import gc
import html
import math
import os
import random
from typing import List, Literal, Optional, Union
import ftfy
import regex as re
import torch
from accelerate.logging import get_logger
logger = get_logger(__name__)
NORM_LAYER_PREFIXES = ["norm_q", "norm_k", "norm_added_q", "norm_added_k"]
# ======================================== memory monitoring ========================================
def get_memory_stats():
if torch.cuda.is_available():
allocated = torch.cuda.memory_allocated() / 1024**3 # GB
reserved = torch.cuda.memory_reserved() / 1024**3 # GB
max_allocated = torch.cuda.max_memory_allocated() / 1024**3
return {"allocated": allocated, "reserved": reserved, "max_allocated": max_allocated}
return None
def reset_memory_stats():
if torch.cuda.is_available():
torch.cuda.empty_cache()
torch.cuda.reset_peak_memory_stats()
gc.collect()
# ======================================== initialize ========================================
def get_config_value(args, name):
if hasattr(args, name):
return getattr(args, name)
elif hasattr(args, "training_config") and hasattr(args.training_config, name):
return getattr(args.training_config, name)
else:
raise AttributeError(f"Neither args nor args.training_config has attribute '{name}'")
def compare_configs(existing_conf, current_conf, path="", ignore_keys=None):
if ignore_keys is None:
ignore_keys = set()
mismatches = []
all_keys = set(existing_conf.keys()) | set(current_conf.keys())
for key in all_keys:
current_path = f"{path}.{key}" if path else key
if current_path in ignore_keys or key in ignore_keys:
continue
if key not in existing_conf:
mismatches.append(f"Key '{current_path}' missing in existing config")
elif key not in current_conf:
mismatches.append(f"Key '{current_path}' missing in current config")
else:
existing_val = existing_conf[key]
current_val = current_conf[key]
if isinstance(existing_val, dict) and isinstance(current_val, dict):
mismatches.extend(compare_configs(existing_val, current_val, current_path, ignore_keys))
elif existing_val != current_val:
mismatches.append(f"Key '{current_path}': existing={existing_val} vs current={current_val}")
return mismatches
def get_optimizer(args, accelerator, params_to_optimize, use_deepspeed: bool = False):
# Use DeepSpeed optimizer
if use_deepspeed:
from accelerate.utils import DummyOptim
return DummyOptim(
params_to_optimize,
lr=args.training_config.learning_rate,
betas=(args.training_config.adam_beta1, args.training_config.adam_beta2),
eps=args.training_config.adam_epsilon,
weight_decay=args.training_config.adam_weight_decay,
)
# Optimizer creation
supported_optimizers = ["adam", "adamw", "prodigy"]
if args.training_config.optimizer.lower() not in supported_optimizers:
accelerator.print(
f"Unsupported choice of optimizer: {args.training_config.optimizer}. Supported optimizers include {supported_optimizers}. Defaulting to AdamW"
)
args.training_config.optimizer = "adamw"
if args.training_config.use_8bit_adam and args.training_config.optimizer.lower() not in ["adam", "adamw"]:
accelerator.print(
f"use_8bit_adam is ignored when optimizer is not set to 'AdamW'. Optimizer was "
f"set to {args.training_config.optimizer.lower()}"
)
if args.training_config.use_8bit_adam:
try:
import bitsandbytes as bnb
except ImportError:
raise ImportError(
"To use 8-bit Adam, please install the bitsandbytes library: `pip install bitsandbytes`."
