r/stable-diffusion-finetune
1
0
Fork 0
Code Issues Pull requests Projects Releases Wiki Activity

Merge remote-tracking branch 'origin/main'

Browse source
This commit is contained in:
Patrick Esser 2022-07-24 20:52:11 +00:00
commit 95dfb95672
8 changed files with 562 additions and 2 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

144
configs/stable-diffusion/inpainting/v1-finetune-for-inpainting-laion-iaesthe.yaml Normal file
View file

@ -0,0 +1,144 @@
model:
base_learning_rate: 7.5e-05
target: ldm.models.diffusion.ddpm.LatentInpaintDiffusion
params:
linear_start: 0.00085
linear_end: 0.0120
num_timesteps_cond: 1
log_every_t: 200
timesteps: 1000
first_stage_key: "jpg"
cond_stage_key: "txt"
image_size: 64
channels: 4
cond_stage_trainable: false # Note: different from the one we trained before
conditioning_key: hybrid # important
monitor: val/loss_simple_ema
scale_factor: 0.18215
ckpt_path: "/fsx/stable-diffusion/stable-diffusion/checkpoints2/v1pp/v1pp-flatline-pruned.ckpt"
scheduler_config: # 10000 warmup steps
target: ldm.lr_scheduler.LambdaLinearScheduler
params:
warm_up_steps: [ 2500 ] # NOTE for resuming. use 10000 if starting from scratch
cycle_lengths: [ 10000000000000 ] # incredibly large number to prevent corner cases
f_start: [ 1.e-6 ]
f_max: [ 1. ]
f_min: [ 1. ]
unet_config:
target: ldm.modules.diffusionmodules.openaimodel.UNetModel
params:
image_size: 32 # unused
in_channels: 9 # 4 data + 4 downscaled image + 1 mask
out_channels: 4
model_channels: 320
attention_resolutions: [ 4, 2, 1 ]
num_res_blocks: 2
channel_mult: [ 1, 2, 4, 4 ]
num_heads: 8
use_spatial_transformer: True
transformer_depth: 1
context_dim: 768
use_checkpoint: True
legacy: False
first_stage_config:
target: ldm.models.autoencoder.AutoencoderKL
params:
embed_dim: 4
monitor: val/rec_loss
ddconfig:
double_z: true
z_channels: 4
resolution: 256
in_channels: 3
out_ch: 3
ch: 128
ch_mult:
- 1
- 2
- 4
- 4
num_res_blocks: 2
attn_resolutions: []
dropout: 0.0
lossconfig:
target: torch.nn.Identity
cond_stage_config:
target: ldm.modules.encoders.modules.FrozenCLIPEmbedder
data:
target: ldm.data.laion.WebDataModuleFromConfig
params:
tar_base: "pipe:aws s3 cp s3://s-datasets/laion-high-resolution/"
batch_size: 4
num_workers: 4
multinode: True
min_size: 512
train:
shards: '{00000..17279}.tar -'
shuffle: 10000
image_key: jpg
image_transforms:
- target: torchvision.transforms.Resize
params:
size: 512
interpolation: 3
- target: torchvision.transforms.RandomCrop
params:
size: 512
postprocess:
target: ldm.data.laion.AddMask
# NOTE use enough shards to avoid empty validation loops in workers
validation:
shards: '{17280..17535}.tar -'
shuffle: 0
image_key: jpg
image_transforms:
- target: torchvision.transforms.Resize
params:
size: 512
interpolation: 3
- target: torchvision.transforms.CenterCrop
params:
size: 512
postprocess:
target: ldm.data.laion.AddMask
lightning:
find_unused_parameters: False
modelcheckpoint:
params:
every_n_train_steps: 2000
callbacks:
image_logger:
target: main.ImageLogger
params:
disabled: False
batch_frequency: 1000
max_images: 4
increase_log_steps: False
log_first_step: False
log_images_kwargs:
use_ema_scope: False
inpaint: False
plot_progressive_rows: False
plot_diffusion_rows: False
N: 4
unconditional_guidance_scale: 3.0
unconditional_guidance_label: [""]
ddim_steps: 100 # todo check these out for inpainting,
ddim_eta: 1.0 # todo check these out for inpainting,
trainer:
benchmark: True
val_check_interval: 5000000 # really sorry
num_sanity_val_steps: 0
accumulate_grad_batches: 2

