| #!/usr/bin/env bash |
| set -euo pipefail |
|
|
| ROOT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")/../.." && pwd)" |
| cd "$ROOT_DIR" |
|
|
| CONFIG="${1:-configs/stage2_distillation.py}" |
| CKPT="${2:-${JOYAI_CHECKPOINT:-}}" |
| : "${CKPT:?Pass a checkpoint as the second argument or set JOYAI_CHECKPOINT.}" |
| DATASET_ROOT="${DATASET_ROOT:-data/OpenVE-Bench}" |
| CSV_PATH="${CSV_PATH:-${DATASET_ROOT}/benchmark_videos.csv}" |
| TASK_TYPES="${TASK_TYPES:-global_style,background_change,local_change,local_remove,local_add}" |
| SAVE="${SAVE:-outputs_eval/openvebench}" |
| NODE_RANK="${NODE_RANK:-0}" |
| NEG_PROMPT="An abstract, computer-generated scene with distorted and blurry visuals. A deformed, disfigured figure without specific features, depicted as an illustration. The background is a collage of grainy textures and striped patterns, lacking clear visual content. The figure moves minimally with weak dynamics and a stuttering effect, displaying distorted and erratic motions. The style incorporates extremely high contrast and extremely high sharpness, combined with low-quality imagery, grainy effects, and includes logos and text elements. The camera employs disjointed and stuttering movements, inconsistent framing, and unstructured composition." |
|
|
| EXTRA_ARGS=() |
| if [[ "${USE_RELATIVE_ROPE:-True}" == "True" ]]; then |
| EXTRA_ARGS+=(--use-relative-rope) |
| fi |
| if [[ "${STORE_CLEAN_ONLY_SELF:-}" == "True" ]]; then |
| EXTRA_ARGS+=(--store-clean-only-self) |
| elif [[ "${STORE_CLEAN_ONLY_SELF:-}" == "False" ]]; then |
| EXTRA_ARGS+=(--no-store-clean-only-self) |
| fi |
| if [[ "${USE_PE:-False}" == "True" ]]; then |
| EXTRA_ARGS+=(--pe-csv-path "${PE_CSV_PATH:-${DATASET_ROOT}/pe_benchmark_videos.csv}") |
| fi |
| if [[ -n "${INFERENCE_STEP:-}" ]]; then |
| EXTRA_ARGS+=(--num-inference-steps "$INFERENCE_STEP") |
| fi |
| if [[ -n "${CFG:-}" ]]; then |
| EXTRA_ARGS+=(--guidance-scale "$CFG") |
| fi |
|
|
| export PYTHONPATH="$ROOT_DIR${PYTHONPATH:+:$PYTHONPATH}" |
|
|
| torchrun \ |
| --nnodes "${NNODES:-1}" \ |
| --nproc_per_node "${GPUS_PER_NODE:-8}" \ |
| --node_rank "$NODE_RANK" \ |
| --master_addr "${MASTER_ADDR:-127.0.0.1}" \ |
| --master_port "${MASTER_PORT:-6000}" \ |
| --module evaluation.benchmark.infer_openvebench_joyomni \ |
| --config "$CONFIG" \ |
| --ckpt-path "$CKPT" \ |
| --save-path "$SAVE" \ |
| --dataset-root "$DATASET_ROOT" \ |
| --csv-path "$CSV_PATH" \ |
| --task-types "$TASK_TYPES" \ |
| --source-guidance-scale "${SRC_CFG:-1.0}" \ |
| --neg-prompt "$NEG_PROMPT" \ |
| --resolution "${RESOLUTION:-720}" \ |
| --max-temporal-ids "${MAX_TEMPORAL_IDS:-8}" \ |
| --seed 42 \ |
| "${EXTRA_ARGS[@]}" "${@:3}" |
|
|
| if [[ "${RUN_EVAL:-False}" == "True" && "$NODE_RANK" == "0" ]]; then |
| : "${GEMINI_API_KEY:?Set GEMINI_API_KEY to run evaluation.}" |
| python evaluation/OpenVE-Bench/eval_openve_gemini.py \ |
| --video_paths "$SAVE" \ |
| --dataset-root "$DATASET_ROOT" \ |
| --csv-path "$CSV_PATH" \ |
| --task-types "$TASK_TYPES" \ |
| --model_id "${MODEL_ID:-Gemini-2.5-pro}" \ |
| --max-workers "${MAX_WORKERS:-32}" |
| fi |
|
|
