Fix prompt2gloss import and refine outputs

One-to-All-Animation/video-generation/inference_1.3b.py

@@ -366,6 +366,7 @@ def unified_worker(rank, world_size, task_list, ckpt_path, use_direct_paths=Fals
 
         alpha = 0.6
         frames_combined = []
+        frames_pred = []
 
         src_uint8 = ((pose_tensor / 2 + 0.5).clamp(0, 1) * 255
                     )[0].byte().permute(1, 2, 3, 0).numpy()     # (T,H,W,C)
@@ -378,6 +379,7 @@ def unified_worker(rank, world_size, task_list, ckpt_path, use_direct_paths=Fals
             blended = (alpha * src + (1 - alpha) * pred).round().astype(np.uint8)
             concat  = np.concatenate([blended, pred.astype(np.uint8)], axis=1)  # (H,2W,C)
             frames_combined.append(concat)
+            frames_pred.append(pred.astype(np.uint8))
 
         imageio.mimwrite(
             final_video_path,
@@ -386,6 +388,15 @@ def unified_worker(rank, world_size, task_list, ckpt_path, use_direct_paths=Fals
             quality=5
         )
 
+        final_pred_path = os.path.join(output_root, f"{job_name}_ref_{ref_cfg}_pose_{pose_cfg}_pred.mp4")
+        if not os.path.exists(final_pred_path):
+            imageio.mimwrite(
+                final_pred_path,
+                frames_pred,
+                fps=pose_fps,
+                quality=5
+            )
+
         print(f"[GPU {rank}] 完成视频: {final_video_path}")
     del pipe
     torch.cuda.empty_cache()

pipeline03_pose2video.sh

@@ -6,44 +6,144 @@ source /research/cbim/vast/sf895/miniforge3/bin/activate one-to-all
 ROOT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
 VIDEO_GEN_DIR="${ROOT_DIR}/One-to-All-Animation/video-generation"
 
-POSE_VIDEO="${POSE_VIDEO:-HELLO_GOOD_MORNING_20251215_142658_openpose_style.mp4}"
+POSE_VIDEO="${POSE_VIDEO:-}"
 REF_IMAGE="${REF_IMAGE:-${ROOT_DIR}/ref_img_dict/man.jpg}"
-GPU_ID="${GPU_ID:-1}"
+GPU_ID="${GPU_ID:-}"
 REF_CFG="${REF_CFG:-2.5}"
 POSE_CFG="${POSE_CFG:-1.5}"
 STAGE1_OUT_DIR="${STAGE1_OUT_DIR:-${ROOT_DIR}/generated_sign_video}"
 STAGE2_OUT_DIR="${STAGE2_OUT_DIR:-${ROOT_DIR}/generated_sign_video_refine}"
+STAGE1_ONLY=0
+STAGE2_ONLY=0
+STAGE2_INPUT=""
 
-echo "Stage1: pose2video"
-(
-  cd "${VIDEO_GEN_DIR}"
-  ONE_TO_ALL_OUTPUT_DIR="${STAGE1_OUT_DIR}" \
-  CUDA_VISIBLE_DEVICES="${GPU_ID}" \
-  python inference_1.3b.py --pose_video "${POSE_VIDEO}" --ref_image "${REF_IMAGE}"
-)
-
-pose_base="$(basename "${POSE_VIDEO}")"
-pose_base="${pose_base%.*}"
-ref_base="$(basename "${REF_IMAGE}")"
-ref_base="${ref_base%.*}"
-
-stage1_pattern="${STAGE1_OUT_DIR}"/*_ref_${REF_CFG}_pose_${POSE_CFG}/"ref_${ref_base}_vid_${pose_base}_ref_${REF_CFG}_pose_${POSE_CFG}.mp4"
-stage1_video="$(ls -t ${stage1_pattern} 2>/dev/null | head -n 1 || true)"
-if [[ -z "${stage1_video}" ]]; then
-  echo "Stage1 output not found. Pattern: ${stage1_pattern}" >&2
+usage() {
+  cat <<'USAGE'
+Usage:
+  pipeline03_pose2video.sh --pose-video <pose.mp4> --ref-image <ref.jpg> [options]
+
+Options:
+  --pose-video PATH     Pose video for stage1 (or refine input if --stage2-only).
+  --ref-image PATH      Reference image for both stages.
+  --stage1-out PATH     Output dir for stage1 (default: generated_sign_video).
+  --stage2-out PATH     Output dir for stage2 (default: generated_sign_video_refine).
+  --stage1-only         Run only stage1.
+  --stage2-only         Run only stage2 (uses --pose-video as refine input unless --stage2-input is provided).
+  --stage2-input PATH   Explicit input video/frames for stage2 only.
+  --gpu ID              GPU id (default: auto-pick free GPU).
+  --ref-cfg FLOAT       Image guidance scale (default: 2.5).
+  --pose-cfg FLOAT      Pose guidance scale (default: 1.5).
+  -h, --help            Show this help.
+USAGE
+}
+
+while [[ $# -gt 0 ]]; do
+  case "$1" in
+    --pose-video) POSE_VIDEO="$2"; shift 2;;
+    --ref-image) REF_IMAGE="$2"; shift 2;;
+    --stage1-out) STAGE1_OUT_DIR="$2"; shift 2;;
+    --stage2-out) STAGE2_OUT_DIR="$2"; shift 2;;
+    --stage1-only) STAGE1_ONLY=1; shift 1;;
+    --stage2-only) STAGE2_ONLY=1; shift 1;;
+    --stage2-input) STAGE2_INPUT="$2"; shift 2;;
+    --gpu) GPU_ID="$2"; shift 2;;
+    --ref-cfg) REF_CFG="$2"; shift 2;;
+    --pose-cfg) POSE_CFG="$2"; shift 2;;
+    -h|--help) usage; exit 0;;
+    *) echo "Unknown option: $1"; usage; exit 1;;
+  esac
+done
+
+if [[ "${STAGE1_ONLY}" -eq 1 && "${STAGE2_ONLY}" -eq 1 ]]; then
+  echo "Cannot use --stage1-only and --stage2-only together." >&2
   exit 1
 fi
 