19
ldm/data/base.py
View file

@ -1,3 +1,5 @@
import os
import numpy as np
from abc import abstractmethod
from torch.utils.data import Dataset, ConcatDataset, ChainDataset, IterableDataset
@ -20,4 +22,19 @@ class Txt2ImgIterableBaseDataset(IterableDataset):
@abstractmethod
def __iter__(self):
pass
pass
class PRNGMixin(object):
"""
Adds a prng property which is a numpy RandomState which gets
reinitialized whenever the pid changes to avoid synchronized sampling
behavior when used in conjunction with multiprocessing.
"""
@property
def prng(self):
currentpid = os.getpid()
if getattr(self, "_initpid", None) != currentpid:
self._initpid = currentpid
self._prng = np.random.RandomState()
return self._prng

0
ldm/data/inpainting/__init__.py Normal file
View file

150
ldm/data/inpainting/synthetic_mask.py Normal file
View file

@ -0,0 +1,150 @@
from PIL import Image, ImageDraw
import numpy as np
settings = {
"256narrow": {
"p_irr": 1,
"min_n_irr": 4,
"max_n_irr": 50,
"max_l_irr": 40,
"max_w_irr": 10,
"min_n_box": None,
"max_n_box": None,
"min_s_box": None,
"max_s_box": None,
"marg": None,
},
"256train": {
"p_irr": 0.5,
"min_n_irr": 1,
"max_n_irr": 5,
"max_l_irr": 200,
"max_w_irr": 100,
"min_n_box": 1,
"max_n_box": 4,
"min_s_box": 30,
"max_s_box": 150,
"marg": 10,
},
"512train": { # TODO: experimental
"p_irr": 0.5,
"min_n_irr": 1,
"max_n_irr": 5,
"max_l_irr": 450,
"max_w_irr": 250,
"min_n_box": 1,
"max_n_box": 4,
"min_s_box": 30,
"max_s_box": 300,
"marg": 10,
},
}
def gen_segment_mask(mask, start, end, brush_width):
mask = mask > 0
mask = (255 * mask).astype(np.uint8)
mask = Image.fromarray(mask)
draw = ImageDraw.Draw(mask)
draw.line([start, end], fill=255, width=brush_width, joint="curve")
mask = np.array(mask) / 255
return mask
def gen_box_mask(mask, masked):
x_0, y_0, w, h = masked
mask[y_0:y_0 + h, x_0:x_0 + w] = 1
return mask
def gen_round_mask(mask, masked, radius):
x_0, y_0, w, h = masked
xy = [(x_0, y_0), (x_0 + w, y_0 + w)]
mask = mask > 0
mask = (255 * mask).astype(np.uint8)
mask = Image.fromarray(mask)
draw = ImageDraw.Draw(mask)
draw.rounded_rectangle(xy, radius=radius, fill=255)
mask = np.array(mask) / 255
return mask
def gen_large_mask(prng, img_h, img_w,
marg, p_irr, min_n_irr, max_n_irr, max_l_irr, max_w_irr,
