Added new timestep weighing strategy

This commit is contained in:
Jaret Burkett
2025-06-04 01:16:02 -06:00
parent adc31ec77d
commit 22cdfadab6
8 changed files with 1348 additions and 9 deletions

View File

@@ -501,13 +501,22 @@ class SDTrainer(BaseSDTrainProcess):
loss = wavelet_loss(pred, batch.latents, noise)
else:
loss = torch.nn.functional.mse_loss(pred.float(), target.float(), reduction="none")
do_weighted_timesteps = False
if self.sd.is_flow_matching:
if self.train_config.linear_timesteps or self.train_config.linear_timesteps2:
do_weighted_timesteps = True
if self.train_config.timestep_type == "weighted":
# use the noise scheduler to get the weights for the timesteps
do_weighted_timesteps = True
# handle linear timesteps and only adjust the weight of the timesteps
if self.sd.is_flow_matching and (self.train_config.linear_timesteps or self.train_config.linear_timesteps2):
if do_weighted_timesteps:
# calculate the weights for the timesteps
timestep_weight = self.sd.noise_scheduler.get_weights_for_timesteps(
timesteps,
v2=self.train_config.linear_timesteps2
v2=self.train_config.linear_timesteps2,
timestep_type=self.train_config.timestep_type
).to(loss.device, dtype=loss.dtype)
timestep_weight = timestep_weight.view(-1, 1, 1, 1).detach()
loss = loss * timestep_weight