)
if args.training_config.optimizer.lower() == "adamw":
optimizer_class = bnb.optim.AdamW8bit if args.training_config.use_8bit_adam else torch.optim.AdamW
optimizer = optimizer_class(
params_to_optimize,
betas=(args.training_config.adam_beta1, args.training_config.adam_beta2),
eps=args.training_config.adam_epsilon,
weight_decay=args.training_config.adam_weight_decay,
)
elif args.training_config.optimizer.lower() == "adam":
optimizer_class = bnb.optim.Adam8bit if args.training_config.use_8bit_adam else torch.optim.Adam
optimizer = optimizer_class(
params_to_optimize,
betas=(args.training_config.adam_beta1, args.training_config.adam_beta2),
eps=args.training_config.adam_epsilon,
weight_decay=args.training_config.adam_weight_decay,
)
elif args.training_config.optimizer.lower() == "prodigy":
try:
import prodigyopt
except ImportError:
raise ImportError("To use Prodigy, please install the prodigyopt library: `pip install prodigyopt`")
optimizer_class = prodigyopt.Prodigy
if args.training_config.learning_rate <= 0.1:
accelerator.print(
"Learning rate is too low. When using prodigy, it's generally better to set learning rate around 1.0"
)
optimizer = optimizer_class(
params_to_optimize,
betas=(args.training_config.adam_beta1, args.training_config.adam_beta2),
beta3=args.training_config.prodigy_beta3,
weight_decay=args.training_config.adam_weight_decay,
eps=args.training_config.adam_epsilon,
decouple=args.training_config.prodigy_decouple,
use_bias_correction=args.training_config.prodigy_use_bias_correction,
safeguard_warmup=args.training_config.prodigy_safeguard_warmup,
)
return optimizer
# ======================================== checkpoints related ========================================
def save_extra_components(args, model=None, model_state_dict=None, output_dir=None):
if model is None and model_state_dict is None:
raise ValueError("Either 'model' or 'model_state_dict' must be provided")
if output_dir is None:
raise ValueError("output_dir must be provided")
os.makedirs(output_dir, exist_ok=True)
state_dict = {}
# Determine whether to use model or model_state_dict
use_state_dict = model_state_dict is not None
# 1. Save patch_short, patch_mid, patch_long (formerly multi_term_memory_patchg)
if args.training_config.is_enable_stage1 and (
args.training_config.is_train_full_multi_term_memory_patchg
or args.training_config.is_train_lora_multi_term_memory_patchg
):
patch_names = ["patch_short", "patch_mid", "patch_long"]
if use_state_dict:
# Extract from state_dict
for k, v in model_state_dict.items():
if any(k.startswith(f"{p}.") for p in patch_names):
state_dict[k] = v.detach().clone().cpu() if torch.is_tensor(v) else v
else:
# Extract from model
for p in patch_names:
if hasattr(model, p):
patch_module = getattr(model, p)
for k, v in patch_module.state_dict().items():
state_dict[f"{p}.{k}"] = v.detach().clone().cpu()
# 2. Save LoRA layers from all transformer blocks
if args.training_config.restrict_self_attn and args.training_config.is_train_restrict_lora:
if use_state_dict:
# Extract LoRA parameters from state_dict
for k, v in model_state_dict.items():
if any(lora_key in k for lora_key in [".q_loras.", ".k_loras.", ".v_loras."]):
state_dict[k] = v.detach().clone().cpu() if torch.is_tensor(v) else v
else:
# Extract from model
for block_idx, block in enumerate(model.blocks):
if hasattr(block.attn1, "q_loras"):
for k, v in block.attn1.q_loras.state_dict().items():
state_dict[f"blocks.{block_idx}.attn1.q_loras.{k}"] = v.detach().clone().cpu()
if hasattr(block.attn1, "k_loras"):
for k, v in block.attn1.k_loras.state_dict().items():
state_dict[f"blocks.{block_idx}.attn1.k_loras.{k}"] = v.detach().clone().cpu()
if hasattr(block.attn1, "v_loras"):
