Skip to content

[SIGGRAPH Asia 2026] UniMate: One Unified Model to Animate Diverse Skeletons

Python

Let num_epochs set the training budget even when num_steps is filled in; label generic bookkeeping bones such as extra_NN as Bone

Friedrich-Me62864e4201c
diff --git a/data_process/joint_annotation/names_clean_rule.py b/data_process/joint_annotation/names_clean_rule.py
index 5a4e222..52c111e 100644
--- a/data_process/joint_annotation/names_clean_rule.py
+++ b/data_process/joint_annotation/names_clean_rule.py
@@ -567,6 +567,10 @@ def _name_side(raw):
return side or extract_inner_side(bare)[0]
+# Bookkeeping names that say nothing about the body part: a placeholder.
+_GENERIC_NAMES = frozenset({'extra', 'helper', 'dummy', 'null', 'joint', 'node', 'bone'})
+
+
def clean_joint_name(raw, animal, refine=True):
"""Clean one joint name: the rules below, then (*refine*) the label
refinements the reviewed annotations support:
@@ -584,6 +588,9 @@ def clean_joint_name(raw, animal, refine=True):
if not refine or not raw or not raw.strip():
return out
cleaned = post_process(out)
+ side_m = re.match(r'^(Left|Right) ', cleaned)
+ if re.sub(r'^(Left|Right) ', '', cleaned).lower() in _GENERIC_NAMES:
+ out = cleaned = with_side(side_m.group(1) if side_m else None, 'Bone')
if cleaned == 'Bone' or not is_canonical_label(cleaned):
vocab = _learned_vocab()
label = next((vocab[k] for k in name_keys(raw) if k in vocab), None)
diff --git a/unimate/configs/schema.py b/unimate/configs/schema.py
index c656d25..0597da6 100644
--- a/unimate/configs/schema.py
+++ b/unimate/configs/schema.py
@@ -331,7 +331,7 @@ class TrainingConfig:
min_lr_ratio: float = 0.01 # Min LR as a fraction of learning_rate
# Duration
- num_epochs: Optional[int] = None
+ num_epochs: Optional[int] = None # when set, the budget (num_steps is then ignored)
num_steps: Optional[int] = 100_000
log_interval: int = 50
save_interval: int = 10_000
diff --git a/unimate/training/train.py b/unimate/training/train.py
index 64d11bf..ef7e3c2 100644
--- a/unimate/training/train.py
+++ b/unimate/training/train.py
@@ -157,9 +157,14 @@ def train_diffusion(args: TrainingArgs, config: MainConfig,
# ---- Training tracker ----
# Optimizer steps per epoch: this rank's batches over the accumulation
# (accelerate also steps on an epoch's last, partial group).
+ # An epoch budget wins over num_steps, which the schema always fills in.
+ epoch_budget = bool(config.training.num_epochs)
+ if epoch_budget and config.training.num_steps and is_main:
+ logger.info(f"num_epochs={config.training.num_epochs} sets the budget; "
+ f"num_steps={config.training.num_steps} is ignored.")
tracker = TrainingTracker(
- max_epochs=config.training.num_epochs,
- max_steps=config.training.num_steps,
+ max_epochs=config.training.num_epochs if epoch_budget else None,
+ max_steps=None if epoch_budget else config.training.num_steps,
steps_per_epoch=math.ceil(
len(dataloader) / config.training.gradient_accumulation_steps),
)