EMG / run_predict_i2v.sh

Upload 2 files

8f06535 verified 7 months ago

13.8 kB

#!/bin/bash

# CogVideoX-Fun I2V Inference Script
# Usage: bash run_predict_i2v.sh [OPTIONS]

# =============================================================================
# 基础设置 (Basic Settings)
# =============================================================================
export CUDA_VISIBLE_DEVICES=6
export PYTHONPATH="/home/xhm001/Desktop/Code/VideoX-Fun-main:${PYTHONPATH}"

# 模型路径 (Model Path)
MODEL_NAME="/home/xhm001/Desktop/Code/VideoX-Fun-main/models/cogvideoxfun/models--alibaba-pai--CogVideoX-Fun-V1.1-5b-InP"

# =============================================================================
# LiLoRA 和 HyperNetwork 路径 (LiLoRA & HyperNetwork Paths)
# =============================================================================
# LiLoRA 权重路径
LILORA_PATH="/home/xhm001/Desktop/Code/VideoX-Fun-main/Hyper_150/checkpoint-14800/checkpoint-checkpoint-14800.safetensors"

# HyperNetwork 权重路径
HYPERNETWORK_PATH="/home/xhm001/Desktop/Code/VideoX-Fun-main/Hyper_150/checkpoint-14800/hypernetwork.safetensors"

# EMG 多模态参考编码器路径
MULTIMODAL_REF_ENCODER_CKPT_PATH=None
MULTIMODAL_REF_ENCODER_SCRIPT=None
EMG2POSE_CKPT_PATH=None
EMG2POSE_REPO=None
HYPERNETWORK_INPUT_DIM=None
EMG_NORMALIZATION_PARTICIPANT=None

# LiLoRA 训练参数 (应与训练配置匹配)
LORA_RANK=1
LORA_DOWN_DIM=100
LORA_UP_DIM=50
LORA_WEIGHT=1


# =============================================================================
# 输入数据路径 (Input Data Paths)
# =============================================================================
# 注意：当使用 --test_json_path 时，以下参数会从 JSON 中读取
# 单样本推理时，可以手动指定以下参数

# 首帧图像路径 (用于图像到视频生成)
FIRST_FRAME_PATH=None

# 参考视频路径 (用于 HyperNetwork 生成权重)
REF_VIDEO_PATH=None

# 测试集 JSON 路径 (可选，批量处理时使用)
TEST_JSON_PATH=/home/xhm001/Desktop/Code/VideoX-Fun-main/Open_VFX_for_evl/cake.json

# =============================================================================
# 视频生成参数 (Video Generation Parameters)
# =============================================================================
VIDEO_LENGTH=49
FPS=8
PARTIAL_VIDEO_LENGTH=None
OVERLAP_VIDEO_LENGTH=4

# =============================================================================
# 推理参数 (Inference Parameters)
# =============================================================================
# 注意：当使用 --test_json_path 时，以下参数会从 JSON 中读取
PROMPT=None
NEGATIVE_PROMPT="The video is not of a high quality, it has a low resolution. Watermark present in each frame. The background is solid. Strange body and strange trajectory. Distortion."
GUIDANCE_SCALE=6.0
SEED=43
NUM_INFERENCE_STEPS=50

# =============================================================================
# GPU 设置 (GPU Settings)
# =============================================================================
# GPU 内存模式: model_full_load, model_full_load_and_qfloat8, model_cpu_offload, model_cpu_offload_and_qfloat8, sequential_cpu_offload
GPU_MEMORY_MODE="model_full_load"

# 多 GPU 设置
ULYSSES_DEGREE=1
RING_DEGREE=1
FSPD_DIT=False
FSPD_TEXT_ENCODER=False
COMPILE_DIT=False

# 权重数据类型: bfloat16, float16, float32
WEIGHT_DTYPE="bfloat16"

# =============================================================================
# 采样器选择 (Sampler Selection)
# =============================================================================
# 可选: Euler, Euler A, DPM++, PNDM, DDIM_Cog, DDIM_Origin
SAMPLER_NAME="DDIM_Origin"

# =============================================================================
# 输出路径 (Output Path)
# =============================================================================
SAVE_PATH="samples/ALL"