for k, v in block.attn1.v_loras.state_dict().items():
state_dict[f"blocks.{block_idx}.attn1.v_loras.{k}"] = v.detach().clone().cpu()
# 3. Save History Scale parameters
if args.training_config.is_amplify_history:
if use_state_dict:
# Extract history_key_scale from state_dict
for k, v in model_state_dict.items():
if "history_key_scale" in k:
state_dict[k] = v.detach().clone().cpu() if torch.is_tensor(v) else v
else:
# Extract from model
for block_idx, block in enumerate(model.blocks):
if hasattr(block.attn1, "history_key_scale"):
state_dict[f"blocks.{block_idx}.attn1.history_key_scale"] = (
block.attn1.history_key_scale.detach().clone().cpu()
)
# 4. Save GAN parameters
if args.training_config.is_use_gan:
if use_state_dict:
# Extract GAN parameters from state_dict
for k, v in model_state_dict.items():
if k.startswith("gan_heads.") or k.startswith("gan_final_head."):
state_dict[k] = v.detach().clone().cpu() if torch.is_tensor(v) else v
else:
# Extract from model
if hasattr(model, "gan_heads"):
for hook_name, gan_head in model.gan_heads.items():
for k, v in gan_head.state_dict().items():
state_dict[f"gan_heads.{hook_name}.{k}"] = v.detach().clone().cpu()
if hasattr(model, "gan_final_head"):
for k, v in model.gan_final_head.state_dict().items():
state_dict[f"gan_final_head.{k}"] = v.detach().clone().cpu()
torch.save(state_dict, os.path.join(output_dir, "transformer_partial.pth"))
print(f"Saved checkpoint with {len(state_dict)} parameters to {output_dir}/transformer_partial.pth")
def load_extra_components(args, model, checkpoint_path):
"""
Load patch_short, patch_mid, patch_long, q_loras, k_loras, v_loras into the model
"""
state_dict = torch.load(checkpoint_path, map_location="cpu")
loaded_keys = set()
# Load patch modules (formerly multi_term_memory_patchg)
if args.training_config.is_enable_stage1:
patch_names = ["patch_short", "patch_mid", "patch_long"]
for p_name in patch_names:
patch_keys_in_sd = [k for k in state_dict.keys() if k.startswith(f"{p_name}.")]
if patch_keys_in_sd and hasattr(model, p_name):
patch_state = {
k.replace(f"{p_name}.", ""): v for k, v in state_dict.items() if k.startswith(f"{p_name}.")
}
patch_module = getattr(model, p_name)
load_info = patch_module.load_state_dict(patch_state, strict=False)
loaded_keys.update(patch_keys_in_sd)
print(f"Loaded {len(patch_keys_in_sd)} parameters for {p_name}")
if load_info.missing_keys:
print(f" Missing keys in {p_name}: {load_info.missing_keys}")
if load_info.unexpected_keys:
print(f" Unexpected keys in {p_name}: {load_info.unexpected_keys}")
# Load LoRA layers
lora_keys_count = 0
if args.training_config.restrict_self_attn:
for block_idx, block in enumerate(model.blocks):
# Load q_loras
q_lora_keys_in_sd = [k for k in state_dict.keys() if k.startswith(f"blocks.{block_idx}.attn1.q_loras.")]
if q_lora_keys_in_sd:
q_lora_state = {
k.replace(f"blocks.{block_idx}.attn1.q_loras.", ""): v
for k, v in state_dict.items()
if k.startswith(f"blocks.{block_idx}.attn1.q_loras.")
}
load_info = block.attn1.q_loras.load_state_dict(q_lora_state, strict=False)
loaded_keys.update(q_lora_keys_in_sd)
lora_keys_count += len(q_lora_keys_in_sd)
if load_info.missing_keys:
print(f" Missing keys in blocks.{block_idx}.attn1.q_loras: {load_info.missing_keys}")
if load_info.unexpected_keys:
print(f" Unexpected keys in blocks.{block_idx}.attn1.q_loras: {load_info.unexpected_keys}")
# Load k_loras
k_lora_keys_in_sd = [k for k in state_dict.keys() if k.startswith(f"blocks.{block_idx}.attn1.k_loras.")]
if k_lora_keys_in_sd:
k_lora_state = {
k.replace(f"blocks.{block_idx}.attn1.k_loras.", ""): v
for k, v in state_dict.items()
if k.startswith(f"blocks.{block_idx}.attn1.k_loras.")
