lune-leco-adapters / flow_leco_trainer.py
AbstractPhil's picture
way better trainer
a59941b verified
"""
LECO Attribute Binding Trainer - COMPLETE WITH PROPER FLOW MATCHING
Complete script with correct flow matching SNR and velocity prediction
"""
import os
import json
import datetime
import random
from dataclasses import dataclass, asdict, field
from typing import List, Tuple
from tqdm.auto import tqdm
from itertools import product
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.utils.tensorboard import SummaryWriter
from safetensors.torch import save_file
from diffusers import UNet2DConditionModel
from transformers import CLIPTextModel, CLIPTokenizer
from huggingface_hub import hf_hub_download
# ============================================================================
# DATA STRUCTURES
# ============================================================================
@dataclass(frozen=True)
class AttributePair:
"""A specific combination that should stay distinct"""
attr1: str
attr2: str
negatives: Tuple[str, ...] = ()
weight: float = 1.0
@dataclass
class AttributeBindingConfig:
"""Config for attribute binding training"""
output_dir: str = "./leco_outputs"
base_model_repo: str = "AbstractPhil/sd15-flow-lune-flux"
base_checkpoint: str = "sd15_flow_flux_t2_6_pose_t4_6_port_t1_4_s18765.pt"
name_prefix: str = "leco"
attribute_pairs: List[AttributePair] = field(default_factory=list)
lora_rank: int = 8
lora_alpha: float = 1.0
training_method: str = "xattn"
seed: int = 42
iterations: int = 500
save_every: int = 250
lr: float = 2e-4
pairs_per_batch: int = 4
negatives_per_positive: int = 2
# Min-SNR parameters
use_min_snr: bool = True
min_snr_gamma: float = 5.0
# Flow matching parameters
shift: float = 2.5
min_timestep: float = 0.0
max_timestep: float = 1000.0
resolution: int = 512
@dataclass
class LECOConfig:
"""Minimal config for LoRA creation"""
lora_rank: int = 4
lora_alpha: float = 1.0
training_method: str = "xattn"
# ============================================================================
# ATTRIBUTE COMBINATION HELPERS
# ============================================================================
def extract_color(text: str) -> str:
"""Extract color from text"""
colors = [
"red", "blue", "green", "yellow", "purple", "orange", "pink",
"black", "white", "brown", "blonde", "silver", "gold", "cyan",
"magenta", "teal", "lavender", "gray", "grey", "beige", "navy",
"maroon", "turquoise", "violet", "indigo", "crimson"
]
text_lower = text.lower()
for color in colors:
if color in text_lower:
return color
return None
def generate_smart_negatives(attr1: str, attr2: str, all_negatives: List[str] = None) -> List[str]:
"""Automatically generate wrong combinations"""
negatives = []
color1 = extract_color(attr1)
color2 = extract_color(attr2)
if color1 and color2 and color1 != color2:
swapped_attr1 = attr1.replace(color1, color2)
swapped_attr2 = attr2.replace(color2, color1)
negatives.append(f"{swapped_attr1}, {swapped_attr2}")
negatives.append(f"{attr1}, {attr2.replace(color2, color1)}")
negatives.append(f"{attr1.replace(color1, color2)}, {attr2}")
# Add universal negatives to combinations
if all_negatives:
for neg in all_negatives:
negatives.append(f"{attr1}, {attr2}, {neg}")
return list(set(negatives))
def create_attribute_combinations(
pair_attr1: List[str],
pair_attr2: List[str],
negatives: List[str] = None,
weight: float = 1.0,
auto_generate_negatives: bool = True
) -> List[AttributePair]:
"""Create all combinations of two attribute lists"""
pairs = []
for attr1, attr2 in product(pair_attr1, pair_attr2):
if auto_generate_negatives:
neg_list = generate_smart_negatives(attr1, attr2, negatives)
else:
neg_list = []
if negatives:
for neg in negatives:
neg_list.append(f"{attr1}, {neg}")
neg_list.append(f"{neg}, {attr2}")
pairs.append(AttributePair(
attr1=attr1,
attr2=attr2,
negatives=tuple(neg_list),
weight=weight
))
return pairs
def combine_attribute_groups(*groups: List[AttributePair]) -> List[AttributePair]:
"""Combine multiple attribute groups"""
combined = []
for group in groups:
combined.extend(group)
return combined
# ============================================================================
# LORA UTILITIES
# ============================================================================
def get_target_modules(training_method: str) -> List[str]:
"""Get layer names to inject LoRA"""
attn1 = ["attn1.to_q", "attn1.to_k", "attn1.to_v", "attn1.to_out.0"]
attn2 = ["attn2.to_q", "attn2.to_k", "attn2.to_v", "attn2.to_out.0"]
method_map = {
"full": attn1 + attn2,
"selfattn": attn1,
"xattn": attn2,
"noxattn": attn1,
"innoxattn": attn2
}
return method_map.get(training_method, attn1 + attn2)
def create_lora_layers(unet: nn.Module, config: LECOConfig):
"""Create LoRA layers"""
target_modules = get_target_modules(config.training_method)
lora_state = {}
trainable_params = []
def get_lora_key(module_path: str) -> str:
return f"lora_unet_{module_path.replace('.', '_')}"
print(f"Creating LoRA layers (method: {config.training_method})...")