# =============================================================================
# 解析命令行参数 (Parse Command Line Arguments)
# =============================================================================
while [[ $# -gt 0 ]]; do
    case $1 in
        --model_name)
            MODEL_NAME="$2"
            shift 2
            ;;
        --lilora_path)
            LILORA_PATH="$2"
            shift 2
            ;;
        --hypernetwork_path)
            HYPERNETWORK_PATH="$2"
            shift 2
            ;;
        --multimodal_ref_encoder_ckpt_path)
            MULTIMODAL_REF_ENCODER_CKPT_PATH="$2"
            shift 2
            ;;
        --multimodal_ref_encoder_script)
            MULTIMODAL_REF_ENCODER_SCRIPT="$2"
            shift 2
            ;;
        --emg2pose_ckpt_path)
            EMG2POSE_CKPT_PATH="$2"
            shift 2
            ;;
        --emg2pose_repo)
            EMG2POSE_REPO="$2"
            shift 2
            ;;
        --hypernetwork_input_dim)
            HYPERNETWORK_INPUT_DIM="$2"
            shift 2
            ;;
        --emg_normalization_participant)
            EMG_NORMALIZATION_PARTICIPANT="$2"
            shift 2
            ;;
        --first_frame_path)
            FIRST_FRAME_PATH="$2"
            shift 2
            ;;
        --ref_video_path)
            REF_VIDEO_PATH="$2"
            shift 2
            ;;
        --test_json_path)
            TEST_JSON_PATH="$2"
            shift 2
            ;;
        --video_length)
            VIDEO_LENGTH="$2"
            shift 2
            ;;
        --fps)
            FPS="$2"
            shift 2
            ;;
        --prompt)
            PROMPT="$2"
            shift 2
            ;;
        --negative_prompt)
            NEGATIVE_PROMPT="$2"
            shift 2
            ;;
        --guidance_scale)
            GUIDANCE_SCALE="$2"
            shift 2
            ;;
        --seed)
            SEED="$2"
            shift 2
            ;;
        --num_inference_steps)
            NUM_INFERENCE_STEPS="$2"
            shift 2
            ;;
        --lora_rank)
            LORA_RANK="$2"
            shift 2
            ;;
        --lora_down_dim)
            LORA_DOWN_DIM="$2"
            shift 2
            ;;
        --lora_up_dim)
            LORA_UP_DIM="$2"
            shift 2
            ;;
        --lora_weight)
            LORA_WEIGHT="$2"
            shift 2
            ;;
        --save_path)
            SAVE_PATH="$2"
            shift 2
            ;;
        --gpu_memory_mode)
            GPU_MEMORY_MODE="$2"
            shift 2
            ;;
        --weight_dtype)
            WEIGHT_DTYPE="$2"
            shift 2
            ;;
        --sampler_name)
            SAMPLER_NAME="$2"
            shift 2
            ;;
        --ulysses_degree)
            ULYSSES_DEGREE="$2"
            shift 2
            ;;
        --ring_degree)
            RING_DEGREE="$2"
            shift 2
            ;;
        --partial_video_length)
            PARTIAL_VIDEO_LENGTH="$2"
            shift 2
            ;;
        --overlap_video_length)
            OVERLAP_VIDEO_LENGTH="$2"
            shift 2
            ;;
        --help|-h)
            echo "Usage: bash run_predict_i2v.sh [OPTIONS]"
            echo ""
            echo "Options:"
            echo "  --model_name            模型路径 (默认: models/Diffusion_Transformer/CogVideoX-Fun-V1.1-2b-InP)"
            echo "  --lilora_path           LiLoRA 权重路径"
            echo "  --hypernetwork_path     HyperNetwork 权重路径"
            echo "  --multimodal_ref_encoder_ckpt_path EMG 多模态参考编码器 checkpoint"
            echo "  --multimodal_ref_encoder_script    EMG 多模态参考编码器脚本"
            echo "  --emg2pose_ckpt_path    EMG2Pose checkpoint"
            echo "  --emg2pose_repo         EMG2Pose repo 路径"
            echo "  --hypernetwork_input_dim HyperNetwork 输入维度"
            echo "  --emg_normalization_participant EMG 归一化参与者目录"
            echo "  --first_frame_path      首帧图像路径"
            echo "  --ref_video_path        参考视频路径 (用于 HyperNetwork)"
            echo "  --test_json_path        测试集 JSON 路径"
            echo "  --video_length          生成视频帧数 (默认: 49)"
            echo "  --fps                   输出视频 FPS (默认: 8)"
            echo "  --prompt                生成提示词"
            echo "  --negative_prompt       负向提示词"
            echo "  --guidance_scale        引导 scale (默认: 6.0)"
            echo "  --seed                  随机种子 (默认: 43)"
            echo "  --num_inference_steps   推理步数 (默认: 50)"
            echo "  --lora_rank             LoRA rank (默认: 1)"
            echo "  --lora_down_dim         LoRA down dimension (默认: 200)"
            echo "  --lora_up_dim           LoRA up dimension (默认: 100)"
            echo "  --lora_weight           LoRA 权重 (默认: 0.55)"
            echo "  --save_path             输出路径 (默认: samples/cogvideox-fun-videos_i2v)"
            echo "  --gpu_memory_mode       GPU 内存模式"
            echo "  --weight_dtype          权重数据类型"
            echo "  --sampler_name          采样器名称"
            echo "  --help                  显示帮助信息"
            exit 0
            ;;
        *)
            echo "Unknown option: $1"
            exit 1
            ;;
    esac
done