}
load_info = block.attn1.k_loras.load_state_dict(k_lora_state, strict=False)
loaded_keys.update(k_lora_keys_in_sd)
lora_keys_count += len(k_lora_keys_in_sd)
if load_info.missing_keys:
print(f" Missing keys in blocks.{block_idx}.attn1.k_loras: {load_info.missing_keys}")
if load_info.unexpected_keys:
print(f" Unexpected keys in blocks.{block_idx}.attn1.k_loras: {load_info.unexpected_keys}")
# Load v_loras
v_lora_keys_in_sd = [k for k in state_dict.keys() if k.startswith(f"blocks.{block_idx}.attn1.v_loras.")]
if v_lora_keys_in_sd:
v_lora_state = {
k.replace(f"blocks.{block_idx}.attn1.v_loras.", ""): v
for k, v in state_dict.items()
if k.startswith(f"blocks.{block_idx}.attn1.v_loras.")
}
load_info = block.attn1.v_loras.load_state_dict(v_lora_state, strict=False)
loaded_keys.update(v_lora_keys_in_sd)
lora_keys_count += len(v_lora_keys_in_sd)
if load_info.missing_keys:
print(f" Missing keys in blocks.{block_idx}.attn1.v_loras: {load_info.missing_keys}")
if load_info.unexpected_keys:
print(f" Unexpected keys in blocks.{block_idx}.attn1.v_loras: {load_info.unexpected_keys}")
print(f"Loaded {lora_keys_count} parameters for Restrict Self Attn LoRA")
# Load History Scale layers
history_keys_count = 0
if args.training_config.is_amplify_history:
for block_idx, block in enumerate(model.blocks):
history_key_scale_key = f"blocks.{block_idx}.attn1.history_key_scale"
if history_key_scale_key in state_dict:
block.attn1.history_key_scale.data = state_dict[history_key_scale_key].to(
block.attn1.history_key_scale.device
)
loaded_keys.add(history_key_scale_key)
history_keys_count += 1
print(f"Loaded {history_keys_count} parameters for History Scale")
# Load GAN
gan_keys_count = 0
if args.training_config.is_use_gan:
# Load intermediate gan_heads
if hasattr(model, "gan_heads"):
for hook_name, gan_head in model.gan_heads.items():
gan_head_prefix = f"gan_heads.{hook_name}."
gan_head_keys_in_sd = [k for k in state_dict.keys() if k.startswith(gan_head_prefix)]
if gan_head_keys_in_sd:
gan_head_state = {
k.replace(gan_head_prefix, ""): v
for k, v in state_dict.items()
if k.startswith(gan_head_prefix)
}
load_info = gan_head.load_state_dict(gan_head_state, strict=False)
loaded_keys.update(gan_head_keys_in_sd)
gan_keys_count += len(gan_head_keys_in_sd)
if load_info.missing_keys:
print(f" Missing keys in gan_heads.{hook_name}: {load_info.missing_keys}")
if load_info.unexpected_keys:
print(f" Unexpected keys in gan_heads.{hook_name}: {load_info.unexpected_keys}")
# Load final gan head
if hasattr(model, "gan_final_head"):
gan_final_keys_in_sd = [k for k in state_dict.keys() if k.startswith("gan_final_head.")]
if gan_final_keys_in_sd:
gan_final_state = {
k.replace("gan_final_head.", ""): v
for k, v in state_dict.items()
if k.startswith("gan_final_head.")