View File

@@ -0,0 +1,228 @@
import gc
import os, sys
from tqdm import tqdm
import numpy as np
import json
sys.path.append(os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
# set visible devices to 0
# os.environ["CUDA_VISIBLE_DEVICES"] = "0"
# protect from formatting
if True:
import torch
from optimum.quanto import freeze, qfloat8, QTensor, qint4
from diffusers import FluxTransformer2DModel, FluxPipeline, AutoencoderKL, FlowMatchEulerDiscreteScheduler
from toolkit.util.quantize import quantize, get_qtype
from transformers import T5EncoderModel, T5TokenizerFast, CLIPTextModel, CLIPTokenizer
from torchvision import transforms
qtype = "qfloat8"
dtype = torch.bfloat16
# base_model_path = "black-forest-labs/FLUX.1-dev"
base_model_path = "ostris/Flex.1-alpha"
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
print("Loading Transformer...")
prompt = "Photo of a man and a woman in a park, sunny day"
output_root = os.path.join(os.path.dirname(os.path.dirname(os.path.abspath(__file__))), "output")
output_path = os.path.join(output_root, "flex_timestep_weights.json")
img_output_path = os.path.join(output_root, "flex_timestep_weights.png")
quantization_type = get_qtype(qtype)
def flush():
torch.cuda.empty_cache()
gc.collect()
pil_to_tensor = transforms.ToTensor()
with torch.no_grad():
transformer = FluxTransformer2DModel.from_pretrained(
base_model_path,
subfolder='transformer',
torch_dtype=dtype
)
transformer.to(device, dtype=dtype)
print("Quantizing Transformer...")
quantize(transformer, weights=quantization_type)
freeze(transformer)
flush()
print("Loading Scheduler...")
scheduler = FlowMatchEulerDiscreteScheduler.from_pretrained(base_model_path, subfolder="scheduler")
print("Loading Autoencoder...")
vae = AutoencoderKL.from_pretrained(base_model_path, subfolder="vae", torch_dtype=dtype)
vae.to(device, dtype=dtype)
flush()
print("Loading Text Encoder...")
tokenizer_2 = T5TokenizerFast.from_pretrained(base_model_path, subfolder="tokenizer_2", torch_dtype=dtype)
text_encoder_2 = T5EncoderModel.from_pretrained(base_model_path, subfolder="text_encoder_2", torch_dtype=dtype)
text_encoder_2.to(device, dtype=dtype)
print("Quantizing Text Encoder...")
quantize(text_encoder_2, weights=get_qtype(qtype))
freeze(text_encoder_2)
flush()
print("Loading CLIP")
text_encoder = CLIPTextModel.from_pretrained(base_model_path, subfolder="text_encoder", torch_dtype=dtype)
tokenizer = CLIPTokenizer.from_pretrained(base_model_path, subfolder="tokenizer", torch_dtype=dtype)
text_encoder.to(device, dtype=dtype)
print("Making pipe")
pipe: FluxPipeline = FluxPipeline(
scheduler=scheduler,
text_encoder=text_encoder,
tokenizer=tokenizer,
text_encoder_2=None,
tokenizer_2=tokenizer_2,
vae=vae,
transformer=None,
)
pipe.text_encoder_2 = text_encoder_2
pipe.transformer = transformer
pipe.to(device, dtype=dtype)
print("Encoding prompt...")
prompt_embeds, pooled_prompt_embeds, text_ids = pipe.encode_prompt(
prompt,
prompt_2=prompt,
device=device
)
generator = torch.manual_seed(42)
height = 1024
width = 1024
print("Generating image...")
# Fix a bug in diffusers/torch
def callback_on_step_end(pipe, i, t, callback_kwargs):
latents = callback_kwargs["latents"]
if latents.dtype != dtype:
latents = latents.to(dtype)
return {"latents": latents}
img = pipe(
prompt_embeds=prompt_embeds,
pooled_prompt_embeds=pooled_prompt_embeds,
height=height,
width=height,
num_inference_steps=30,
guidance_scale=3.5,
generator=generator,
callback_on_step_end=callback_on_step_end,
).images[0]
img.save(img_output_path)
print(f"Image saved to {img_output_path}")
print("Encoding image...")
# img is a PIL image. convert it to a -1 to 1 tensor
img = pil_to_tensor(img)
img = img.unsqueeze(0) # add batch dimension
img = img * 2 - 1 # convert to -1 to 1 range
img = img.to(device, dtype=dtype)
latents = vae.encode(img).latent_dist.sample()
shift = vae.config['shift_factor'] if vae.config['shift_factor'] is not None else 0
latents = vae.config['scaling_factor'] * (latents - shift)
num_channels_latents = pipe.transformer.config.in_channels // 4
l_height = 2 * (int(height) // (pipe.vae_scale_factor * 2))
l_width = 2 * (int(width) // (pipe.vae_scale_factor * 2))
packed_latents = pipe._pack_latents(latents, 1, num_channels_latents, l_height, l_width)
packed_latents, latent_image_ids = pipe.prepare_latents(
1,
num_channels_latents,
height,
width,
prompt_embeds.dtype,
device,
generator,
packed_latents,
)
print("Calculating timestep weights...")
torch.manual_seed(8675309)
noise = torch.randn_like(packed_latents, device=device, dtype=dtype)
# Create linear timesteps from 1000 to 0
num_train_timesteps = 1000
timesteps_torch = torch.linspace(1000, 1, num_train_timesteps, device='cpu')
timesteps = np.linspace(1, num_train_timesteps, num_train_timesteps, dtype=np.float32)[::-1].copy()
timesteps = torch.from_numpy(timesteps).to(dtype=torch.float32)
timestep_weights = torch.zeros(num_train_timesteps, dtype=torch.float32, device=device)
guidance = torch.full([1], 1.0, device=device, dtype=torch.float32)
guidance = guidance.expand(latents.shape[0])
pbar = tqdm(range(num_train_timesteps), desc="loss: 0.000000 scaler: 0.0000")
for i in pbar:
timestep = timesteps[i:i+1].to(device)
t_01 = (timestep / 1000).to(device)
t_01 = t_01.reshape(-1, 1, 1)
noisy_latents = (1.0 - t_01) * packed_latents + t_01 * noise
noise_pred = pipe.transformer(
hidden_states=noisy_latents, # torch.Size([1, 4096, 64])
timestep=timestep / 1000,
guidance=guidance,
pooled_projections=pooled_prompt_embeds,
encoder_hidden_states=prompt_embeds,
txt_ids=text_ids,
img_ids=latent_image_ids,
return_dict=False,
)[0]
target = noise - packed_latents
loss = torch.nn.functional.mse_loss(noise_pred.float(), target.float())
loss = loss
# determine scaler to multiply loss by to make it 1
scaler = 1.0 / (loss + 1e-6)
timestep_weights[i] = scaler
pbar.set_description(f"loss: {loss.item():.6f} scaler: {scaler.item():.4f}")
print("normalizing timestep weights...")
# normalize the timestep weights so they are a mean of 1.0
timestep_weights = timestep_weights / timestep_weights.mean()
timestep_weights = timestep_weights.cpu().numpy().tolist()
print("Saving timestep weights...")
with open(output_path, 'w') as f:
json.dump(timestep_weights, f)
print(f"Timestep weights saved to {output_path}")
print("Done!")
flush()

