels/
*.egg-info/ *.egg-info/
.installed.cfg .installed.cfg
@ -86,6 +95,7 @@ ipython_config.py
# For a library or package, you might want to ignore these files since the code is # For a library or package, you might want to ignore these files since the code is
# intended to run in multiple environments; otherwise, check them in: # intended to run in multiple environments; otherwise, check them in:
# .python-version # .python-version
.python-version
# pipenv # pipenv
# According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control. # According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.
@ -109,7 +119,7 @@ ipython_config.py
# https://pdm.fming.dev/#use-with-ide # https://pdm.fming.dev/#use-with-ide
.pdm.toml .pdm.toml
# PEP 582; used by e.g. github.com/David-OConnor/pyflow and github.com/pdm-project/pdm # PEP 582; used by e.g. github.com/David-OConnor/pyflow
__pypackages__/ __pypackages__/
# Celery stuff # Celery stuff
@ -159,8 +169,7 @@ cython_debug/
# option (not recommended) you can uncomment the following to ignore the entire idea folder. # option (not recommended) you can uncomment the following to ignore the entire idea folder.
#.idea/ #.idea/
**/*.ckpt
src/ src/
logs/ logs/
**/__pycache__/ **/__pycache__/
outputs outputs

View File

@ -17,9 +17,6 @@ class DDIMSampler(object):
self.schedule = schedule self.schedule = schedule
def register_buffer(self, name, attr): def register_buffer(self, name, attr):
if type(attr) == torch.Tensor:
if attr.device != torch.device("cuda"):
attr = attr.to(torch.device("cuda"))
setattr(self, name, attr) setattr(self, name, attr)
def make_schedule(self, ddim_num_steps, ddim_discretize="uniform", ddim_eta=0., verbose=True): def make_schedule(self, ddim_num_steps, ddim_discretize="uniform", ddim_eta=0., verbose=True):

View File

@ -16,9 +16,6 @@ class PLMSSampler(object):
self.schedule = schedule self.schedule = schedule
def register_buffer(self, name, attr): def register_buffer(self, name, attr):
if type(attr) == torch.Tensor:
if attr.device != torch.device("cuda"):
attr = attr.to(torch.device("cuda"))
setattr(self, name, attr) setattr(self, name, attr)
def make_schedule(self, ddim_num_steps, ddim_discretize="uniform", ddim_eta=0., verbose=True): def make_schedule(self, ddim_num_steps, ddim_discretize="uniform", ddim_eta=0., verbose=True):