}
load_info = model.gan_final_head.load_state_dict(gan_final_state, strict=False)
loaded_keys.update(gan_final_keys_in_sd)
gan_keys_count += len(gan_final_keys_in_sd)
if load_info.missing_keys:
print(f" Missing keys in gan_final_head: {load_info.missing_keys}")
if load_info.unexpected_keys:
print(f" Unexpected keys in gan_final_head: {load_info.unexpected_keys}")
if gan_keys_count > 0:
print(f"Loaded {gan_keys_count} parameters for GAN components")
if not loaded_keys:
print("No extra components were loaded from the checkpoint.")
return
all_sd_keys = set(state_dict.keys())
unmatched_keys = all_sd_keys - loaded_keys
print("\nCheckpoint loading completed.")
print(f"Total loaded keys: {len(loaded_keys)}")
if unmatched_keys:
print(f"The following keys in the checkpoint were not loaded into the model: {sorted(unmatched_keys)}\n")
else:
print("Load extra module successfully! All keys in the checkpoint were successfully processed or matched.\n")
def save_model_checkpoint(
transformer,
args,
save_path,
weight_dtype=None,
unwrap_model_fn=None,
get_peft_model_state_dict_fn=None,
collate_lora_metadata_fn=None,
save_extra_components_fn=None,
pipeline_class=None,
norm_layer_prefixes=None,
):
modules_to_save = {}
model_to_save = unwrap_model_fn(transformer) if unwrap_model_fn else transformer
transformer_lora_layers = get_peft_model_state_dict_fn(model_to_save)
if args.model_config.train_norm_layers:
norm_prefixes = norm_layer_prefixes or []
transformer_norm_layers = {
f"transformer.{name}": param
for name, param in model_to_save.named_parameters()
if any(k in name for k in norm_prefixes)
}
transformer_lora_layers = {
**transformer_lora_layers,
**transformer_norm_layers,
}
modules_to_save["transformer"] = model_to_save
if pipeline_class and hasattr(pipeline_class, "save_lora_weights"):
lora_metadata = collate_lora_metadata_fn(modules_to_save) if collate_lora_metadata_fn else {}
pipeline_class.save_lora_weights(
save_directory=save_path,
transformer_lora_layers=transformer_lora_layers,
**lora_metadata,
)
if save_extra_components_fn:
save_extra_components_fn(args=args, model=model_to_save, output_dir=save_path)
modules_to_save = None
lora_metadata = None
transformer_norm_layers = None
transformer_lora_layers = None
del modules_to_save
del lora_metadata
del transformer_norm_layers
del transformer_lora_layers
def load_model_checkpoint(
args,
checkpoint_path,
transformer,
pipeline_class=None,
norm_layer_prefixes=None,
convert_unet_state_dict_to_peft_fn=None,
set_peft_model_state_dict_fn=None,
cast_training_params_fn=None,
):
if not os.path.exists(checkpoint_path):
raise ValueError(f"Checkpoint path does not exist: {checkpoint_path}")
lora_state_dict = None
if pipeline_class and hasattr(pipeline_class, "load_lora_weights"):
lora_state_dict = pipeline_class.lora_state_dict(checkpoint_path)
transformer_state_dict = {
f"{k.replace('transformer.', '')}": v for k, v in lora_state_dict.items() if k.startswith("transformer.")
}
transformer_state_dict = convert_unet_state_dict_to_peft_fn(transformer_state_dict)
incompatible_keys = set_peft_model_state_dict_fn(transformer, transformer_state_dict, adapter_name="default")
if incompatible_keys is not None:
unexpected_keys = getattr(incompatible_keys, "unexpected_keys", None)
if unexpected_keys:
print(
f"Loading adapter weights from state_dict led to unexpected keys not found in the model: "
f" {unexpected_keys}. "
)
print(f"load lora from {checkpoint_path} successfully!")
