Update

1 files changed, 14 insertions, 7 deletions

diff --git a/training/functional.py b/training/functional.py
index 4d0cf0e..85dd884 100644
--- a/training/functional.py
+++ b/training/functional.py
@@ -12,7 +12,7 @@ from torch.utils.data import DataLoader
 
 from accelerate import Accelerator
 from transformers import CLIPTextModel
-from diffusers import AutoencoderKL, UNet2DConditionModel, UniPCMultistepScheduler
+from diffusers import AutoencoderKL, UNet2DConditionModel, DDPMScheduler, UniPCMultistepScheduler, SchedulerMixin
 
 from tqdm.auto import tqdm
 from PIL import Image
@@ -74,12 +74,19 @@ def make_grid(images, rows, cols):
     return grid
 
 
-def get_models(pretrained_model_name_or_path: str):
+def get_models(pretrained_model_name_or_path: str, noise_scheduler: str = "ddpm"):
+    if noise_scheduler == "deis":
+        noise_scheduler_cls = DEISMultistepScheduler
+    elif noise_scheduler == "ddpm":
+        noise_scheduler_cls = DDPMScheduler
+    else:
+        raise ValueError(f"noise_scheduler must be one of [\"ddpm\", \"deis\"], got {noise_scheduler}")
+
     tokenizer = MultiCLIPTokenizer.from_pretrained(pretrained_model_name_or_path, subfolder='tokenizer')
     text_encoder = CLIPTextModel.from_pretrained(pretrained_model_name_or_path, subfolder='text_encoder')
     vae = AutoencoderKL.from_pretrained(pretrained_model_name_or_path, subfolder='vae')
     unet = UNet2DConditionModel.from_pretrained(pretrained_model_name_or_path, subfolder='unet')
-    noise_scheduler = DEISMultistepScheduler.from_pretrained(pretrained_model_name_or_path, subfolder='scheduler')
+    noise_scheduler = noise_scheduler_cls.from_pretrained(pretrained_model_name_or_path, subfolder='scheduler')
     sample_scheduler = UniPCMultistepScheduler.from_pretrained(
         pretrained_model_name_or_path, subfolder='scheduler')
 
@@ -94,7 +101,7 @@ def save_samples(
     text_encoder: CLIPTextModel,
     tokenizer: MultiCLIPTokenizer,
     vae: AutoencoderKL,
-    sample_scheduler: UniPCMultistepScheduler,
+    sample_scheduler: SchedulerMixin,
     train_dataloader: DataLoader,
     val_dataloader: Optional[DataLoader],
     output_dir: Path,
@@ -181,7 +188,7 @@ def generate_class_images(
     vae: AutoencoderKL,
     unet: UNet2DConditionModel,
     tokenizer: MultiCLIPTokenizer,
-    sample_scheduler: UniPCMultistepScheduler,
+    sample_scheduler: SchedulerMixin,
     train_dataset: VlpnDataset,
     sample_batch_size: int,
     sample_image_size: int,
@@ -252,7 +259,7 @@ def add_placeholder_tokens(
 
 def loss_step(
     vae: AutoencoderKL,
-    noise_scheduler: DEISMultistepScheduler,
+    noise_scheduler: SchedulerMixin,
     unet: UNet2DConditionModel,
     text_encoder: CLIPTextModel,
     with_prior_preservation: bool,
@@ -552,7 +559,7 @@ def train(
     unet: UNet2DConditionModel,
     text_encoder: CLIPTextModel,
     vae: AutoencoderKL,
-    noise_scheduler: DEISMultistepScheduler,
+    noise_scheduler: SchedulerMixin,
     dtype: torch.dtype,
     seed: int,
     project: str,