View File

@ -123,7 +123,8 @@ The vast majority of these arguments default to reasonable values.
full_precision=False, full_precision=False,
strength=0.75, # default in scripts/img2img.py strength=0.75, # default in scripts/img2img.py
embedding_path=None, embedding_path=None,
latent_diffusion_weights=False # just to keep track of this parameter when regenerating prompt latent_diffusion_weights=False, # just to keep track of this parameter when regenerating prompt
device='cuda'
): ):
self.outdir = outdir self.outdir = outdir
self.batch_size = batch_size self.batch_size = batch_size
@ -147,11 +148,13 @@ The vast majority of these arguments default to reasonable values.
self.model = None # empty for now self.model = None # empty for now
self.sampler = None self.sampler = None
self.latent_diffusion_weights=latent_diffusion_weights self.latent_diffusion_weights=latent_diffusion_weights
self.device = device
if seed is None: if seed is None:
self.seed = self._new_seed() self.seed = self._new_seed()
else: else:
self.seed = seed self.seed = seed
@torch.no_grad()
def txt2img(self,prompt,outdir=None,batch_size=None,iterations=None, def txt2img(self,prompt,outdir=None,batch_size=None,iterations=None,
steps=None,seed=None,grid=None,individual=None,width=None,height=None, steps=None,seed=None,grid=None,individual=None,width=None,height=None,
cfg_scale=None,ddim_eta=None,strength=None,embedding_path=None,init_img=None,skip_normalize=False): cfg_scale=None,ddim_eta=None,strength=None,embedding_path=None,init_img=None,skip_normalize=False):
@ -206,69 +209,67 @@ The vast majority of these arguments default to reasonable values.
# Gawd. Too many levels of indent here. Need to refactor into smaller routines! # Gawd. Too many levels of indent here. Need to refactor into smaller routines!
try: try:
with torch.no_grad(): with precision_scope(self.device.type), model.ema_scope():
with precision_scope("cuda"): all_samples = list()
with model.ema_scope(): for n in trange(iterations, desc="Sampling"):
all_samples = list() seed_everything(seed)
for n in trange(iterations, desc="Sampling"): for prompts in tqdm(data, desc="data", dynamic_ncols=True):
seed_everything(seed) uc = None
for prompts in tqdm(data, desc="data", dynamic_ncols=True): if cfg_scale != 1.0:
uc = None uc = model.get_learned_conditioning(batch_size * [""])
if cfg_scale != 1.0: if isinstance(prompts, tuple):
uc = model.get_learned_conditioning(batch_size * [""]) prompts = list(prompts)
if isinstance(prompts, tuple):
prompts = list(prompts)
# weighted sub-prompts # weighted sub-prompts
subprompts,weights = T2I._split_weighted_subprompts(prompts[0]) subprompts,weights = T2I._split_weighted_subprompts(prompts[0])
if len(subprompts) > 1: if len(subprompts) > 1:
# i dont know if this is correct.. but it works # i dont know if this is correct.. but it works
c = torch.zeros_like(uc) c = torch.zeros_like(uc)
# get total weight for normalizing # get total weight for normalizing
totalWeight = sum(weights) totalWeight = sum(weights)
# normalize each "sub prompt" and add it # normalize each "sub prompt" and add it
for i in range(0,len(subprompts)): for i in range(0,len(subprompts)):
weight = weights[i] weight = weights[i]
if not skip_normalize: if not skip_normalize:
weight = weight / totalWeight weight = weight / totalWeight
c = torch.add(c,model.get_learned_conditioning(subprompts[i]), alpha=weight) c = torch.add(c,model.get_learned_conditioning(subprompts[i]), alpha=weight)
else: # just standard 1 prompt else: # just standard 1 prompt
c = model.get_learned_conditioning(prompts) c = model.get_learned_conditioning(prompts)
shape = [self.latent_channels, height // self.downsampling_factor, width // self.downsampling_factor] shape = [self.latent_channels, height // self.downsampling_factor, width // self.downsampling_factor]
samples_ddim, _ = sampler.sample(S=steps, samples_ddim, _ = sampler.sample(S=steps,
conditioning=c, conditioning=c,
batch_size=batch_size, batch_size=batch_size,
shape=shape, shape=shape,
verbose=False, verbose=False,
unconditional_guidance_scale=cfg_scale, unconditional_guidance_scale=cfg_scale,
unconditional_conditioning=uc, unconditional_conditioning=uc,
eta=ddim_eta, eta=ddim_eta,
x_T=start_code) x_T=start_code)
x_samples_ddim = model.decode_first_stage(samples_ddim) x_samples_ddim = model.decode_first_stage(samples_ddim)
x_samples_ddim = torch.clamp((x_samples_ddim + 1.0) / 2.0, min=0.0, max=1.0) x_samples_ddim = torch.clamp((x_samples_ddim + 1.0) / 2.0, min=0.0, max=1.0)
if not grid: if not grid:
for x_sample in x_samples_ddim: for x_sample in x_samples_ddim:
x_sample = 255. * rearrange(x_sample.cpu().numpy(), 'c h w -> h w c') x_sample = 255. * rearrange(x_sample.cpu().numpy(), 'c h w -> h w c')
filename = self._unique_filename(outdir,previousname=filename, filename = self._unique_filename(outdir,previousname=filename,
seed=seed,isbatch=(batch_size>1)) seed=seed,isbatch=(batch_size>1))
assert not os.path.exists(filename) assert not os.path.exists(filename)
Image.fromarray(x_sample.astype(np.uint8)).save(filename) Image.fromarray(x_sample.astype(np.uint8)).save(filename)
images.append([filename,seed]) images.append([filename,seed])
else: else:
all_samples.append(x_samples_ddim) all_samples.append(x_samples_ddim)
seeds.append(seed) seeds.append(seed)
image_count += 1 image_count += 1
seed = self._new_seed() seed = self._new_seed()
if grid: if grid:
images = self._make_grid(samples=all_samples, images = self._make_grid(samples=all_samples,
seeds=seeds, seeds=seeds,
batch_size=batch_size, batch_size=batch_size,
iterations=iterations, iterations=iterations,
outdir=outdir) outdir=outdir)
except KeyboardInterrupt: except KeyboardInterrupt:
print('*interrupted*') print('*interrupted*')
print('Partial results will be returned; if --grid was requested, nothing will be returned.') print('Partial results will be returned; if --grid was requested, nothing will be returned.')
@ -281,6 +282,7 @@ The vast majority of these arguments default to reasonable values.
return images return images
# There is lots of shared code between this and txt2img and should be refactored. # There is lots of shared code between this and txt2img and should be refactored.
@torch.no_grad()
def img2img(self,prompt,outdir=None,init_img=None,batch_size=None,iterations=None, def img2img(self,prompt,outdir=None,init_img=None,batch_size=None,iterations=None,
steps=None,seed=None,grid=None,individual=None,width=None,height=None, steps=None,seed=None,grid=None,individual=None,width=None,height=None,
