Added Perlin noise to training

3 files changed, 28 insertions, 8 deletions

diff --git a/pipelines/stable_diffusion/vlpn_stable_diffusion.py b/pipelines/stable_diffusion/vlpn_stable_diffusion.py
index f02dd72..5f4fc38 100644
--- a/pipelines/stable_diffusion/vlpn_stable_diffusion.py
+++ b/pipelines/stable_diffusion/vlpn_stable_diffusion.py
@@ -22,7 +22,7 @@ from diffusers import (
     PNDMScheduler,
 )
 from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion import StableDiffusionPipelineOutput
-from diffusers.utils import logging, randn_tensor
+from diffusers.utils import logging
 from transformers import CLIPTextModel, CLIPTokenizer
 
 from models.clip.util import unify_input_ids, get_extended_embeddings
@@ -308,7 +308,7 @@ class VlpnStableDiffusion(DiffusionPipeline):
 
     def prepare_image(self, batch_size, width, height, dtype, device, generator=None):
         noise = perlin_noise(
-            batch_size, width, height, res=1, octaves=4, generator=generator, dtype=dtype, device=device
+            batch_size, 1, width, height, res=1, octaves=4, generator=generator, dtype=dtype, device=device
         ).expand(batch_size, 3, width, height)
         return (1.4 * noise).clamp(-1, 1)
 
diff --git a/training/functional.py b/training/functional.py
index 1c38635..db46766 100644
--- a/training/functional.py
+++ b/training/functional.py
@@ -23,6 +23,7 @@ from models.clip.embeddings import ManagedCLIPTextEmbeddings, patch_managed_embe
 from models.clip.util import get_extended_embeddings
 from models.clip.tokenizer import MultiCLIPTokenizer
 from training.util import AverageMeter
+from util.noise import perlin_noise
 
 
 def const(result=None):
@@ -253,6 +254,7 @@ def loss_step(
     text_encoder: CLIPTextModel,
     with_prior_preservation: bool,
     prior_loss_weight: float,
+    perlin_strength: float,
     seed: int,
     step: int,
     batch: dict[str, Any],
@@ -275,6 +277,19 @@ def loss_step(
         generator=generator
     )
 
+    if perlin_strength != 0:
+        noise += perlin_strength * perlin_noise(
+            latents.shape[0],
+            latents.shape[1],
+            latents.shape[2],
+            latents.shape[3],
+            res=1,
+            octaves=4,
+            dtype=latents.dtype,
+            device=latents.device,
+            generator=generator
+        )
+
     # Sample a random timestep for each image
     timesteps = torch.randint(
         0,
@@ -559,6 +574,7 @@ def train(
     global_step_offset: int = 0,
     with_prior_preservation: bool = False,
     prior_loss_weight: float = 1.0,
+    perlin_strength: float = 0.1,
     **kwargs,
 ):
     text_encoder, unet, optimizer, train_dataloader, val_dataloader, lr_scheduler, extra = strategy.prepare(
@@ -593,6 +609,7 @@ def train(
         text_encoder,
         with_prior_preservation,
         prior_loss_weight,
+        perlin_strength,
         seed,
     )
 
diff --git a/util/noise.py b/util/noise.py
index 38ab172..3c4f82d 100644
--- a/util/noise.py
+++ b/util/noise.py
@@ -1,4 +1,5 @@
 import math
+
 import torch
 
 # 2D Perlin noise in PyTorch https://gist.github.com/vadimkantorov/ac1b097753f217c5c11bc2ff396e0a57
@@ -47,11 +48,13 @@ def rand_perlin_2d_octaves(shape, res, octaves=1, persistence=0.5, dtype=None, d
     return noise
 
 
-def perlin_noise(batch_size: int, width: int, height: int, res=8, octaves=1, dtype=None, device=None, generator=None):
+def perlin_noise(batch_size: int, channels: int, width: int, height: int, res=8, octaves=1, dtype=None, device=None, generator=None):
     return torch.stack([
-        rand_perlin_2d_octaves(
-            (width, height), (res, res), octaves, dtype=dtype, device=device, generator=generator
-        ).unsqueeze(0)
-        for _
-        in range(batch_size)
+        torch.stack([
+            rand_perlin_2d_octaves(
+                (width, height), (res, res), octaves, dtype=dtype, device=device, generator=generator
+            )
+            for _ in range(channels)
+        ])
+        for _ in range(batch_size)
     ])