for name, module in unet.named_modules():
if not any(target in name for target in target_modules):
continue
if not isinstance(module, nn.Linear):
continue
lora_key = get_lora_key(name)
in_dim = module.in_features
out_dim = module.out_features
rank = config.lora_rank
lora_down = nn.Parameter(torch.zeros(rank, in_dim))
lora_up = nn.Parameter(torch.zeros(out_dim, rank))
nn.init.kaiming_uniform_(lora_down, a=1.0)
nn.init.zeros_(lora_up)
lora_state[f"{lora_key}.lora_down.weight"] = lora_down
lora_state[f"{lora_key}.lora_up.weight"] = lora_up
lora_state[f"{lora_key}.alpha"] = torch.tensor(config.lora_alpha)
lora_state[f"{lora_key}._module"] = module
trainable_params.extend([lora_down, lora_up])
print(f"✓ Created {len(trainable_params)//2} LoRA layers ({len(trainable_params)} parameters)")
return lora_state, trainable_params
def apply_lora_hooks(unet: nn.Module, lora_state: dict, scale: float = 1.0) -> list:
"""Apply LoRA using forward hooks"""
handles = []
for key in lora_state:
if not key.endswith(".lora_down.weight"):
continue
base_key = key.replace(".lora_down.weight", "")
module = lora_state[f"{base_key}._module"]
lora_down = lora_state[f"{base_key}.lora_down.weight"]
lora_up = lora_state[f"{base_key}.lora_up.weight"]
alpha = lora_state[f"{base_key}.alpha"].item()
rank = lora_down.shape[0]
scaling = (alpha / rank) * scale
def make_hook(down, up, s):
def forward_hook(mod, inp, out):
x = inp[0]
lora_out = F.linear(F.linear(x, down), up)
return out + lora_out * s
return forward_hook
handle = module.register_forward_hook(make_hook(lora_down, lora_up, scaling))
handles.append(handle)
return handles
def remove_lora_hooks(handles: list):
"""Remove all LoRA hooks"""
for handle in handles:
handle.remove()
# ============================================================================
# TRAINING LOSS WITH PROPER FLOW MATCHING
# ============================================================================
def compute_attribute_binding_loss_batched(
unet,
lora_state,
positive_pairs: List[AttributePair],
tokenizer,
text_encoder,
config: AttributeBindingConfig,
device: str = "cuda"
):
"""Batched attribute binding with PROPER FLOW MATCHING"""
# 1. Sample sigma with constrained range (matching your training code)
min_sigma = config.min_timestep / 1000.0
max_sigma = config.max_timestep / 1000.0
sigma = torch.rand(1, device=device)
sigma = min_sigma + sigma * (max_sigma - min_sigma) # Constrain to range
# Apply shift transformation
sigma = (config.shift * sigma) / (1 + (config.shift - 1) * sigma)
timestep = sigma * 1000.0
sigma_expanded = sigma.view(1, 1, 1, 1)
# 2. Flow matching: x_t = sigma * noise + (1 - sigma) * x_0
# For LECO: we use pure noise as x_0 (no clean latents available)
noise = torch.randn(1, 4, config.resolution // 8, config.resolution // 8, device=device)
noisy_input = sigma_expanded * noise # Simplified since x_0 = 0 (centered)
# Build prompts
positive_prompts = []
negative_prompts = []
pair_weights = []
for pair in positive_pairs:
correct = f"{pair.attr1}, {pair.attr2}"
positive_prompts.append(correct)
pair_weights.append(pair.weight)
if pair.negatives:
sampled_negs = random.sample(
list(pair.negatives),
min(config.negatives_per_positive, len(pair.negatives))
)
negative_prompts.extend(sampled_negs)
if not positive_prompts:
return torch.tensor(0.0, device=device), {
"positive_loss": 0, "negative_loss": 0,
"positive_count": 0, "negative_count": 0,
"timestep": 0.0, "snr_weight": 1.0
}
neutral_prompt = ""
all_prompts = [neutral_prompt] + positive_prompts + negative_prompts
text_inputs = tokenizer(
all_prompts,
padding="max_length",
max_length=tokenizer.model_max_length,
truncation=True,
return_tensors="pt"
).to(device)