cfg_scale=None,ddim_eta=None,strength=None,embedding_path=None,skip_normalize=False): cfg_scale=None,ddim_eta=None,strength=None,embedding_path=None,skip_normalize=False):
@ -331,7 +333,7 @@ The vast majority of these arguments default to reasonable values.
assert os.path.isfile(init_img) assert os.path.isfile(init_img)
init_image = self._load_img(init_img).to(self.device) init_image = self._load_img(init_img).to(self.device)
init_image = repeat(init_image, '1 ... -> b ...', b=batch_size) init_image = repeat(init_image, '1 ... -> b ...', b=batch_size)
with precision_scope("cuda"): with precision_scope(self.device.type):
init_latent = model.get_first_stage_encoding(model.encode_first_stage(init_image)) # move to latent space init_latent = model.get_first_stage_encoding(model.encode_first_stage(init_image)) # move to latent space
sampler.make_schedule(ddim_num_steps=steps, ddim_eta=ddim_eta, verbose=False) sampler.make_schedule(ddim_num_steps=steps, ddim_eta=ddim_eta, verbose=False)
@ -353,63 +355,61 @@ The vast majority of these arguments default to reasonable values.
# Gawd. Too many levels of indent here. Need to refactor into smaller routines! # Gawd. Too many levels of indent here. Need to refactor into smaller routines!
try: try:
with torch.no_grad(): with precision_scope(self.device.type), model.ema_scope():
with precision_scope("cuda"): all_samples = list()
with model.ema_scope(): for n in trange(iterations, desc="Sampling"):
all_samples = list() seed_everything(seed)
for n in trange(iterations, desc="Sampling"): for prompts in tqdm(data, desc="data", dynamic_ncols=True):
seed_everything(seed) uc = None
for prompts in tqdm(data, desc="data", dynamic_ncols=True): if cfg_scale != 1.0:
uc = None uc = model.get_learned_conditioning(batch_size * [""])
if cfg_scale != 1.0: if isinstance(prompts, tuple):
uc = model.get_learned_conditioning(batch_size * [""]) prompts = list(prompts)
if isinstance(prompts, tuple):
prompts = list(prompts)
# weighted sub-prompts # weighted sub-prompts
subprompts,weights = T2I._split_weighted_subprompts(prompts[0]) subprompts,weights = T2I._split_weighted_subprompts(prompts[0])
if len(subprompts) > 1: if len(subprompts) > 1:
# i dont know if this is correct.. but it works # i dont know if this is correct.. but it works
c = torch.zeros_like(uc) c = torch.zeros_like(uc)
# get total weight for normalizing # get total weight for normalizing
totalWeight = sum(weights) totalWeight = sum(weights)
# normalize each "sub prompt" and add it # normalize each "sub prompt" and add it
for i in range(0,len(subprompts)): for i in range(0,len(subprompts)):
weight = weights[i] weight = weights[i]
if not skip_normalize: if not skip_normalize:
weight = weight / totalWeight weight = weight / totalWeight
c = torch.add(c,model.get_learned_conditioning(subprompts[i]), alpha=weight) c = torch.add(c,model.get_learned_conditioning(subprompts[i]), alpha=weight)
else: # just standard 1 prompt else: # just standard 1 prompt
c = model.get_learned_conditioning(prompts) c = model.get_learned_conditioning(prompts)
# encode (scaled latent) # encode (scaled latent)
z_enc = sampler.stochastic_encode(init_latent, torch.tensor([t_enc]*batch_size).to(self.device)) z_enc = sampler.stochastic_encode(init_latent, torch.tensor([t_enc]*batch_size).to(self.device))
# decode it # decode it
samples = sampler.decode(z_enc, c, t_enc, unconditional_guidance_scale=cfg_scale, samples = sampler.decode(z_enc, c, t_enc, unconditional_guidance_scale=cfg_scale,
unconditional_conditioning=uc,) unconditional_conditioning=uc,)
x_samples = model.decode_first_stage(samples) x_samples = model.decode_first_stage(samples)
x_samples = torch.clamp((x_samples + 1.0) / 2.0, min=0.0, max=1.0) x_samples = torch.clamp((x_samples + 1.0) / 2.0, min=0.0, max=1.0)
if not grid: if not grid:
for x_sample in x_samples: for x_sample in x_samples:
x_sample = 255. * rearrange(x_sample.cpu().numpy(), 'c h w -> h w c') x_sample = 255. * rearrange(x_sample.cpu().numpy(), 'c h w -> h w c')
filename = self._unique_filename(outdir,previousname=filename, filename = self._unique_filename(outdir,previousname=filename,
seed=seed,isbatch=(batch_size>1)) seed=seed,isbatch=(batch_size>1))
assert not os.path.exists(filename) assert not os.path.exists(filename)
Image.fromarray(x_sample.astype(np.uint8)).save(filename) Image.fromarray(x_sample.astype(np.uint8)).save(filename)
images.append([filename,seed]) images.append([filename,seed])
else: else:
all_samples.append(x_samples) all_samples.append(x_samples)
seeds.append(seed) seeds.append(seed)
image_count +=1 image_count +=1
seed = self._new_seed() seed = self._new_seed()
if grid: if grid:
images = self._make_grid(samples=all_samples, images = self._make_grid(samples=all_samples,
seeds=seeds, seeds=seeds,
batch_size=batch_size, batch_size=batch_size,
iterations=iterations, iterations=iterations,
outdir=outdir) outdir=outdir)
except KeyboardInterrupt: except KeyboardInterrupt:
print('*interrupted*') print('*interrupted*')
@ -448,11 +448,13 @@ The vast majority of these arguments default to reasonable values.
seed_everything(self.seed) seed_everything(self.seed)
try: try:
config = OmegaConf.load(self.config) config = OmegaConf.load(self.config)
self.device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu") self.device = torch.device(self.device) if torch.cuda.is_available() else torch.device("cpu")
model = self._load_model_from_config(config,self.weights) model = self._load_model_from_config(config,self.weights)
if self.embedding_path is not None: if self.embedding_path is not None:
model.embedding_manager.load(self.embedding_path) model.embedding_manager.load(self.embedding_path)
self.model = model.to(self.device) self.model = model.to(self.device)
# model.to doesn't change the cond_stage_model.device used to move the tokenizer output, so set it here
self.model.cond_stage_model.device = self.device
except AttributeError: except AttributeError:
raise SystemExit raise SystemExit
@ -489,7 +491,6 @@ The vast majority of these arguments default to reasonable values.
sd = pl_sd["state_dict"] sd = pl_sd["state_dict"]
model = instantiate_from_config(config.model) model = instantiate_from_config(config.model)
m, u = model.load_state_dict(sd, strict=False) m, u = model.load_state_dict(sd, strict=False)
model.cuda()
model.eval() model.eval()
if self.full_precision: if self.full_precision:
print('Using slower but more accurate full-precision math (--full_precision)') print('Using slower but more accurate full-precision math (--full_precision)')