# =============================================================================
# 构建命令行参数 (Build Command Line Arguments)
# =============================================================================
PYTHON_ARGS=()

# GPU 设置
PYTHON_ARGS+=("--gpu_memory_mode" "${GPU_MEMORY_MODE}")
PYTHON_ARGS+=("--ulysses_degree" "${ULYSSES_DEGREE}")
PYTHON_ARGS+=("--ring_degree" "${RING_DEGREE}")

[[ "${FSPD_DIT}" == "True" ]] && PYTHON_ARGS+=("--fsdp_dit")
[[ "${FSPD_TEXT_ENCODER}" == "True" ]] && PYTHON_ARGS+=("--fsdp_text_encoder")
[[ "${COMPILE_DIT}" == "True" ]] && PYTHON_ARGS+=("--compile_dit")

# 模型设置
PYTHON_ARGS+=("--model_name" "${MODEL_NAME}")
PYTHON_ARGS+=("--sampler_name" "${SAMPLER_NAME}")

# LiLoRA 和 HyperNetwork
[[ -n "${LILORA_PATH}" && "${LILORA_PATH}" != "None" ]] && PYTHON_ARGS+=("--lilora_path" "${LILORA_PATH}")
[[ -n "${HYPERNETWORK_PATH}" && "${HYPERNETWORK_PATH}" != "None" ]] && PYTHON_ARGS+=("--hypernetwork_path" "${HYPERNETWORK_PATH}")
[[ -n "${MULTIMODAL_REF_ENCODER_CKPT_PATH}" && "${MULTIMODAL_REF_ENCODER_CKPT_PATH}" != "None" ]] && PYTHON_ARGS+=("--multimodal_ref_encoder_ckpt_path" "${MULTIMODAL_REF_ENCODER_CKPT_PATH}")
[[ -n "${MULTIMODAL_REF_ENCODER_SCRIPT}" && "${MULTIMODAL_REF_ENCODER_SCRIPT}" != "None" ]] && PYTHON_ARGS+=("--multimodal_ref_encoder_script" "${MULTIMODAL_REF_ENCODER_SCRIPT}")
[[ -n "${EMG2POSE_CKPT_PATH}" && "${EMG2POSE_CKPT_PATH}" != "None" ]] && PYTHON_ARGS+=("--emg2pose_ckpt_path" "${EMG2POSE_CKPT_PATH}")
[[ -n "${EMG2POSE_REPO}" && "${EMG2POSE_REPO}" != "None" ]] && PYTHON_ARGS+=("--emg2pose_repo" "${EMG2POSE_REPO}")
[[ -n "${HYPERNETWORK_INPUT_DIM}" && "${HYPERNETWORK_INPUT_DIM}" != "None" ]] && PYTHON_ARGS+=("--hypernetwork_input_dim" "${HYPERNETWORK_INPUT_DIM}")
[[ -n "${EMG_NORMALIZATION_PARTICIPANT}" && "${EMG_NORMALIZATION_PARTICIPANT}" != "None" ]] && PYTHON_ARGS+=("--emg_normalization_participant" "${EMG_NORMALIZATION_PARTICIPANT}")