-echo "Stage2: refine on ${stage1_video}"
-(
-  cd "${VIDEO_GEN_DIR}"
-  ./infer_refine.sh \
-    --input "${stage1_video}" \
-    --ref "${REF_IMAGE}" \
-    --ref-cfg "${REF_CFG}" \
-    --pose-cfg "${POSE_CFG}" \
-    --out-dir "${STAGE2_OUT_DIR}" \
-    --gpu "${GPU_ID}"
-)
+pick_free_gpu() {
+  if command -v nvidia-smi >/dev/null 2>&1; then
+    local best
+    best="$(nvidia-smi --query-gpu=index,memory.free --format=csv,noheader,nounits 2>/dev/null | \
+      awk -F',' '{gsub(/ /,"",$2); if($2>max){max=$2; idx=$1}} END{print idx}')"
+    if [[ -n "${best}" ]]; then
+      echo "${best}"
+      return
+    fi
+  fi
+  echo "0"
+}
+
+if [[ -z "${GPU_ID}" ]]; then
+  GPU_ID="$(pick_free_gpu)"
+  echo "Using GPU: ${GPU_ID}"
+fi
+
+if [[ "${STAGE2_ONLY}" -eq 1 ]]; then
+  if [[ -z "${STAGE2_INPUT}" ]]; then
+    if [[ -n "${POSE_VIDEO}" ]]; then
+      STAGE2_INPUT="${POSE_VIDEO}"
+      echo "Stage2-only: using --pose-video as refine input."
+    else
+      echo "Stage2-only requires --stage2-input (or --pose-video)." >&2
+      exit 1
+    fi
+  fi
+else
+  if [[ -z "${POSE_VIDEO}" ]]; then
+    echo "Stage1 requires --pose-video." >&2
+    exit 1
+  fi
+fi
+
+run_stage1() {
+  echo "Stage1: pose2video"
+  (
+    cd "${VIDEO_GEN_DIR}"
+    ONE_TO_ALL_OUTPUT_DIR="${STAGE1_OUT_DIR}" \
+    CUDA_VISIBLE_DEVICES="${GPU_ID}" \
+    python inference_1.3b.py --pose_video "${POSE_VIDEO}" --ref_image "${REF_IMAGE}"
+  )
+}
+
+resolve_stage1_output() {
+  local pose_base ref_base stage1_pattern stage1_video
+  pose_base="$(basename "${POSE_VIDEO}")"
+  pose_base="${pose_base%.*}"
+  ref_base="$(basename "${REF_IMAGE}")"
+  ref_base="${ref_base%.*}"
+
+  stage1_pattern="${STAGE1_OUT_DIR}"/*_ref_${REF_CFG}_pose_${POSE_CFG}/"ref_${ref_base}_vid_${pose_base}_ref_${REF_CFG}_pose_${POSE_CFG}.mp4"
+  stage1_video="$(ls -t ${stage1_pattern} 2>/dev/null | head -n 1 || true)"
+  if [[ -z "${stage1_video}" ]]; then
+    echo "Stage1 output not found. Pattern: ${stage1_pattern}" >&2
+    exit 1
+  fi
+  echo "${stage1_video}"
+}
+
+run_stage2() {
+  local input_video="$1"
+  echo "Stage2: refine on ${input_video}"
+  (
+    cd "${VIDEO_GEN_DIR}"
+    ./infer_refine.sh \
+      --input "${input_video}" \
+      --ref "${REF_IMAGE}" \
+      --ref-cfg "${REF_CFG}" \
+      --pose-cfg "${POSE_CFG}" \
+      --out-dir "${STAGE2_OUT_DIR}" \
+      --gpu "${GPU_ID}"
+  )
+}
+
+if [[ "${STAGE2_ONLY}" -ne 1 ]]; then
+  run_stage1
+  if [[ "${STAGE1_ONLY}" -eq 1 ]]; then
+    echo "Done (stage1 only)."
+    exit 0
+  fi
+  STAGE2_INPUT="$(resolve_stage1_output)"
+fi
+
+run_stage2 "${STAGE2_INPUT}"
 
 echo "Done."

prompt2gloss/__init__.py

@@ -0,0 +1,5 @@
+"""Prompt2Gloss package helpers."""
+
+from .prompt2gloss import build_input  # re-export for plugin importer
+
+__all__ = ["build_input"]