View File

@ -60,7 +60,8 @@ def main():
full_precision=opt.full_precision, full_precision=opt.full_precision,
config=config, config=config,
latent_diffusion_weights=opt.laion400m, # this is solely for recreating the prompt latent_diffusion_weights=opt.laion400m, # this is solely for recreating the prompt
embedding_path=opt.embedding_path embedding_path=opt.embedding_path,
device=opt.device
) )
# make sure the output directory exists # make sure the output directory exists
@ -282,10 +283,14 @@ def create_argv_parser():
type=str, type=str,
default="outputs/img-samples", default="outputs/img-samples",
help="directory in which to place generated images and a log of prompts and seeds") help="directory in which to place generated images and a log of prompts and seeds")
parser.add_argument('--embedding_path', parser.add_argument('--embedding_path',
type=str, type=str,
help="Path to a pre-trained embedding manager checkpoint - can only be set on command line") help="Path to a pre-trained embedding manager checkpoint - can only be set on command line")
parser.add_argument('--device',
'-d',
type=str,
default="cuda",
help="device to run stable diffusion on. defaults to cuda `torch.cuda.current_device()` if avalible")
return parser return parser

1
src/clip Submodule

@ -0,0 +1 @@
Subproject commit d50d76daa670286dd6cacf3bcd80b5e4823fc8e1

1
src/k-diffusion Submodule

@ -0,0 +1 @@
Subproject commit db5799068749bf3a6d5845120ed32df16b7d883b

@ -0,0 +1 @@
Subproject commit 24268930bf1dce879235a7fddd0b2355b84d7ea6