# LiLoRA 参数
PYTHON_ARGS+=("--lora_rank" "${LORA_RANK}")
PYTHON_ARGS+=("--lora_down_dim" "${LORA_DOWN_DIM}")
PYTHON_ARGS+=("--lora_up_dim" "${LORA_UP_DIM}")
[[ -n "${LORA_WEIGHT}" && "${LORA_WEIGHT}" != "None" ]] && PYTHON_ARGS+=("--lora_weight" "${LORA_WEIGHT}")

# 视频参数
PYTHON_ARGS+=("--video_length" "${VIDEO_LENGTH}")
PYTHON_ARGS+=("--fps" "${FPS}")
[[ -n "${PARTIAL_VIDEO_LENGTH}" && "${PARTIAL_VIDEO_LENGTH}" != "None" ]] && PYTHON_ARGS+=("--partial_video_length" "${PARTIAL_VIDEO_LENGTH}")
PYTHON_ARGS+=("--overlap_video_length" "${OVERLAP_VIDEO_LENGTH}")

# 权重类型
PYTHON_ARGS+=("--weight_dtype" "${WEIGHT_DTYPE}")

# 推理参数 (单样本时使用，批量时从 JSON 读取)
if [[ -z "${TEST_JSON_PATH}" || "${TEST_JSON_PATH}" == "None" ]]; then
    [[ -n "${FIRST_FRAME_PATH}" && "${FIRST_FRAME_PATH}" != "None" ]] && PYTHON_ARGS+=("--first_frame_path" "${FIRST_FRAME_PATH}")
    [[ -n "${REF_VIDEO_PATH}" && "${REF_VIDEO_PATH}" != "None" ]] && PYTHON_ARGS+=("--ref_video_path" "${REF_VIDEO_PATH}")
    [[ -n "${PROMPT}" && "${PROMPT}" != "None" ]] && PYTHON_ARGS+=("--prompt" "${PROMPT}")
    PYTHON_ARGS+=("--negative_prompt" "${NEGATIVE_PROMPT}")
    PYTHON_ARGS+=("--guidance_scale" "${GUIDANCE_SCALE}")
    PYTHON_ARGS+=("--seed" "${SEED}")
    PYTHON_ARGS+=("--num_inference_steps" "${NUM_INFERENCE_STEPS}")
else
    PYTHON_ARGS+=("--test_json_path" "${TEST_JSON_PATH}")
fi

# 输出路径
PYTHON_ARGS+=("--save_path" "${SAVE_PATH}")

# =============================================================================
# 运行推理脚本 (Run Inference Script)
# =============================================================================
echo "========================================"
echo "CogVideoX-Fun I2V Inference"
echo "========================================"
echo "Model: ${MODEL_NAME}"
echo "LiLoRA: ${LILORA_PATH}"
echo "HyperNetwork: ${HYPERNETWORK_PATH}"
echo "Multimodal Ref Encoder: ${MULTIMODAL_REF_ENCODER_CKPT_PATH}"
echo "Video Length: ${VIDEO_LENGTH}"
if [[ -n "${TEST_JSON_PATH}" && "${TEST_JSON_PATH}" != "None" ]]; then
    echo "Test Set: ${TEST_JSON_PATH}"
else
    echo "First Frame: ${FIRST_FRAME_PATH}"
    echo "Prompt: ${PROMPT}"
fi
echo "========================================"

torchrun --nproc-per-node=1 --master_port=29501\
    /home/xhm001/Desktop/Code/VideoX-Fun-main/examples/cogvideox_fun/predict_i2v.py \
    "${PYTHON_ARGS[@]}"


echo ""
echo "Inference completed!"

