Implement SD3 loss weighting (#8528)

* Add lognorm and cosmap weighting

* Implement mode sampling

* Update examples/dreambooth/train_dreambooth_lora_sd3.py

* Update examples/dreambooth/train_dreambooth_lora_sd3.py

* Update examples/dreambooth/train_dreambooth_sd3.py

* Update examples/dreambooth/train_dreambooth_sd3.py

* Update examples/dreambooth/train_dreambooth_sd3.py

* Update examples/dreambooth/train_dreambooth_lora_sd3.py

* Update examples/dreambooth/train_dreambooth_sd3.py

* Update examples/dreambooth/train_dreambooth_sd3.py

* Update examples/dreambooth/train_dreambooth_lora_sd3.py

* keep timestamp sampling fully on cpu

---------

Co-authored-by: Kashif Rasul <[email protected]>
Co-authored-by: Sayak Paul <[email protected]>

Author: 3 people

examples/dreambooth/train_dreambooth_lora_sd3.py

@@ -1462,7 +1462,18 @@ def get_sigmas(timesteps, n_dim=4, dtype=torch.float32):
                 bsz = model_input.shape[0]
 
                 # Sample a random timestep for each image
-                indices = torch.randint(0, noise_scheduler_copy.config.num_train_timesteps, (bsz,))
+                # for weighting schemes where we sample timesteps non-uniformly
+                if args.weighting_scheme == "logit_normal":
+                    # See 3.1 in the SD3 paper ($rf/lognorm(0.00,1.00)$).
+                    u = torch.normal(mean=args.logit_mean, std=args.logit_std, size=(bsz,), device="cpu")
+                    u = torch.nn.functional.sigmoid(u)
+                elif args.weighting_scheme == "mode":
+                    u = torch.rand(size=(bsz,), device="cpu")
+                    u = 1 - u - args.mode_scale * (torch.cos(math.pi * u / 2) ** 2 - 1 + u)
+                else:
+                    u = torch.rand(size=(bsz,), device="cpu")
+
+                indices = (u * noise_scheduler_copy.config.num_train_timesteps).long()
                 timesteps = noise_scheduler_copy.timesteps[indices].to(device=model_input.device)
 
                 # Add noise according to flow matching.
@@ -1483,16 +1494,15 @@ def get_sigmas(timesteps, n_dim=4, dtype=torch.float32):
                 model_pred = model_pred * (-sigmas) + noisy_model_input
 
                 # TODO (kashif, sayakpaul): weighting sceme needs to be experimented with :)
+                # these weighting schemes use a uniform timestep sampling
+                # and instead post-weight the loss
                 if args.weighting_scheme == "sigma_sqrt":
                     weighting = (sigmas**-2.0).float()
-                elif args.weighting_scheme == "logit_normal":
-                    # See 3.1 in the SD3 paper ($rf/lognorm(0.00,1.00)$).
-                    u = torch.normal(mean=args.logit_mean, std=args.logit_std, size=(bsz,), device=accelerator.device)
-                    weighting = torch.nn.functional.sigmoid(u)
-                elif args.weighting_scheme == "mode":
-                    # See sec 3.1 in the SD3 paper (20).
-                    u = torch.rand(size=(bsz,), device=accelerator.device)
-                    weighting = 1 - u - args.mode_scale * (torch.cos(math.pi * u / 2) ** 2 - 1 + u)
+                elif args.weighting_scheme == "cosmap":
+                    bot = 1 - 2 * sigmas + 2 * sigmas**2
+                    weighting = 2 / (math.pi * bot)
+                else:
+                    weighting = torch.ones_like(sigmas)
 
                 # simplified flow matching aka 0-rectified flow matching loss
                 # target = model_input - noise

examples/dreambooth/train_dreambooth_sd3.py

@@ -1526,7 +1526,18 @@ def get_sigmas(timesteps, n_dim=4, dtype=torch.float32):
                 bsz = model_input.shape[0]
 
                 # Sample a random timestep for each image
-                indices = torch.randint(0, noise_scheduler_copy.config.num_train_timesteps, (bsz,))
+                # for weighting schemes where we sample timesteps non-uniformly
+                if args.weighting_scheme == "logit_normal":
+                    # See 3.1 in the SD3 paper ($rf/lognorm(0.00,1.00)$).
+                    u = torch.normal(mean=args.logit_mean, std=args.logit_std, size=(bsz,), device="cpu")
+                    u = torch.nn.functional.sigmoid(u)
+                elif args.weighting_scheme == "mode":
+                    u = torch.rand(size=(bsz,), device="cpu")
+                    u = 1 - u - args.mode_scale * (torch.cos(math.pi * u / 2) ** 2 - 1 + u)
+                else:
+                    u = torch.rand(size=(bsz,), device="cpu")
+
+                indices = (u * noise_scheduler_copy.config.num_train_timesteps).long()
                 timesteps = noise_scheduler_copy.timesteps[indices].to(device=model_input.device)
 
                 # Add noise according to flow matching.
@@ -1560,18 +1571,15 @@ def get_sigmas(timesteps, n_dim=4, dtype=torch.float32):
                 # Follow: Section 5 of https://arxiv.org/abs/2206.00364.
                 # Preconditioning of the model outputs.
                 model_pred = model_pred * (-sigmas) + noisy_model_input
-
-                # TODO (kashif, sayakpaul): weighting sceme needs to be experimented with :)
+                # these weighting schemes use a uniform timestep sampling
+                # and instead post-weight the loss
                 if args.weighting_scheme == "sigma_sqrt":
                     weighting = (sigmas**-2.0).float()
-                elif args.weighting_scheme == "logit_normal":
-                    # See 3.1 in the SD3 paper ($rf/lognorm(0.00,1.00)$).
-                    u = torch.normal(mean=args.logit_mean, std=args.logit_std, size=(bsz,), device=accelerator.device)
-                    weighting = torch.nn.functional.sigmoid(u)
-                elif args.weighting_scheme == "mode":
-                    # See sec 3.1 in the SD3 paper (20).
-                    u = torch.rand(size=(bsz,), device=accelerator.device)
-                    weighting = 1 - u - args.mode_scale * (torch.cos(math.pi * u / 2) ** 2 - 1 + u)
+                elif args.weighting_scheme == "cosmap":
+                    bot = 1 - 2 * sigmas + 2 * sigmas**2
+                    weighting = 2 / (math.pi * bot)
+                else:
+                    weighting = torch.ones_like(sigmas)
 
                 # simplified flow matching aka 0-rectified flow matching loss
                 # target = model_input - noise