src/loss.py

_A=None
import torch
from tqdm import tqdm
import output
class LossSchedulerModel(torch.nn.Module):
	def __init__(A,wx,we):super(LossSchedulerModel,A).__init__();assert len(wx.shape)==1 and len(we.shape)==2;B=wx.shape[0];assert B==we.shape[0]and B==we.shape[1];A.register_parameter('wx',torch.nn.Parameter(wx));A.register_parameter('we',torch.nn.Parameter(we))
	def forward(A,t,xT,e_prev):
		B=e_prev;assert t-len(B)+1==0;C=xT*A.wx[t]
		for(D,E)in zip(B,A.we[t]):C+=D*E
		return C.to(xT.dtype)
class LossScheduler:
	def __init__(A,timesteps,model):A.timesteps=timesteps;A.model=model;A.init_noise_sigma=1.;A.order=1
	@staticmethod
	def load(path):A,B,C=torch.load(path,map_location='cpu');D=LossSchedulerModel(B,C);return LossScheduler(A,D)
	def save(A,path):B,C,D=A.timesteps,A.model.wx,A.model.we;torch.save((B,C,D),path)
	def set_timesteps(A,num_inference_steps,device='cuda'):B=device;A.xT=_A;A.e_prev=[];A.t_prev=-1;A.model=A.model.to(B);A.timesteps=A.timesteps.to(B)
	def scale_model_input(A,sample,*B,**C):return sample
	@torch.no_grad()
	def step(self,model_output,timestep,sample,*D,**E):
		A=self;B=A.timesteps.tolist().index(timestep);assert A.t_prev==-1 or B==A.t_prev+1
		if A.t_prev==-1:A.xT=sample
		A.e_prev.append(model_output);C=A.model(B,A.xT,A.e_prev)
		if B+1==len(A.timesteps):A.xT=_A;A.e_prev=[];A.t_prev=-1
		else:A.t_prev=B
		return C,
class SchedulerWrapper:
	def __init__(A,scheduler,loss_params_path='optimized_loss_params_re_hooked_model.pth'):A.scheduler=scheduler;A.catch_x,A.catch_e,A.catch_x_={},{},{};A.loss_scheduler=_A;A.loss_params_path=loss_params_path
	def set_timesteps(A,num_inference_steps,**C):
		D=num_inference_steps
		if A.loss_scheduler is _A:B=A.scheduler.set_timesteps(D,**C);A.timesteps=A.scheduler.timesteps;A.init_noise_sigma=A.scheduler.init_noise_sigma;A.order=A.scheduler.order;return B
		else:B=A.loss_scheduler.set_timesteps(D,**C);A.timesteps=A.loss_scheduler.timesteps;A.init_noise_sigma=A.scheduler.init_noise_sigma;A.order=A.scheduler.order;return B
	def step(B,model_output,timestep,sample,**F):
		D=sample;E=model_output;A=timestep
		if B.loss_scheduler is _A:
			C=B.scheduler.step(E,A,D,**F);A=A.tolist()
			if A not in B.catch_x:B.catch_x[A]=[];B.catch_e[A]=[];B.catch_x_[A]=[]
			B.catch_x[A].append(D.clone().detach().cpu());B.catch_e[A].append(E.clone().detach().cpu());B.catch_x_[A].append(C[0].clone().detach().cpu());return C
		else:C=B.loss_scheduler.step(E,A,D,**F);return C
	def scale_model_input(A,sample,timestep):return sample
	def add_noise(A,original_samples,noise,timesteps):B=A.scheduler.add_noise(original_samples,noise,timesteps);return B
	def get_path(C):
		A=sorted([A for A in C.catch_x],reverse=True);B,D=[],[]
		for E in A:F=torch.cat(C.catch_x[E],dim=0);B.append(F);G=torch.cat(C.catch_e[E],dim=0);D.append(G)
		H=A[-1];I=torch.cat(C.catch_x_[H],dim=0);B.append(I);A=torch.tensor(A,dtype=torch.int32);B=torch.stack(B);D=torch.stack(D);return A,B,D
	def load_loss_params(A):B,C,D=torch.load(A.loss_params_path,map_location='cpu');A.loss_model=LossSchedulerModel(C,D);A.loss_scheduler=LossScheduler(B,A.loss_model);print("loaded ",A.loss_params_path)
	def prepare_loss(A,num_accelerate_steps=15):A.load_loss_params()
def get_instance(device):
    compress = output.Decoder().to(device).requires_grad_(False)
    compress.load_state_dict(torch.load("taesdxl_decoder.pth", map_location=device, weights_only=True)) # taesdxl_decoder.pth
    return compress
@torch.no_grad()
def hook_pipe(pipe, compress, mul, sub, scaling_factor):
    def compress_machine(magic, *args, **kwargs):
        magic = magic.float().mul(scaling_factor) #pipe.vae.config.scaling_factor
        out_magic = compress(magic).mul_(mul).sub_(sub).cpu()#.mul_(1.2).sub_(0.75).cpu()
        return (out_magic, )
    pipe.vae.decode = compress_machine