	#!/bin/bash

	# CogVideoX-Fun I2V Inference Script
	# Usage: bash run_predict_i2v.sh [OPTIONS]

	# =============================================================================
	# 基础设置 (Basic Settings)
	# =============================================================================
	export CUDA_VISIBLE_DEVICES=6
	export PYTHONPATH="/home/xhm001/Desktop/Code/VideoX-Fun-main:${PYTHONPATH}"

	# 模型路径 (Model Path)
	MODEL_NAME="/home/xhm001/Desktop/Code/VideoX-Fun-main/models/cogvideoxfun/models--alibaba-pai--CogVideoX-Fun-V1.1-5b-InP"

	# =============================================================================
	# LiLoRA 和 HyperNetwork 路径 (LiLoRA & HyperNetwork Paths)
	# =============================================================================
	# LiLoRA 权重路径
	LILORA_PATH="/home/xhm001/Desktop/Code/VideoX-Fun-main/Hyper_150/checkpoint-14800/checkpoint-checkpoint-14800.safetensors"

	# HyperNetwork 权重路径
	HYPERNETWORK_PATH="/home/xhm001/Desktop/Code/VideoX-Fun-main/Hyper_150/checkpoint-14800/hypernetwork.safetensors"

	# EMG 多模态参考编码器路径
	MULTIMODAL_REF_ENCODER_CKPT_PATH=None
	MULTIMODAL_REF_ENCODER_SCRIPT=None
	EMG2POSE_CKPT_PATH=None
	EMG2POSE_REPO=None
	HYPERNETWORK_INPUT_DIM=None
	EMG_NORMALIZATION_PARTICIPANT=None

	# LiLoRA 训练参数 (应与训练配置匹配)
	LORA_RANK=1
	LORA_DOWN_DIM=100
	LORA_UP_DIM=50
	LORA_WEIGHT=1


	# =============================================================================
	# 输入数据路径 (Input Data Paths)
	# =============================================================================
	# 注意：当使用 --test_json_path 时，以下参数会从 JSON 中读取
	# 单样本推理时，可以手动指定以下参数

	# 首帧图像路径 (用于图像到视频生成)
	FIRST_FRAME_PATH=None

	# 参考视频路径 (用于 HyperNetwork 生成权重)
	REF_VIDEO_PATH=None

	# 测试集 JSON 路径 (可选，批量处理时使用)
	TEST_JSON_PATH=/home/xhm001/Desktop/Code/VideoX-Fun-main/Open_VFX_for_evl/cake.json

	# =============================================================================
	# 视频生成参数 (Video Generation Parameters)
	# =============================================================================
	VIDEO_LENGTH=49
	FPS=8
	PARTIAL_VIDEO_LENGTH=None
	OVERLAP_VIDEO_LENGTH=4

	# =============================================================================
	# 推理参数 (Inference Parameters)
	# =============================================================================
	# 注意：当使用 --test_json_path 时，以下参数会从 JSON 中读取
	PROMPT=None
	NEGATIVE_PROMPT="The video is not of a high quality, it has a low resolution. Watermark present in each frame. The background is solid. Strange body and strange trajectory. Distortion."
	GUIDANCE_SCALE=6.0
	SEED=43
	NUM_INFERENCE_STEPS=50

	# =============================================================================
	# GPU 设置 (GPU Settings)
	# =============================================================================
	# GPU 内存模式: model_full_load, model_full_load_and_qfloat8, model_cpu_offload, model_cpu_offload_and_qfloat8, sequential_cpu_offload
	GPU_MEMORY_MODE="model_full_load"

	# 多 GPU 设置
	ULYSSES_DEGREE=1
	RING_DEGREE=1
	FSPD_DIT=False
	FSPD_TEXT_ENCODER=False
	COMPILE_DIT=False

	# 权重数据类型: bfloat16, float16, float32
	WEIGHT_DTYPE="bfloat16"

	# =============================================================================
	# 采样器选择 (Sampler Selection)
	# =============================================================================
	# 可选: Euler, Euler A, DPM++, PNDM, DDIM_Cog, DDIM_Origin
	SAMPLER_NAME="DDIM_Origin"

	# =============================================================================
	# 输出路径 (Output Path)
	# =============================================================================
	SAVE_PATH="samples/ALL"

