Euler_a: Re-introduce generator arg for reproducible output

2 files changed, 10 insertions, 3 deletions

diff --git a/dreambooth.py b/dreambooth.py
index a181293..db097e5 100644
--- a/dreambooth.py
+++ b/dreambooth.py
@@ -193,7 +193,7 @@ def parse_args():
     parser.add_argument(
         "--ema_power",
         type=float,
-        default=9 / 10
+        default=5 / 6
     )
     parser.add_argument(
         "--ema_max_decay",
@@ -957,7 +957,7 @@ def main():
                     "lr/text": lr_scheduler.get_last_lr()[1]
                 }
                 if args.use_ema:
-                    logs["ema_decay"] = ema_unet.decay
+                    logs["ema_decay"] = 1 - ema_unet.decay
 
                 accelerator.log(logs, step=global_step)
 
diff --git a/schedulers/scheduling_euler_ancestral_discrete.py b/schedulers/scheduling_euler_ancestral_discrete.py
index 3a2de68..828e0dd 100644
--- a/schedulers/scheduling_euler_ancestral_discrete.py
+++ b/schedulers/scheduling_euler_ancestral_discrete.py
@@ -130,6 +130,7 @@ class EulerAncestralDiscreteScheduler(SchedulerMixin, ConfigMixin):
         timestep: Union[float, torch.FloatTensor],
         step_index: Union[int, torch.IntTensor],
         sample: Union[torch.FloatTensor, np.ndarray],
+        generator: torch.Generator = None,
         return_dict: bool = True,
     ) -> Union[SchedulerOutput, Tuple]:
         """
@@ -165,7 +166,13 @@ class EulerAncestralDiscreteScheduler(SchedulerMixin, ConfigMixin):
 
         prev_sample = sample + derivative * dt
 
-        prev_sample = prev_sample + torch.randn_like(prev_sample) * sigma_up
+        prev_sample = prev_sample + torch.randn(
+            prev_sample.shape,
+            layout=prev_sample.layout,
+            device=prev_sample.device,
+            dtype=prev_sample.dtype,
+            generator=generator
+        ) * sigma_up
 
         if not return_dict:
             return (prev_sample,)