View File

@@ -437,7 +437,7 @@ class TrainConfig:
# adds an additional loss to the network to encourage it output a normalized standard deviation
self.target_norm_std = kwargs.get('target_norm_std', None)
self.target_norm_std_value = kwargs.get('target_norm_std_value', 1.0)
self.timestep_type = kwargs.get('timestep_type', 'sigmoid') # sigmoid, linear, lognorm_blend, next_sample
self.timestep_type = kwargs.get('timestep_type', 'sigmoid') # sigmoid, linear, lognorm_blend, next_sample, weighted
self.next_sample_timesteps = kwargs.get('next_sample_timesteps', 8)
self.linear_timesteps = kwargs.get('linear_timesteps', False)
self.linear_timesteps2 = kwargs.get('linear_timesteps2', False)

View File

@@ -1773,6 +1773,97 @@ class LatentCachingMixin:
self.sd.restore_device_state()
class TextEmbeddingCachingMixin:
def __init__(self: 'AiToolkitDataset', **kwargs):
# if we have super, call it
if hasattr(super(), '__init__'):
super().__init__(**kwargs)
self.is_caching_text_embeddings = self.dataset_config.cache_text_embeddings
def cache_text_embeddings(self: 'AiToolkitDataset'):
with accelerator.main_process_first():
print_acc(f"Caching text_embeddings for {self.dataset_path}")
# cache all latents to disk
to_disk = self.is_caching_latents_to_disk
to_memory = self.is_caching_latents_to_memory
print_acc(" - Saving text embeddings to disk")
# move sd items to cpu except for vae
self.sd.set_device_state_preset('cache_latents')
# use tqdm to show progress
i = 0
for file_item in tqdm(self.file_list, desc=f'Caching latents{" to disk" if to_disk else ""}'):
# set latent space version
if self.sd.model_config.latent_space_version is not None:
file_item.latent_space_version = self.sd.model_config.latent_space_version
elif self.sd.is_xl:
file_item.latent_space_version = 'sdxl'
elif self.sd.is_v3:
file_item.latent_space_version = 'sd3'
elif self.sd.is_auraflow:
file_item.latent_space_version = 'sdxl'
elif self.sd.is_flux:
file_item.latent_space_version = 'flux1'
elif self.sd.model_config.is_pixart_sigma:
file_item.latent_space_version = 'sdxl'
else:
file_item.latent_space_version = self.sd.model_config.arch
file_item.is_caching_to_disk = to_disk
file_item.is_caching_to_memory = to_memory
file_item.latent_load_device = self.sd.device
latent_path = file_item.get_latent_path(recalculate=True)
# check if it is saved to disk already
if os.path.exists(latent_path):
if to_memory:
# load it into memory
state_dict = load_file(latent_path, device='cpu')
file_item._encoded_latent = state_dict['latent'].to('cpu', dtype=self.sd.torch_dtype)
else:
# not saved to disk, calculate
# load the image first
file_item.load_and_process_image(self.transform, only_load_latents=True)
dtype = self.sd.torch_dtype
device = self.sd.device_torch
# add batch dimension
try:
imgs = file_item.tensor.unsqueeze(0).to(device, dtype=dtype)
latent = self.sd.encode_images(imgs).squeeze(0)
except Exception as e:
print_acc(f"Error processing image: {file_item.path}")
print_acc(f"Error: {str(e)}")
raise e
# save_latent
if to_disk:
state_dict = OrderedDict([
('latent', latent.clone().detach().cpu()),
])
# metadata
meta = get_meta_for_safetensors(file_item.get_latent_info_dict())
os.makedirs(os.path.dirname(latent_path), exist_ok=True)
save_file(state_dict, latent_path, metadata=meta)
if to_memory:
# keep it in memory
file_item._encoded_latent = latent.to('cpu', dtype=self.sd.torch_dtype)
del imgs
del latent
del file_item.tensor
# flush(garbage_collect=False)
file_item.is_latent_cached = True
i += 1
# flush every 100
# if i % 100 == 0:
# flush()
# restore device state
self.sd.restore_device_state()
class CLIPCachingMixin:
def __init__(self: 'AiToolkitDataset', **kwargs):
# if we have super, call it