all_embeddings = text_encoder(text_inputs.input_ids)[0]
neutral_emb = all_embeddings[0:1]
positive_embs = all_embeddings[1:1+len(positive_prompts)]
negative_embs = all_embeddings[1+len(positive_prompts):]
batch_size = len(all_prompts) - 1
noisy_input_batch = noisy_input.repeat(batch_size, 1, 1, 1)
timestep_batch = timestep.repeat(batch_size)
combined_embs = torch.cat([positive_embs, negative_embs], dim=0)
# Get VELOCITY predictions
with torch.no_grad():
vel_neutral = unet(
noisy_input, timestep_batch[0:1],
encoder_hidden_states=neutral_emb,
return_dict=False
)[0]
vel_baseline = unet(
noisy_input_batch, timestep_batch,
encoder_hidden_states=combined_embs,
return_dict=False
)[0]
vel_positive_baseline = vel_baseline[:len(positive_prompts)]
vel_negative_baseline = vel_baseline[len(positive_prompts):]
handles = apply_lora_hooks(unet, lora_state, scale=1.0)
try:
vel_with_lora = unet(
noisy_input_batch, timestep_batch,
encoder_hidden_states=combined_embs,
return_dict=False
)[0]
finally:
remove_lora_hooks(handles)
vel_positive_lora = vel_with_lora[:len(positive_prompts)]
vel_negative_lora = vel_with_lora[len(positive_prompts):]
# 3. Compute FLOW MATCHING SNR (not DDPM)
snr_weight = 1.0
if config.use_min_snr:
# Flow matching SNR: ((1 - sigma)^2) / (sigma^2)
sigma_sq = sigma.squeeze() ** 2
snr = ((1 - sigma.squeeze()) ** 2) / (sigma_sq + 1e-8)
# Min-SNR clamping
snr_clamped = torch.minimum(snr, torch.tensor(config.min_snr_gamma, device=device))
snr_weight_tensor = snr_clamped / snr
# Velocity prediction adjustment: divide by (SNR + 1)
snr_weight_tensor = snr_weight_tensor / (snr + 1)
snr_weight = snr_weight_tensor.item()
else:
snr_weight_tensor = torch.ones(1, device=device)
# Compute losses
vel_neutral_expanded = vel_neutral.expand_as(vel_positive_baseline)
target_positive_direction = vel_positive_baseline - vel_neutral_expanded
lora_positive_delta = vel_positive_lora - vel_positive_baseline
positive_loss_per_sample = F.mse_loss(
lora_positive_delta,
target_positive_direction * 0.3,
reduction='none'
).mean(dim=(1,2,3))
# Apply both pair weights and SNR weights
pair_weights_tensor = torch.tensor(pair_weights, device=device)
weighted_positive_loss = (positive_loss_per_sample * pair_weights_tensor * snr_weight_tensor).mean()
negative_loss = torch.tensor(0.0, device=device)
lora_negative_norm = 0.0
if len(negative_prompts) > 0:
vel_neutral_expanded_neg = vel_neutral.expand_as(vel_negative_baseline)
target_negative_direction = vel_neutral_expanded_neg - vel_negative_baseline
lora_negative_delta = vel_negative_lora - vel_negative_baseline
negative_loss = F.mse_loss(lora_negative_delta, target_negative_direction * 0.2, reduction='mean')
negative_loss = negative_loss * snr_weight_tensor
lora_negative_norm = lora_negative_delta.norm().item()
total_loss = weighted_positive_loss + negative_loss * 0.5
metrics = {
"positive_loss": weighted_positive_loss.item(),
"negative_loss": negative_loss.item() if isinstance(negative_loss, torch.Tensor) else 0.0,
"positive_count": len(positive_prompts),
"negative_count": len(negative_prompts),
"timestep": timestep.item(),
"sigma": sigma.item(),
"snr_weight": snr_weight,
"lora_positive_norm": lora_positive_delta.norm().item(),
"lora_negative_norm": lora_negative_norm
}
return total_loss, metrics
# ============================================================================
# TRAINING FUNCTION
# ============================================================================
def train_attribute_binding(config: AttributeBindingConfig):
"""Fast training for attribute binding with Min-SNR"""
device = "cuda"
torch.manual_seed(config.seed)
if not config.attribute_pairs:
raise ValueError("No attribute pairs specified!")