min_n_box, max_n_box, min_s_box, max_s_box):
"""
img_h: int, an image height
img_w: int, an image width
marg: int, a margin for a box starting coordinate
p_irr: float, 0 <= p_irr <= 1, a probability of a polygonal chain mask
min_n_irr: int, min number of segments
max_n_irr: int, max number of segments
max_l_irr: max length of a segment in polygonal chain
max_w_irr: max width of a segment in polygonal chain
min_n_box: int, min bound for the number of box primitives
max_n_box: int, max bound for the number of box primitives
min_s_box: int, min length of a box side
max_s_box: int, max length of a box side
"""
mask = np.zeros((img_h, img_w))
uniform = prng.randint
if np.random.uniform(0, 1) < p_irr: # generate polygonal chain
n = uniform(min_n_irr, max_n_irr) # sample number of segments
for _ in range(n):
y = uniform(0, img_h) # sample a starting point
x = uniform(0, img_w)
a = uniform(0, 360) # sample angle
l = uniform(10, max_l_irr) # sample segment length
w = uniform(5, max_w_irr) # sample a segment width
# draw segment starting from (x,y) to (x_,y_) using brush of width w
x_ = x + l * np.sin(a)
y_ = y + l * np.cos(a)
mask = gen_segment_mask(mask, start=(x, y), end=(x_, y_), brush_width=w)
x, y = x_, y_
else: # generate Box masks
n = uniform(min_n_box, max_n_box) # sample number of rectangles
for _ in range(n):
h = uniform(min_s_box, max_s_box) # sample box shape
w = uniform(min_s_box, max_s_box)
x_0 = uniform(marg, img_w - marg - w) # sample upper-left coordinates of box
y_0 = uniform(marg, img_h - marg - h)
if np.random.uniform(0, 1) < 0.5:
mask = gen_box_mask(mask, masked=(x_0, y_0, w, h))
else:
r = uniform(0, 60) # sample radius
mask = gen_round_mask(mask, masked=(x_0, y_0, w, h), radius=r)
return mask
make_lama_mask = lambda prng, h, w: gen_large_mask(prng, h, w,
**settings["256train"])
make_narrow_lama_mask = lambda prng, h, w: gen_large_mask(prng, h, w,
**settings["256narrow"])
make_512_lama_mask = lambda prng, h, w: gen_large_mask(prng, h, w,
**settings["512train"])
if __name__ == "__main__":
import sys
out = sys.argv[1]
prng = np.random.RandomState(1)
kwargs = settings["256train"]
mask = gen_large_mask(prng, 256, 256, **kwargs)
mask = (255 * mask).astype(np.uint8)
mask = Image.fromarray(mask)
mask.save(out)

