Skip to content

Open-Sora: Democratizing Efficient Video Production for All

Python

rflow scheduler unfinished(40%)

ChangeFWorld9a14b9e23cac
diff --git a/opensora/schedulers/rf/__init__.py b/opensora/schedulers/rf/__init__.py
index e69de29..e8bdc45 100644
--- a/opensora/schedulers/rf/__init__.py
+++ b/opensora/schedulers/rf/__init__.py
@@ -0,0 +1,10 @@
+# should have property num_timesteps,
+# method sample() training_losses()
+import torch
+from .rectified_flow import RFlowScheduler
+from functools import partial
+
+from opensora.registry import SCHEDULERS
+
+# @SCHEDULERS.register_module("rflow")
+# class RFLOW:
\ No newline at end of file
diff --git a/opensora/schedulers/rf/rectified_flow.py b/opensora/schedulers/rf/rectified_flow.py
new file mode 100644
index 0000000..80732e5
--- /dev/null
+++ b/opensora/schedulers/rf/rectified_flow.py
@@ -0,0 +1,49 @@
+import torch
+import numpy as np
+
+# some code are inspired by https://github.com/magic-research/piecewise-rectified-flow/blob/main/scripts/train_perflow.py
+
+
+class RFlowScheduler:
+ def __init__(
+ self,
+ num_timesteps = 1000,
+ ):
+ self.num_timesteps = num_timesteps
+
+
+ def training_losses(self, model, x_start, t, model_kwargs=None, noise = None, mask = None, weights = None):
+ '''
+ Compute training losses for a single timestep.
+ Arguments format copied from opensora/schedulers/iddpm/gaussian_diffusion.py/training_losses
+ Note: t is int tensor and should be rescaled from [0, num_timesteps-1] to [1,0]
+ '''
+ assert mask is None, "mask not support for rectified flow yet"
+ assert weights is None, "weights not support for rectified flow yet"
+
+ if model_kwargs is None:
+ model_kwargs = {}
+ if noise is None:
+ noise = torch.randn_like(x_start)
+
+ x_t = self.add_noise(x_start, noise, t)
+
+ terms = {}
+ velocity_pred = model(x_t, t, **model_kwargs)
+ loss = (velocity_pred - (x_start - noise)).pow(2).mean()
+ terms['loss'] = loss
+
+ return terms
+
+ def add_noise(self, x0, x1, t):
+ '''
+ x0: sample of dataset
+ x1: sample of gaussian distribution
+ '''
+ # rescale t from [0,num_timesteps] to [0,1]
+ t = t / self.num_timesteps
+ return t * x1 + (1 - t) * x0
+
+ def step():
+ pass
+
\ No newline at end of file