Added input pretubation (huggingface#3292)

* Added input pretubation

* Fixed spelling

Author: isamu-isozaki

examples/text_to_image/train_text_to_image.py

@@ -112,6 +112,9 @@ def log_validation(vae, text_encoder, tokenizer, unet, args, accelerator, weight
 
 def parse_args():
     parser = argparse.ArgumentParser(description="Simple example of a training script.")
+    parser.add_argument(
+        "--input_pertubation", type=float, default=0, help="The scale of input pretubation. Recommended 0.1."
+    )
     parser.add_argument(
         "--pretrained_model_name_or_path",
         type=str,
@@ -801,15 +804,19 @@ def collate_fn(examples):
                     noise += args.noise_offset * torch.randn(
                         (latents.shape[0], latents.shape[1], 1, 1), device=latents.device
                     )
-
+                if args.input_pertubation:
+                    new_noise = noise + args.input_pertubation * torch.randn_like(noise)
                 bsz = latents.shape[0]
                 # Sample a random timestep for each image
                 timesteps = torch.randint(0, noise_scheduler.config.num_train_timesteps, (bsz,), device=latents.device)
                 timesteps = timesteps.long()
 
                 # Add noise to the latents according to the noise magnitude at each timestep
                 # (this is the forward diffusion process)
-                noisy_latents = noise_scheduler.add_noise(latents, noise, timesteps)
+                if args.input_pertubation:
+                    noisy_latents = noise_scheduler.add_noise(latents, new_noise, timesteps)
+                else:
+                    noisy_latents = noise_scheduler.add_noise(latents, noise, timesteps)
 
                 # Get the text embedding for conditioning
                 encoder_hidden_states = text_encoder(batch["input_ids"])[0]