	# =============================================================================
	# 解析命令行参数 (Parse Command Line Arguments)
	# =============================================================================
	while [[ $# -gt 0 ]]; do
	case $1 in
	--model_name)
	MODEL_NAME="$2"
	shift 2
	;;
	--lilora_path)
	LILORA_PATH="$2"
	shift 2
	;;
	--hypernetwork_path)
	HYPERNETWORK_PATH="$2"
	shift 2
	;;
	--multimodal_ref_encoder_ckpt_path)
	MULTIMODAL_REF_ENCODER_CKPT_PATH="$2"
	shift 2
	;;
	--multimodal_ref_encoder_script)
	MULTIMODAL_REF_ENCODER_SCRIPT="$2"
	shift 2
	;;
	--emg2pose_ckpt_path)
	EMG2POSE_CKPT_PATH="$2"
	shift 2
	;;
	--emg2pose_repo)
	EMG2POSE_REPO="$2"
	shift 2
	;;
	--hypernetwork_input_dim)
	HYPERNETWORK_INPUT_DIM="$2"
	shift 2
	;;
	--emg_normalization_participant)
	EMG_NORMALIZATION_PARTICIPANT="$2"
	shift 2
	;;
	--first_frame_path)
	FIRST_FRAME_PATH="$2"
	shift 2
	;;
	--ref_video_path)
	REF_VIDEO_PATH="$2"
	shift 2
	;;
	--test_json_path)
	TEST_JSON_PATH="$2"
	shift 2
	;;
	--video_length)
	VIDEO_LENGTH="$2"
	shift 2
	;;
	--fps)
	FPS="$2"
	shift 2
	;;
	--prompt)
	PROMPT="$2"
	shift 2
	;;
	--negative_prompt)
	NEGATIVE_PROMPT="$2"
	shift 2
	;;
	--guidance_scale)
	GUIDANCE_SCALE="$2"
	shift 2
	;;
	--seed)
	SEED="$2"
	shift 2
	;;
	--num_inference_steps)
	NUM_INFERENCE_STEPS="$2"
	shift 2
	;;
	--lora_rank)
	LORA_RANK="$2"
	shift 2
	;;
	--lora_down_dim)
	LORA_DOWN_DIM="$2"
	shift 2
	;;
	--lora_up_dim)
	LORA_UP_DIM="$2"
	shift 2
	;;
	--lora_weight)
	LORA_WEIGHT="$2"
	shift 2
	;;
	--save_path)
	SAVE_PATH="$2"
	shift 2
	;;
	--gpu_memory_mode)
	GPU_MEMORY_MODE="$2"
	shift 2
	;;
	--weight_dtype)
	WEIGHT_DTYPE="$2"
	shift 2
	;;
	--sampler_name)
	SAMPLER_NAME="$2"
	shift 2
	;;
	--ulysses_degree)
	ULYSSES_DEGREE="$2"
	shift 2
	;;
	--ring_degree)
	RING_DEGREE="$2"
	shift 2
	;;
	--partial_video_length)
	PARTIAL_VIDEO_LENGTH="$2"
	shift 2
	;;
	--overlap_video_length)
	OVERLAP_VIDEO_LENGTH="$2"
	shift 2
	;;
	--help\|-h)
	echo "Usage: bash run_predict_i2v.sh [OPTIONS]"
	echo ""
	echo "Options:"
	echo " --model_name 模型路径 (默认: models/Diffusion_Transformer/CogVideoX-Fun-V1.1-2b-InP)"
	echo " --lilora_path LiLoRA 权重路径"
	echo " --hypernetwork_path HyperNetwork 权重路径"
	echo " --multimodal_ref_encoder_ckpt_path EMG 多模态参考编码器 checkpoint"
	echo " --multimodal_ref_encoder_script EMG 多模态参考编码器脚本"
	echo " --emg2pose_ckpt_path EMG2Pose checkpoint"
	echo " --emg2pose_repo EMG2Pose repo 路径"
	echo " --hypernetwork_input_dim HyperNetwork 输入维度"
	echo " --emg_normalization_participant EMG 归一化参与者目录"
	echo " --first_frame_path 首帧图像路径"
	echo " --ref_video_path 参考视频路径 (用于 HyperNetwork)"
	echo " --test_json_path 测试集 JSON 路径"
	echo " --video_length 生成视频帧数 (默认: 49)"
	echo " --fps 输出视频 FPS (默认: 8)"
	echo " --prompt 生成提示词"
	echo " --negative_prompt 负向提示词"
	echo " --guidance_scale 引导 scale (默认: 6.0)"
	echo " --seed 随机种子 (默认: 43)"
	echo " --num_inference_steps 推理步数 (默认: 50)"
	echo " --lora_rank LoRA rank (默认: 1)"
	echo " --lora_down_dim LoRA down dimension (默认: 200)"
	echo " --lora_up_dim LoRA up dimension (默认: 100)"
	echo " --lora_weight LoRA 权重 (默认: 0.55)"
	echo " --save_path 输出路径 (默认: samples/cogvideox-fun-videos_i2v)"
	echo " --gpu_memory_mode GPU 内存模式"
	echo " --weight_dtype 权重数据类型"
	echo " --sampler_name 采样器名称"
	echo " --help 显示帮助信息"
	exit 0
	;;
	*)
	echo "Unknown option: $1"
	exit 1
	;;
	esac
	done

