[SIGGRAPH Asia 2026] UniMate: One Unified Model to Animate Diverse Skeletons
Python
#!/bin/bash
# Train a UniMate model on a single GPU.
#
# Usage:
# bash scripts/run_train.sh [config] [-- extra args...]
#
# Arguments:
# config training config JSON (default: configs/uniml3d_60frames_graph_adaln.json)
# extra anything after `--` is forwarded to unimate.training.train,
# e.g. --batch_size 8 --resume outputs/<run>/checkpoints/<ckpt>.pt
#
# Env overrides:
# OUTPUT_DIR (default: the config's experiment.output_dir)
# CONDA_ENV conda environment to activate (default: unimate)
# CUDA_VISIBLE_DEVICES GPU to use (default: the one with the most free memory)
#
# Multi-GPU: accelerate launch --num_processes <N> -m unimate.training.train --config <config>
set -euo pipefail
source "$(dirname "${BASH_SOURCE[0]}")/_common.sh"
handle_help "$@"
CONFIG=configs/uniml3d_60frames_graph_adaln.json
if [[ $# -gt 0 && "$1" != "--" ]]; then CONFIG=$1; shift; fi
[[ "${1:-}" == "--" ]] && shift
[[ -f "$CONFIG" ]] || { echo "No such config: $CONFIG" >&2; exit 1; }
select_gpu
echo "Config: $CONFIG"
echo "Output dir: ${OUTPUT_DIR:-<config default>}"
CMD=(accelerate launch --num_processes 1 -m unimate.training.train --config "$CONFIG")
[[ -n "${OUTPUT_DIR:-}" ]] && CMD+=(--output_dir "$OUTPUT_DIR")
"${CMD[@]}" "$@"