if args.model_config.train_norm_layers and lora_state_dict and norm_layer_prefixes:
transformer_norm_state_dict = {
k: v
for k, v in lora_state_dict.items()
if k.startswith("transformer.") and any(norm_k in k for norm_k in norm_layer_prefixes)
}
transformer._transformer_norm_layers = pipeline_class._load_norm_into_transformer(
transformer_norm_state_dict,
transformer=transformer,
discard_original_layers=False,
)
load_extra_components(args, transformer, os.path.join(checkpoint_path, "transformer_partial.pth"))
if args.training_config.mixed_precision != "fp32":
models = [transformer]
cast_training_params_fn(models)
# ======================================== sigmas & timesteps ========================================
def get_sigmas(noise_scheduler, timesteps, n_dim=4, device="cuda", dtype=torch.float32):
sigmas = noise_scheduler.sigmas.to(device=device, dtype=dtype)
schedule_timesteps = noise_scheduler.timesteps.to(device)
timesteps = timesteps.to(device)
step_indices = [(schedule_timesteps == t).nonzero().item() for t in timesteps]
sigma = sigmas[step_indices].flatten()
while len(sigma.shape) < n_dim:
sigma = sigma.unsqueeze(-1)
return sigma
def calculate_shift(
image_seq_len,
base_seq_len: int = 256,
max_seq_len: int = 4096,
base_shift: float = 0.5,
max_shift: float = 1.15,
):
m = (max_shift - base_shift) / (max_seq_len - base_seq_len)
b = base_shift - m * base_seq_len
mu = image_seq_len * m + b
return mu
def apply_schedule_shift(
sigmas,
noise,
sigmas_two=None,
base_seq_len: int = 256,
max_seq_len: int = 4096,
base_shift: float = 0.5,
max_shift: float = 1.15,
exp_max: float = 7.0,
time_shift_type: Literal["exponential", "linear"] = "linear",
mu: float = None,
return_mu: bool = False,
):
if mu is None:
# Resolution-dependent shifting of timestep schedules as per section 5.3.2 of SD3 paper
image_seq_len = (noise.shape[-1] * noise.shape[-2] * noise.shape[-3]) // 4 # patch size 1,2,2
mu = calculate_shift(
image_seq_len,
base_seq_len if base_seq_len is not None else 256,
max_seq_len if max_seq_len is not None else 4096,
base_shift if base_shift is not None else 0.5,
max_shift if max_shift is not None else 1.15,
)
if time_shift_type == "exponential":
mu = min(mu, math.log(exp_max))
mu = math.exp(mu)
if sigmas_two is not None:
sigmas = (sigmas * mu) / (1 + (mu - 1) * sigmas)
sigmas_two = (sigmas_two * mu) / (1 + (mu - 1) * sigmas_two)
if return_mu:
return sigmas, sigmas_two, mu
else:
return sigmas, sigmas_two
else:
sigmas = (sigmas * mu) / (1 + (mu - 1) * sigmas)
if return_mu:
return sigmas, mu
else:
return sigmas
# ======================================== clean prompt ========================================
def basic_clean(text):
text = ftfy.fix_text(text)
text = html.unescape(html.unescape(text))
return text.strip()
def whitespace_clean(text):
text = re.sub(r"\s+", " ", text)
text = text.strip()
return text
def prompt_clean(text):
text = whitespace_clean(basic_clean(text))
return text
def _get_t5_prompt_embeds(
tokenizer,
text_encoder,
prompt: Union[str, List[str]] = None,
num_videos_per_prompt: int = 1,
max_sequence_length: int = 512,
caption_dropout_p: float = 0.0,
device: Optional[torch.device] = "cuda",
dtype: Optional[torch.dtype] = torch.bfloat16,
):
device = device
dtype = dtype
prompt = [prompt] if isinstance(prompt, str) else prompt
prompt = [prompt_clean(u) for u in prompt]
batch_size = len(prompt)
text_inputs = tokenizer(
prompt,
padding="max_length",
max_length=max_sequence_length,
truncation=True,
add_special_tokens=True,
return_attention_mask=True,
return_tensors="pt",
)
text_input_ids, mask = text_inputs.input_ids, text_inputs.attention_mask
prompt_embeds = text_encoder(text_input_ids.to(device), mask.to(device)).last_hidden_state
