github/ai-toolkit
1
0
Fork 0
Code Issues Pull Requests Actions 1 Packages Projects Releases Wiki Activity

Added experimental concept replacer, replicate converter, bucket maker, and other goodies

Browse Source
This commit is contained in:
Jaret Burkett
2023-09-06 18:50:32 -06:00
parent f84500159c
commit 436bf0c6a3
8 changed files with 517 additions and 561 deletions
Download Patch File Download Diff File Expand all files Collapse all files

159
extensions_built_in/concept_replacer/ConceptReplacer.py Normal file
View File

@@ -0,0 +1,159 @@
import random
from collections import OrderedDict
from torch.utils.data import DataLoader
from toolkit.prompt_utils import concat_prompt_embeds, split_prompt_embeds
from toolkit.stable_diffusion_model import StableDiffusion, BlankNetwork
from toolkit.train_tools import get_torch_dtype, apply_snr_weight
import gc
import torch
from jobs.process import BaseSDTrainProcess
def flush():
torch.cuda.empty_cache()
gc.collect()
class ConceptReplacementConfig:
def __init__(self, **kwargs):
self.concept: str = kwargs.get('concept', '')
self.replacement: str = kwargs.get('replacement', '')
class ConceptReplacer(BaseSDTrainProcess):
def __init__(self, process_id: int, job, config: OrderedDict, **kwargs):
super().__init__(process_id, job, config, **kwargs)
replacement_list = self.config.get('replacements', [])
self.replacement_list = [ConceptReplacementConfig(**x) for x in replacement_list]
def before_model_load(self):
pass
def hook_before_train_loop(self):
self.sd.vae.eval()
self.sd.vae.to(self.device_torch)
# textual inversion
if self.embedding is not None:
# keep original embeddings as reference
self.orig_embeds_params = self.sd.text_encoder.get_input_embeddings().weight.data.clone()
# set text encoder to train. Not sure if this is necessary but diffusers example did it
self.sd.text_encoder.train()
def hook_train_loop(self, batch):
with torch.no_grad():
dtype = get_torch_dtype(self.train_config.dtype)
noisy_latents, noise, timesteps, conditioned_prompts, imgs = self.process_general_training_batch(batch)
network_weight_list = batch.get_network_weight_list()
# have a blank network so we can wrap it in a context and set multipliers without checking every time
if self.network is not None:
network = self.network
else:
network = BlankNetwork()
batch_replacement_list = []
# get a random replacement for each prompt
for prompt in conditioned_prompts:
replacement = random.choice(self.replacement_list)
batch_replacement_list.append(replacement)
# build out prompts
concept_prompts = []
replacement_prompts = []
for idx, replacement in enumerate(batch_replacement_list):
prompt = conditioned_prompts[idx]
# insert shuffled concept at beginning and end of prompt
shuffled_concept = [x.strip() for x in replacement.concept.split(',')]
random.shuffle(shuffled_concept)
shuffled_concept = ', '.join(shuffled_concept)
concept_prompts.append(f"{shuffled_concept}, {prompt}, {shuffled_concept}")
# insert replacement at beginning and end of prompt
shuffled_replacement = [x.strip() for x in replacement.replacement.split(',')]
random.shuffle(shuffled_replacement)
shuffled_replacement = ', '.join(shuffled_replacement)
replacement_prompts.append(f"{shuffled_replacement}, {prompt}, {shuffled_replacement}")
# predict the replacement without network
conditional_embeds = self.sd.encode_prompt(replacement_prompts).to(self.device_torch, dtype=dtype)
replacement_pred = self.sd.predict_noise(
latents=noisy_latents.to(self.device_torch, dtype=dtype),
conditional_embeddings=conditional_embeds.to(self.device_torch, dtype=dtype),
timestep=timesteps,
guidance_scale=1.0,
)
del conditional_embeds
replacement_pred = replacement_pred.detach()
self.optimizer.zero_grad()
flush()
# text encoding
grad_on_text_encoder = False
if self.train_config.train_text_encoder:
grad_on_text_encoder = True
if self.embedding:
grad_on_text_encoder = True
# set the weights
network.multiplier = network_weight_list
# activate network if it exits
with network:
with torch.set_grad_enabled(grad_on_text_encoder):
# embed the prompts
conditional_embeds = self.sd.encode_prompt(concept_prompts).to(self.device_torch, dtype=dtype)
if not grad_on_text_encoder:
# detach the embeddings
conditional_embeds = conditional_embeds.detach()
self.optimizer.zero_grad()
flush()
noise_pred = self.sd.predict_noise(
latents=noisy_latents.to(self.device_torch, dtype=dtype),
conditional_embeddings=conditional_embeds.to(self.device_torch, dtype=dtype),
timestep=timesteps,
guidance_scale=1.0,
)
loss = torch.nn.functional.mse_loss(noise_pred.float(), replacement_pred.float(), reduction="none")
loss = loss.mean([1, 2, 3])
if self.train_config.min_snr_gamma is not None and self.train_config.min_snr_gamma > 0.000001:
# add min_snr_gamma
loss = apply_snr_weight(loss, timesteps, self.sd.noise_scheduler, self.train_config.min_snr_gamma)
loss = loss.mean()
# back propagate loss to free ram
loss.backward()
flush()
# apply gradients
self.optimizer.step()
self.optimizer.zero_grad()
self.lr_scheduler.step()
if self.embedding is not None:
# Let's make sure we don't update any embedding weights besides the newly added token
index_no_updates = torch.ones((len(self.sd.tokenizer),), dtype=torch.bool)
index_no_updates[
min(self.embedding.placeholder_token_ids): max(self.embedding.placeholder_token_ids) + 1] = False
with torch.no_grad():
self.sd.text_encoder.get_input_embeddings().weight[
index_no_updates
] = self.orig_embeds_params[index_no_updates]
loss_dict = OrderedDict(
{'loss': loss.item()}
)
# reset network multiplier
network.multiplier = 1.0
return loss_dict