19
ldm/data/laion.py
View file

@ -16,6 +16,8 @@ from webdataset.handlers import warn_and_continue
from ldm.util import instantiate_from_config
from ldm.data.inpainting.synthetic_mask import gen_large_mask, make_lama_mask, make_narrow_lama_mask, make_512_lama_mask
from ldm.data.base import PRNGMixin
class DataWithWings(torch.utils.data.IterableDataset):
@ -229,6 +231,23 @@ class AddLR(object):
return sample
class AddMask(PRNGMixin):
def __init__(self, size=512):
super().__init__()
self.make_mask = make_512_lama_mask if size == 512 else make_lama_mask
def __call__(self, sample):
# sample['jpg'] is tensor hwc in [-1, 1] at this point
x = sample['jpg']
mask = self.make_mask(self.prng, x.shape[0], x.shape[1])
mask[mask < 0.5] = 0
mask[mask > 0.5] = 1
mask = torch.from_numpy(mask[..., None])
sample['mask'] = mask
sample['masked_image'] = x * (mask < 0.5)
return sample
def example00():
url = "pipe:aws s3 cp s3://s-datasets/laion5b/laion2B-data/000000.tar -"
dataset = wds.WebDataset(url)

163
ldm/models/diffusion/ddpm.py
View file

@ -1260,7 +1260,6 @@ class LatentDiffusion(DDPM):
use_ema_scope=True,
**kwargs):
ema_scope = self.ema_scope if use_ema_scope else nullcontext
use_ddim = ddim_steps is not None
log = dict()
@ -1582,6 +1581,168 @@ class LatentUpscaleDiffusion(LatentDiffusion):
return log
class LatentInpaintDiffusion(LatentDiffusion):
"""
can either run as pure inpainting model (only concat mode) or with mixed conditionings,
e.g. mask as concat and text via cross-attn.
To disable finetuning mode, set finetune_keys to None
"""
def __init__(self,
finetune_keys=("model.diffusion_model.input_blocks.0.0.weight", "model_ema.diffusion_modelinput_blocks00weight"),
concat_keys=("mask", "masked_image"),
masked_image_key="masked_image",
keep_finetune_dims=4, # if model was trained without concat mode before and we would like to keep these channels
*args, **kwargs
):
ckpt_path = kwargs.pop("ckpt_path", None)
ignore_keys = kwargs.pop("ignore_keys", list())
super().__init__(*args, **kwargs)
self.masked_image_key = masked_image_key
assert self.masked_image_key in concat_keys
self.finetune_keys = finetune_keys
self.concat_keys = concat_keys
self.keep_dims = keep_finetune_dims
if exists(self.finetune_keys): assert exists(ckpt_path), 'can only finetune from a given checkpoint'
if exists(ckpt_path):
self.init_from_ckpt(ckpt_path, ignore_keys)
def init_from_ckpt(self, path, ignore_keys=list(), only_model=False):
sd = torch.load(path, map_location="cpu")
if "state_dict" in list(sd.keys()):
sd = sd["state_dict"]
keys = list(sd.keys())
for k in keys:
for ik in ignore_keys:
if k.startswith(ik):
print("Deleting key {} from state_dict.".format(k))
del sd[k]
# make it explicit, finetune by including extra input channels
if exists(self.finetune_keys) and k in self.finetune_keys:
new_entry = None
for name, param in self.named_parameters():
if name in self.finetune_keys:
print(f"modifying key '{name}' and keeping its original {self.keep_dims} (channels) dimensions only")
new_entry = torch.zeros_like(param) # zero init
assert exists(new_entry), 'did not find matching parameter to modify'
new_entry[:, :self.keep_dims, ...] = sd[k]
sd[k] = new_entry
missing, unexpected = self.load_state_dict(sd, strict=False) if not only_model else self.model.load_state_dict(
sd, strict=False)
print(f"Restored from {path} with {len(missing)} missing and {len(unexpected)} unexpected keys")
if len(missing) > 0:
print(f"Missing Keys: {missing}")
if len(unexpected) > 0:
print(f"Unexpected Keys: {unexpected}")
@torch.no_grad()
def get_input(self, batch, k, cond_key=None, bs=None, return_first_stage_outputs=False):
# note: restricted to non-trainable encoders currently
assert not self.cond_stage_trainable, 'trainable cond stages not yet supported for inpaiting'
z, c, x, xrec, xc = super().get_input(batch, self.first_stage_key, return_first_stage_outputs=True,
force_c_encode=True, return_original_cond=True, bs=bs)
assert exists(self.concat_keys)