prompt_embeds = prompt_embeds.to(dtype=dtype, device=device)
if random.random() < caption_dropout_p:
prompt_embeds.fill_(0)
mask.fill_(False)
seq_lens = mask.gt(0).sum(dim=1).long()
prompt_embeds = [u[:v] for u, v in zip(prompt_embeds, seq_lens)]
prompt_embeds = torch.stack(
[torch.cat([u, u.new_zeros(max_sequence_length - u.size(0), u.size(1))]) for u in prompt_embeds], dim=0
)
# duplicate text embeddings for each generation per prompt, using mps friendly method
_, seq_len, _ = prompt_embeds.shape
prompt_embeds = prompt_embeds.repeat(1, num_videos_per_prompt, 1)
prompt_embeds = prompt_embeds.view(batch_size * num_videos_per_prompt, seq_len, -1)
return prompt_embeds, text_inputs.attention_mask
def encode_prompt(
tokenizer,
text_encoder,
prompt: Union[str, List[str]],
num_videos_per_prompt: int = 1,
prompt_embeds: Optional[torch.Tensor] = None,
max_sequence_length: int = 512,
caption_dropout_p: float = 0.0,
device: Optional[torch.device] = "cuda",
dtype: Optional[torch.dtype] = torch.bfloat16,
):
prompt = [prompt] if isinstance(prompt, str) else prompt
if prompt_embeds is None:
prompt_embeds, prompt_attention_mask = _get_t5_prompt_embeds(
tokenizer,
text_encoder,
prompt=prompt,
num_videos_per_prompt=num_videos_per_prompt,
max_sequence_length=max_sequence_length,
caption_dropout_p=caption_dropout_p,
device=device,
dtype=dtype,
)
return prompt_embeds, prompt_attention_mask
# ======================================== other techniques ========================================
class AdaptiveAntiDrifting:
def __init__(
self,
rho_mu: float = 0.9,
rho_sigma: float = 0.9,
delta_mu: float = 0.15,
delta_sigma: float = 0.15,
device: torch.device = None,
dtype: torch.dtype = torch.float32,
):
"""
Args:
rho_mu: EMA coefficient for mean (momentum parameter)
rho_sigma: EMA coefficient for variance (momentum parameter)
delta_mu: Threshold for mean drift detection
delta_sigma: Threshold for variance drift detection
device: Device for tensor operations
dtype: Data type for tensors
"""
self.rho_mu = rho_mu
self.rho_sigma = rho_sigma
self.delta_mu = delta_mu
self.delta_sigma = delta_sigma
self.device = device
self.dtype = dtype
# Global statistics (initialized on first chunk)
self.global_mean = None
self.global_var = None
self.is_initialized = False
def compute_latent_statistics(self, latent_chunk: torch.Tensor) -> tuple:
# Shape: (B, C, T, H, W) -> (B, C)
mean = latent_chunk.mean(dim=[2, 3, 4])
var = latent_chunk.var(dim=[2, 3, 4])
return mean, var
def update_global_statistics(self, current_mean: torch.Tensor, current_var: torch.Tensor):
if not self.is_initialized:
self.global_mean = current_mean.clone()
self.global_var = current_var.clone()
self.is_initialized = True
else:
self.global_mean = self.rho_mu * self.global_mean + (1 - self.rho_mu) * current_mean
self.global_var = self.rho_sigma * self.global_var + (1 - self.rho_sigma) * current_var
def detect_drift(self, current_mean: torch.Tensor, current_var: torch.Tensor) -> bool:
if not self.is_initialized:
return False
mean_drift = torch.norm(current_mean - self.global_mean, p=2, dim=-1).mean().item()
var_drift = torch.norm(current_var - self.global_var, p=2, dim=-1).mean().item()
has_drift = (mean_drift > self.delta_mu) and (var_drift > self.delta_sigma)
return has_drift
def apply_frame_aware_corruption(
self,
history_latents: torch.Tensor,
corruption_strength: float = 0.1,
generator: Optional[torch.Generator] = None,
) -> torch.Tensor:
noise = torch.randn_like(history_latents, generator=generator, device=history_latents.device)
corrupted_latents = history_latents + corruption_strength * noise
return corrupted_latents
def reset(self):
self.global_mean = None
self.global_var = None
self.is_initialized = False