26
extensions_built_in/concept_replacer/__init__.py Normal file
View File

@@ -0,0 +1,26 @@
# This is an example extension for custom training. It is great for experimenting with new ideas.
from toolkit.extension import Extension
# This is for generic training (LoRA, Dreambooth, FineTuning)
class ConceptReplacerExtension(Extension):
# uid must be unique, it is how the extension is identified
uid = "concept_replacer"
# name is the name of the extension for printing
name = "Concept Replacer"
# This is where your process class is loaded
# keep your imports in here so they don't slow down the rest of the program
@classmethod
def get_process(cls):
# import your process class here so it is only loaded when needed and return it
from .ConceptReplacer import ConceptReplacer
return ConceptReplacer
AI_TOOLKIT_EXTENSIONS = [
# you can put a list of extensions here
ConceptReplacerExtension,
]

91
extensions_built_in/concept_replacer/config/train.example.yaml Normal file
View File

@@ -0,0 +1,91 @@
---
job: extension
config:
name: test_v1
process:
- type: 'textual_inversion_trainer'
training_folder: "out/TI"
device: cuda:0
# for tensorboard logging
log_dir: "out/.tensorboard"
embedding:
trigger: "your_trigger_here"
tokens: 12
init_words: "man with short brown hair"
save_format: "safetensors" # 'safetensors' or 'pt'
save:
dtype: float16 # precision to save
save_every: 100 # save every this many steps
max_step_saves_to_keep: 5 # only affects step counts
datasets:
- folder_path: "/path/to/dataset"
caption_ext: "txt"
default_caption: "[trigger]"
buckets: true
resolution: 512
train:
noise_scheduler: "ddpm" # or "ddpm", "lms", "euler_a"
steps: 3000
weight_jitter: 0.0
lr: 5e-5
train_unet: false
gradient_checkpointing: true
train_text_encoder: false
optimizer: "adamw"
# optimizer: "prodigy"
optimizer_params:
weight_decay: 1e-2
lr_scheduler: "constant"
max_denoising_steps: 1000
batch_size: 4
dtype: bf16
xformers: true
min_snr_gamma: 5.0
# skip_first_sample: true
noise_offset: 0.0 # not needed for this
model:
# objective reality v2
name_or_path: "https://civitai.com/models/128453?modelVersionId=142465"
is_v2: false # for v2 models
is_xl: false # for SDXL models
is_v_pred: false # for v-prediction models (most v2 models)
sample:
sampler: "ddpm" # must match train.noise_scheduler
sample_every: 100 # sample every this many steps
width: 512
height: 512
prompts:
- "photo of [trigger] laughing"
- "photo of [trigger] smiling"
- "[trigger] close up"
- "dark scene [trigger] frozen"
- "[trigger] nighttime"
- "a painting of [trigger]"
- "a drawing of [trigger]"
- "a cartoon of [trigger]"
- "[trigger] pixar style"
- "[trigger] costume"
neg: ""
seed: 42
walk_seed: false
guidance_scale: 7
sample_steps: 20
network_multiplier: 1.0
logging:
log_every: 10 # log every this many steps
use_wandb: false # not supported yet
verbose: false
# You can put any information you want here, and it will be saved in the model.
# The below is an example, but you can put your grocery list in it if you want.
# It is saved in the model so be aware of that. The software will include this
# plus some other information for you automatically
meta:
# [name] gets replaced with the name above
name: "[name]"
# version: '1.0'
# creator:
# name: Your Name
# email: your@gmail.com
# website: https://your.website