pairs_with_negatives = sum(1 for p in config.attribute_pairs if p.negatives)
print(f"Pairs with explicit negatives: {pairs_with_negatives}/{len(config.attribute_pairs)}")
timestamp = datetime.datetime.now().strftime("%Y%m%d_%H%M%S")
output_dir = os.path.join(config.output_dir, f"attribute_binding_{timestamp}")
os.makedirs(output_dir, exist_ok=True)
writer = SummaryWriter(log_dir=output_dir, flush_secs=60)
with open(os.path.join(output_dir, "config.json"), "w") as f:
json.dump(asdict(config), f, indent=2)
print("="*80)
print("ATTRIBUTE BINDING TRAINING")
if config.use_min_snr:
print(f"Using Min-SNR Weighting (gamma={config.min_snr_gamma})")
print("="*80)
# VERIFY UNET LOADING
print("\nVerifying UNet loading...")
print("Loading base SD1.5 UNet for comparison...")
unet_base = UNet2DConditionModel.from_pretrained(
"runwayml/stable-diffusion-v1-5",
subfolder="unet",
torch_dtype=torch.float32
).to(device)
# Create test inputs
test_latents = torch.randn(1, 4, 64, 64, device=device)
test_timestep = torch.tensor([500], device=device)
test_encoder = torch.randn(1, 77, 768, device=device)
with torch.no_grad():
baseline_out = unet_base(test_latents, test_timestep, encoder_hidden_states=test_encoder, return_dict=False)[0]
print(f"Baseline output norm: {baseline_out.norm().item():.6f}")
del unet_base
torch.cuda.empty_cache()
print("\nLoading Lune flow-matching model...")
checkpoint_path = hf_hub_download(
repo_id=config.base_model_repo,
filename=config.base_checkpoint,
repo_type="model"
)
checkpoint = torch.load(checkpoint_path, map_location="cpu")
unet = UNet2DConditionModel.from_pretrained(
"runwayml/stable-diffusion-v1-5",
subfolder="unet",
torch_dtype=torch.float32
)
student_dict = checkpoint["student"]
cleaned_dict = {k[5:] if k.startswith("unet.") else k: v for k, v in student_dict.items()}
missing, unexpected = unet.load_state_dict(cleaned_dict, strict=False)
print(f"Missing keys: {len(missing)}, Unexpected keys: {len(unexpected)}")
unet = unet.to(device)
unet.requires_grad_(False)
unet.eval()
# Verify Lune loaded correctly
with torch.no_grad():
lune_out = unet(test_latents, test_timestep, encoder_hidden_states=test_encoder, return_dict=False)[0]
print(f"Lune output norm: {lune_out.norm().item():.6f}")
diff = (lune_out - baseline_out).abs().mean().item()
print(f"Difference from baseline: {diff:.6f}")
if diff < 1e-4:
print("⚠️ WARNING: Outputs are nearly identical - checkpoint may not have loaded!")
else:
print("✓ Lune checkpoint loaded correctly (outputs differ)")
print("\nLoading CLIP...")
tokenizer = CLIPTokenizer.from_pretrained(
"runwayml/stable-diffusion-v1-5", subfolder="tokenizer"
)
text_encoder = CLIPTextModel.from_pretrained(
"runwayml/stable-diffusion-v1-5", subfolder="text_encoder",
torch_dtype=torch.float32
).to(device)
text_encoder.requires_grad_(False)
text_encoder.eval()
print("✓ Loaded CLIP")
print(f"\nCreating LoRA (rank={config.lora_rank})...")
leco_config = LECOConfig(
lora_rank=config.lora_rank,
lora_alpha=config.lora_alpha,
training_method=config.training_method
)
lora_state, trainable_params = create_lora_layers(unet, leco_config)
print(f"Moving LoRA parameters to {device}...")