c_cat = list()
for ck in self.concat_keys:
cc = rearrange(batch[ck], 'b h w c -> b c h w').to(memory_format=torch.contiguous_format).float()
if bs is not None:
cc = cc[:bs]
cc = cc.to(self.device)
bchw = z.shape
if ck != self.masked_image_key:
cc = torch.nn.functional.interpolate(cc, size=bchw[-2:])
else:
cc = self.get_first_stage_encoding(self.encode_first_stage(cc))
c_cat.append(cc)
c_cat = torch.cat(c_cat, dim=1)
all_conds = {"c_concat": [c_cat], "c_crossattn": [c]}
if return_first_stage_outputs:
return z, all_conds, x, xrec, xc
return z, all_conds
@torch.no_grad()
def log_images(self, batch, N=8, n_row=4, sample=True, ddim_steps=200, ddim_eta=1., return_keys=None,
quantize_denoised=True, inpaint=True, plot_denoise_rows=False, plot_progressive_rows=True,
plot_diffusion_rows=True, unconditional_guidance_scale=1., unconditional_guidance_label=None,
use_ema_scope=True,
**kwargs):
ema_scope = self.ema_scope if use_ema_scope else nullcontext
use_ddim = ddim_steps is not None
log = dict()
z, c, x, xrec, xc = self.get_input(batch, self.first_stage_key, bs=N, return_first_stage_outputs=True)
c_cat, c = c["c_concat"][0], c["c_crossattn"][0]
N = min(x.shape[0], N)
n_row = min(x.shape[0], n_row)
log["inputs"] = x
log["reconstruction"] = xrec
if self.model.conditioning_key is not None:
if hasattr(self.cond_stage_model, "decode"):
xc = self.cond_stage_model.decode(c)
log["conditioning"] = xc
elif self.cond_stage_key in ["caption", "txt"]:
xc = log_txt_as_img((x.shape[2], x.shape[3]), batch[self.cond_stage_key], size=x.shape[2] // 25)
log["conditioning"] = xc
elif self.cond_stage_key == 'class_label':
xc = log_txt_as_img((x.shape[2], x.shape[3]), batch["human_label"], size=x.shape[2] // 25)
log['conditioning'] = xc
elif isimage(xc):
log["conditioning"] = xc
if ismap(xc):
log["original_conditioning"] = self.to_rgb(xc)
if plot_diffusion_rows:
# get diffusion row
diffusion_row = list()
z_start = z[:n_row]
for t in range(self.num_timesteps):
if t % self.log_every_t == 0 or t == self.num_timesteps - 1:
t = repeat(torch.tensor([t]), '1 -> b', b=n_row)
t = t.to(self.device).long()
noise = torch.randn_like(z_start)
z_noisy = self.q_sample(x_start=z_start, t=t, noise=noise)
diffusion_row.append(self.decode_first_stage(z_noisy))
diffusion_row = torch.stack(diffusion_row) # n_log_step, n_row, C, H, W
diffusion_grid = rearrange(diffusion_row, 'n b c h w -> b n c h w')
diffusion_grid = rearrange(diffusion_grid, 'b n c h w -> (b n) c h w')
diffusion_grid = make_grid(diffusion_grid, nrow=diffusion_row.shape[0])
log["diffusion_row"] = diffusion_grid
if sample:
# get denoise row
with ema_scope("Sampling"):
samples, z_denoise_row = self.sample_log(cond={"c_concat": [c_cat], "c_crossattn": [c]},
batch_size=N, ddim=use_ddim,
ddim_steps=ddim_steps, eta=ddim_eta)
# samples, z_denoise_row = self.sample(cond=c, batch_size=N, return_intermediates=True)
x_samples = self.decode_first_stage(samples)
log["samples"] = x_samples
if plot_denoise_rows:
denoise_grid = self._get_denoise_row_from_list(z_denoise_row)
log["denoise_row"] = denoise_grid
if unconditional_guidance_scale > 1.0:
uc_cross = self.get_unconditional_conditioning(N, unconditional_guidance_label)
uc_cat = c_cat
uc_full = {"c_concat": [uc_cat], "c_crossattn": [uc_cross]}
with ema_scope("Sampling with classifier-free guidance"):
samples_cfg, _ = self.sample_log(cond={"c_concat": [c_cat], "c_crossattn": [c]},
batch_size=N, ddim=use_ddim,
ddim_steps=ddim_steps, eta=ddim_eta,
unconditional_guidance_scale=unconditional_guidance_scale,
unconditional_conditioning=uc_full,
)
x_samples_cfg = self.decode_first_stage(samples_cfg)
log[f"samples_cfg_scale_{unconditional_guidance_scale:.2f}"] = x_samples_cfg
log["masked_image"] = rearrange(batch["masked_image"],
'b h w c -> b c h w').to(memory_format=torch.contiguous_format).float()
return log
class Layout2ImgDiffusion(LatentDiffusion):
# TODO: move all layout-specific hacks to this class
def __init__(self, cond_stage_key, *args, **kwargs):

