debug-noise

VADER-VideoCrafter/lvdm/models/samplers/ddim.py

@@ -5,18 +5,9 @@ import torch
 from lvdm.models.utils_diffusion import make_ddim_sampling_parameters, make_ddim_timesteps
 from lvdm.common import noise_like
 import random
-import os
 # import ipdb
 # st = ipdb.set_trace
 
-def seed_everything_self(TORCH_SEED):
-	random.seed(TORCH_SEED)
-	os.environ['PYTHONHASHSEED'] = str(TORCH_SEED)
-	np.random.seed(TORCH_SEED)
-	torch.manual_seed(TORCH_SEED)
-	torch.cuda.manual_seed_all(TORCH_SEED)
-	torch.backends.cudnn.deterministic = True
-	torch.backends.cudnn.benchmark = False
 
 class DDIMSampler(object):
     def __init__(self, model, schedule="linear", **kwargs):
@@ -97,7 +88,6 @@ class DDIMSampler(object):
                log_every_t=100,
                unconditional_guidance_scale=1.,
                unconditional_conditioning=None,
-               seed=0,
                # this has to come in the same format as the conditioning, # e.g. as encoded tokens, ...
                **kwargs
                ):
@@ -143,7 +133,6 @@ class DDIMSampler(object):
                                                     unconditional_guidance_scale=unconditional_guidance_scale,
                                                     unconditional_conditioning=unconditional_conditioning,
                                                     verbose=verbose,
-                                                    seed=seed,
                                                     **kwargs)
         return samples, intermediates
 
@@ -154,11 +143,10 @@ class DDIMSampler(object):
                       mask=None, x0=None, img_callback=None, log_every_t=100,
                       temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None,
                       unconditional_guidance_scale=1., unconditional_conditioning=None, verbose=True,
-                      cond_tau=1., target_size=None, start_timesteps=None, seed=0,
+                      cond_tau=1., target_size=None, start_timesteps=None,
                       **kwargs):
         device = self.model.betas.device        
         # print('ddim device', device)
-        seed_everything_self(seed)
         b = shape[0]
         if x_T is None:
             img = torch.randn(shape, device=device)
@@ -168,8 +156,8 @@ class DDIMSampler(object):
         print("x_T: ", x_T)
         print("shape: ", shape)
         print('random seed debug: ', torch.randn(100, device=device).sum())
-        print("Debug initial noise1: ", img.sum().item())
-        print("Debug initial noise2: ", torch.randn(shape, device=device).sum().item())
+        print("Debug initial noise: ", torch.randn(shape, device=device).sum().item())
+        print("Debug initial noise: ", torch.randn(shape, device=device).sum().item())
         print("noise device: ", img.device)
         
         if timesteps is None:

VADER-VideoCrafter/scripts/main/funcs.py

@@ -14,7 +14,7 @@ from lvdm.models.samplers.ddim import DDIMSampler
 # st = ipdb.set_trace
 
 def batch_ddim_sampling(model, cond, noise_shape, n_samples=1, ddim_steps=50, ddim_eta=1.0,\
-                        cfg_scale=1.0, temporal_cfg_scale=None, backprop_mode=None, decode_frame='-1', seed=0, **kwargs):
+                        cfg_scale=1.0, temporal_cfg_scale=None, backprop_mode=None, decode_frame='-1', **kwargs):
     ddim_sampler = DDIMSampler(model)
     if backprop_mode is not None:   # it is for training now, backprop_mode != None also means vader training mode
         ddim_sampler.backprop_mode = backprop_mode
@@ -64,7 +64,6 @@ def batch_ddim_sampling(model, cond, noise_shape, n_samples=1, ddim_steps=50, dd
                                             temporal_length=noise_shape[2],
                                             conditional_guidance_scale_temporal=temporal_cfg_scale,
                                             x_T=x_T,
-                                            seed=seed,
                                             **kwargs
                                             )
             

VADER-VideoCrafter/scripts/main/train_t2v_lora.py

@@ -655,10 +655,10 @@ def run_training(args, model, **kwargs):
                 seed_everything_self(args.seed)
                 if isinstance(peft_model, torch.nn.parallel.DistributedDataParallel):
                     batch_samples = batch_ddim_sampling(peft_model.module, cond, noise_shape, args.n_samples, \
-                                                        args.ddim_steps, args.ddim_eta, args.unconditional_guidance_scale, None, decode_frame=args.decode_frame, seed=args.seed, **kwargs)
+                                                        args.ddim_steps, args.ddim_eta, args.unconditional_guidance_scale, None, decode_frame=args.decode_frame, **kwargs)
                 else:
                     batch_samples = batch_ddim_sampling(peft_model, cond, noise_shape, args.n_samples, \
-                                                            args.ddim_steps, args.ddim_eta, args.unconditional_guidance_scale, None, decode_frame=args.decode_frame, seed=args.seed, **kwargs)
+                                                            args.ddim_steps, args.ddim_eta, args.unconditional_guidance_scale, None, decode_frame=args.decode_frame, **kwargs)
 
                 print("batch_samples dtype: ", batch_samples.dtype)
                 print("batch_samples device: ", batch_samples.device)