for param in trainable_params:
param.data = param.data.to(device)
for key, value in lora_state.items():
if isinstance(value, torch.Tensor) and not isinstance(value, nn.Parameter):
lora_state[key] = value.to(device)
optimizer = torch.optim.AdamW(trainable_params, lr=config.lr, weight_decay=0.01)
print(f"\nTraining Configuration:")
print(f" Attribute pairs: {len(config.attribute_pairs)}")
for i, pair in enumerate(config.attribute_pairs[:3], 1):
print(f" {i}. {pair.attr1} + {pair.attr2} (weight: {pair.weight})")
if pair.negatives:
print(f" Negatives: {len(pair.negatives)} total")
if len(config.attribute_pairs) > 3:
print(f" ... and {len(config.attribute_pairs)-3} more")
print(f"\n Iterations: {config.iterations}")
print(f" Pairs per batch: {config.pairs_per_batch}")
print(f" Negatives per positive: {config.negatives_per_positive}")
print(f" Learning rate: {config.lr}")
print("="*80 + "\n")
progress = tqdm(range(config.iterations), desc="Training")
for step in progress:
sampled_pairs = random.sample(
config.attribute_pairs,
min(config.pairs_per_batch, len(config.attribute_pairs))
)
loss, metrics = compute_attribute_binding_loss_batched(
unet, lora_state,
sampled_pairs,
tokenizer, text_encoder,
config,
device
)
loss.backward()
grad_norm = torch.nn.utils.clip_grad_norm_(trainable_params, max_norm=1.0)
optimizer.step()
optimizer.zero_grad()
writer.add_scalar("loss/total", loss.item(), step)
writer.add_scalar("loss/positive", metrics["positive_loss"], step)
writer.add_scalar("loss/negative", metrics["negative_loss"], step)
writer.add_scalar("grad_norm", grad_norm.item(), step)
writer.add_scalar("snr_weight", metrics["snr_weight"], step)
progress.set_postfix({
"loss": f"{loss.item():.4f}",
"pos": f"{metrics['positive_loss']:.3f}",
"neg": f"{metrics['negative_loss']:.3f}",
"snr": f"{metrics['snr_weight']:.2f}",
"grad": f"{grad_norm.item():.3f}"
})
if (step + 1) % config.save_every == 0 or step == config.iterations - 1:
save_dict = {}
for key, value in lora_state.items():
if isinstance(value, torch.Tensor) and not key.endswith("._module"):
save_dict[key] = value.detach().cpu()
metadata = {
"ss_network_module": "networks.lora",
"ss_network_dim": str(config.lora_rank),
"ss_network_alpha": str(config.lora_alpha),
"ss_training_method": config.training_method,
"leco_action": "attribute_binding",
"leco_num_pairs": str(len(config.attribute_pairs)),
"leco_step": str(step + 1),
"leco_min_snr": str(config.use_min_snr),
"leco_min_snr_gamma": str(config.min_snr_gamma)
}
filename = f"{config.name_prefix}_r{config.lora_rank}_s{step+1}.safetensors"
filepath = os.path.join(output_dir, filename)
save_file(save_dict, filepath, metadata=metadata)
print(f"\n✓ Saved: {filename}")
writer.close()
print("\n" + "="*80)
print("✅ Training complete!")
print(f"Output: {output_dir}")
print("="*80)
return output_dir
if __name__ == "__main__":
# Example 1: Hair + Clothes colors
universal_negs = ["ugly, duplicate, morbid, mutilated, blurry, fuzzy, out of frame, gross"]
hair_colors = ["red hair", "blue hair", "green hair"]
clothes = ["red shirt", "blue shirt", "green shirt"]
hair_clothes_pairs = create_attribute_combinations(
pair_attr1=hair_colors,
pair_attr2=clothes,
negatives=universal_negs,
weight=1.0,
auto_generate_negatives=True
)
print(f"Generated {len(hair_clothes_pairs)} hair+clothes pairs")
# Training config
config = AttributeBindingConfig(
name_prefix="color_clothes_test",
attribute_pairs=hair_clothes_pairs,
iterations=5000,
lora_rank=16,
lr=2e-4,
pairs_per_batch=4,
negatives_per_positive=3,
training_method="xattn",
save_every=250,
# Flow matching parameters
shift=2.5,
min_timestep=0.0,
max_timestep=1000.0,
# Min-SNR enabled
use_min_snr=True,
min_snr_gamma=5.0
)
train_attribute_binding(config)