27
scripts/slurm/v1_inpainting_improvedaesthetics_torch111/launcher.sh Executable file
View file

@ -0,0 +1,27 @@
#!/bin/bash
# mpi version for node rank
H=`hostname`
THEID=`echo -e $HOSTNAMES | python3 -c "import sys;[sys.stdout.write(str(i)) for i,line in enumerate(next(sys.stdin).split(' ')) if line.strip() == '$H'.strip()]"`
export NODE_RANK=${THEID}
echo THEID=$THEID
echo "##########################################"
echo MASTER_ADDR=${MASTER_ADDR}
echo MASTER_PORT=${MASTER_PORT}
echo NODE_RANK=${NODE_RANK}
echo WORLD_SIZE=${WORLD_SIZE}
echo "##########################################"
# debug environment worked great so we stick with it
# no magic there, just a miniconda python=3.9, pytorch=1.12, cudatoolkit=11.3
# env with pip dependencies from stable diffusion's requirements.txt
eval "$(/fsx/stable-diffusion/debug/miniconda3/bin/conda shell.bash hook)"
conda activate torch111
cd /fsx/robin/stable-diffusion/stable-diffusion
CONFIG="/fsx/robin/stable-diffusion/stable-diffusion/configs/stable-diffusion/inpainting/v1-finetune-for-inpainting-laion-iaesthe.yaml"
# debugging
#EXTRA="${EXTRA} -d True lightning.callbacks.image_logger.params.batch_frequency=50"
python main.py --base $CONFIG --gpus 0,1,2,3,4,5,6,7 -t --num_nodes ${WORLD_SIZE} --scale_lr False #$EXTRA

42
scripts/slurm/v1_inpainting_improvedaesthetics_torch111/sbatch.sh Executable file
View file

@ -0,0 +1,42 @@
#!/bin/bash
#SBATCH --partition=compute-od-gpu
#SBATCH --job-name=stable-diffusion-v1-inpainting-improvedaesthetics-torch111
#SBATCH --nodes 8
#SBATCH --ntasks-per-node 1
#SBATCH --cpus-per-gpu=4
#SBATCH --gres=gpu:8
#SBATCH --exclusive
#SBATCH --output=%x_%j.out
#SBATCH --comment "Key=Monitoring,Value=ON"
module load intelmpi
source /opt/intel/mpi/latest/env/vars.sh
export LD_LIBRARY_PATH=/opt/aws-ofi-nccl/lib:/opt/amazon/efa/lib64:/usr/local/cuda-11.0/efa/lib:/usr/local/cuda-11.0/lib:/usr/local/cuda-11.0/lib64:/usr/local/cuda-11.0:/opt/nccl/build/lib:/opt/aws-ofi-nccl-inst
all/lib:/opt/aws-ofi-nccl/lib:$LD_LIBRARY_PATH
export NCCL_PROTO=simple
export PATH=/opt/amazon/efa/bin:$PATH
export LD_PRELOAD="/opt/nccl/build/lib/libnccl.so"
export FI_EFA_FORK_SAFE=1
export FI_LOG_LEVEL=1
export FI_EFA_USE_DEVICE_RDMA=1 # use for p4dn
export NCCL_DEBUG=info
export PYTHONFAULTHANDLER=1
export CUDA_LAUNCH_BLOCKING=0
export OMPI_MCA_mtl_base_verbose=1
export FI_EFA_ENABLE_SHM_TRANSFER=0
export FI_PROVIDER=efa
export FI_EFA_TX_MIN_CREDITS=64
export NCCL_TREE_THRESHOLD=0
# sent to sub script
export HOSTNAMES=`scontrol show hostnames "$SLURM_JOB_NODELIST"`
export MASTER_ADDR=$(scontrol show hostnames "$SLURM_JOB_NODELIST" | head -n 1)
export MASTER_PORT=12802
export COUNT_NODE=`scontrol show hostnames "$SLURM_JOB_NODELIST" | wc -l`
export WORLD_SIZE=$COUNT_NODE
echo go $COUNT_NODE
echo $HOSTNAMES
echo $WORLD_SIZE
mpirun -n $COUNT_NODE -perhost 1 /fsx/robin/stable-diffusion/stable-diffusion/scripts/slurm/v1_inpainting_improvedaesthetics_torch111/launcher.sh