View File

@@ -4,6 +4,7 @@ from torch.distributions import LogNormal
from diffusers import FlowMatchEulerDiscreteScheduler
import torch
import numpy as np
from toolkit.timestep_weighing.default_weighing_scheme import default_weighing_scheme
def calculate_shift(
@@ -47,20 +48,26 @@ class CustomFlowMatchEulerDiscreteScheduler(FlowMatchEulerDiscreteScheduler):
hbsmntw_weighing[num_timesteps //
2:] = hbsmntw_weighing[num_timesteps // 2:].max()
# Create linear timesteps from 1000 to 0
timesteps = torch.linspace(1000, 0, num_timesteps, device='cpu')
# Create linear timesteps from 1000 to 1
timesteps = torch.linspace(1000, 1, num_timesteps, device='cpu')
self.linear_timesteps = timesteps
self.linear_timesteps_weights = bsmntw_weighing
self.linear_timesteps_weights2 = hbsmntw_weighing
pass
def get_weights_for_timesteps(self, timesteps: torch.Tensor, v2=False) -> torch.Tensor:
def get_weights_for_timesteps(self, timesteps: torch.Tensor, v2=False, timestep_type="linear") -> torch.Tensor:
# Get the indices of the timesteps
step_indices = [(self.timesteps == t).nonzero().item()
for t in timesteps]
# Get the weights for the timesteps
if timestep_type == "weighted":
weights = torch.tensor(
[default_weighing_scheme[i] for i in step_indices],
device=timesteps.device,
dtype=timesteps.dtype
)
if v2:
weights = self.linear_timesteps_weights2[step_indices].flatten()
else:
@@ -106,8 +113,8 @@ class CustomFlowMatchEulerDiscreteScheduler(FlowMatchEulerDiscreteScheduler):
patch_size=1
):
self.timestep_type = timestep_type
if timestep_type == 'linear':
timesteps = torch.linspace(1000, 0, num_timesteps, device=device)
if timestep_type == 'linear' or timestep_type == 'weighted':
timesteps = torch.linspace(1000, 1, num_timesteps, device=device)
self.timesteps = timesteps
return timesteps
elif timestep_type == 'sigmoid':
@@ -198,7 +205,7 @@ class CustomFlowMatchEulerDiscreteScheduler(FlowMatchEulerDiscreteScheduler):
t1 = ((1 - t1/t1.max()) * 1000)
# add half of linear
t2 = torch.linspace(1000, 0, int(
t2 = torch.linspace(1000, 1, int(
num_timesteps * (1 - alpha)), device=device)
timesteps = torch.cat((t1, t2))

View File

File diff suppressed because it is too large Load Diff

Binary file not shown.

After

Width:  |  Height:  |  Size: 190 KiB