manbeast3b's picture
Update src/loss.py
0e4c4ae verified
_A=None
count=0
image_count=0
import torch
from tqdm import tqdm
import numpy as np
class ExponentialScalingMatrix(torch.nn.Module):
def __init__(self, size):
super().__init__()
self.base = torch.nn.Parameter(torch.ones(size, size)).to('cuda')
def forward(self, timestep):
return self.base * torch.exp(-torch.tensor(timestep) / 10.0).to('cuda')
class RecurrentDynamicScaling(torch.nn.Module):
def __init__(self, size):
super().__init__()
self.rnn = torch.nn.GRUCell(size, size)
self.state = torch.nn.Parameter(torch.zeros(size)) # Learnable state
def forward(self, timestep, x, correction):
self.state = self.rnn(self.state, correction) # Gradients flow through GRU
scaled_correction = correction * torch.sigmoid(self.state) # Scale dynamically
return x + scaled_correction
class AttentionDynamicScaling(torch.nn.Module):
def __init__(self, size):
super().__init__()
self.attention = torch.nn.MultiheadAttention(size, num_heads=1)
def forward(self, timestep, x, correction):
query = correction.unsqueeze(0) # Shape (1, batch, size)
key = x.unsqueeze(0)
value = correction.unsqueeze(0)
scaled_correction, _ = self.attention(query, key, value)
return x + scaled_correction.squeeze(0)
class LossSchedulerModel(torch.nn.Module):
def __init__(A,wx,we,scale_factor=None,strategy="exponential_scaling"):
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))
size=13
if type(scale_factor)!=None:
A.register_parameter('scale_factor',torch.nn.Parameter(scale_factor))
else:
A.scale_factor = torch.nn.Parameter(torch.ones(size))
A.decay_rate = 0.1
if strategy == "exponential_scaling":
A.matrix_generator = ExponentialScalingMatrix(size)
else:
raise ValueError("Unknown strategy")
def forward_other(A,t,xT,e_prev):
B=e_prev
assert t-len(B)+1==0;C=xT*A.wx[t]
jitter=A.matrix_generator.forward(t)
for(D,E,J)in zip(B,A.we[t],jitter[t]):
dynamic_scale = torch.tanh(A.decay_rate * torch.tensor(t)).to(D.device)
correction=D*(E+J.to(D.device))
C += correction * dynamic_scale
return C.to(xT.dtype) #C.to(xT.dtype)
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, lr=0.01):
A.timesteps=timesteps
A.model=model
A.init_noise_sigma=1.
A.order=1
# A.optimizer = torch.optim.SGD([A.model.scale_factor, A.model.wx, A.model.we], lr=lr)
A.optimizer = torch.optim.SGD([A.model.wx, A.model.we], lr=lr)
A.model.train()
@staticmethod
def load(path):A,B,C,E=torch.load(path,map_location='cpu');D=LossSchedulerModel(B,C,scale_factor=E);return LossScheduler(A,D)
def save(A,path):B,C,D,E=A.timesteps,A.model.wx,A.model.we,A.model.scale_factor;torch.save((B,C,D,E),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
def step_orig(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,
def step(self,model_output,timestep,sample,*D,**E):
global image_count
global count
A=self;B=A.timesteps.tolist().index(timestep);assert A.t_prev==-1 or B==A.t_prev+1; count+=1
if timestep==A.timesteps[0]: print("resetting"); image_count+=1;count=0;
print(timestep)
if A.t_prev==-1:A.xT=sample
A.e_prev.append(model_output)
if timestep==51:
with torch.enable_grad():
orig_output = torch.load(f"/home/mbhat/edge-maxxing/miner/miner/latents/latent_orig_{image_count}_20.pth")
A.optimizer.zero_grad()
C=A.model(B,A.xT,A.e_prev)
L=torch.nn.functional.mse_loss(C, orig_output)
L.backward()
A.optimizer.step(); print("optimized")
else:
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
A.save(f"/home/mbhat/weights/weights_latest_{image_count}_{int(timestep)}.pth")
else:A.t_prev=B
return C,
def step_other(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; count+=1
if A.t_prev==-1:A.xT=sample
A.e_prev.append(model_output)
if timestep==A.timesteps.tolist()[-1]:
with torch.enable_grad():
A.optimizer.zero_grad()
C=A.model(B,A.xT,A.e_prev)
L=torch.nn.functional.mse_loss(C, model_output)
L.backward()
A.optimizer.step()
if B+1==len(A.timesteps):
A.xT=_A;A.e_prev=[];
A.t_prev=-1
# A.save("/home/mbhat/weights.pth")
else:A.t_prev=B
return C,
class SchedulerWrapper:
def __init__(A,scheduler,loss_params_path='/home/mbhat/weights.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=11
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,E=torch.load(A.loss_params_path,map_location='cpu');A.loss_model=LossSchedulerModel(C,D,scale_factor=E);A.loss_scheduler=LossScheduler(B,A.loss_model)
def prepare_loss(A,num_accelerate_steps=15):A.load_loss_params()