	# =============================================================================
	# 构建命令行参数 (Build Command Line Arguments)
	# =============================================================================
	PYTHON_ARGS=()

	# GPU 设置
	PYTHON_ARGS+=("--gpu_memory_mode" "${GPU_MEMORY_MODE}")
	PYTHON_ARGS+=("--ulysses_degree" "${ULYSSES_DEGREE}")
	PYTHON_ARGS+=("--ring_degree" "${RING_DEGREE}")

	[[ "${FSPD_DIT}" == "True" ]] && PYTHON_ARGS+=("--fsdp_dit")
	[[ "${FSPD_TEXT_ENCODER}" == "True" ]] && PYTHON_ARGS+=("--fsdp_text_encoder")
	[[ "${COMPILE_DIT}" == "True" ]] && PYTHON_ARGS+=("--compile_dit")

	# 模型设置
	PYTHON_ARGS+=("--model_name" "${MODEL_NAME}")
	PYTHON_ARGS+=("--sampler_name" "${SAMPLER_NAME}")

	# LiLoRA 和 HyperNetwork
	[[ -n "${LILORA_PATH}" && "${LILORA_PATH}" != "None" ]] && PYTHON_ARGS+=("--lilora_path" "${LILORA_PATH}")
	[[ -n "${HYPERNETWORK_PATH}" && "${HYPERNETWORK_PATH}" != "None" ]] && PYTHON_ARGS+=("--hypernetwork_path" "${HYPERNETWORK_PATH}")
	[[ -n "${MULTIMODAL_REF_ENCODER_CKPT_PATH}" && "${MULTIMODAL_REF_ENCODER_CKPT_PATH}" != "None" ]] && PYTHON_ARGS+=("--multimodal_ref_encoder_ckpt_path" "${MULTIMODAL_REF_ENCODER_CKPT_PATH}")
	[[ -n "${MULTIMODAL_REF_ENCODER_SCRIPT}" && "${MULTIMODAL_REF_ENCODER_SCRIPT}" != "None" ]] && PYTHON_ARGS+=("--multimodal_ref_encoder_script" "${MULTIMODAL_REF_ENCODER_SCRIPT}")
	[[ -n "${EMG2POSE_CKPT_PATH}" && "${EMG2POSE_CKPT_PATH}" != "None" ]] && PYTHON_ARGS+=("--emg2pose_ckpt_path" "${EMG2POSE_CKPT_PATH}")
	[[ -n "${EMG2POSE_REPO}" && "${EMG2POSE_REPO}" != "None" ]] && PYTHON_ARGS+=("--emg2pose_repo" "${EMG2POSE_REPO}")
	[[ -n "${HYPERNETWORK_INPUT_DIM}" && "${HYPERNETWORK_INPUT_DIM}" != "None" ]] && PYTHON_ARGS+=("--hypernetwork_input_dim" "${HYPERNETWORK_INPUT_DIM}")
	[[ -n "${EMG_NORMALIZATION_PARTICIPANT}" && "${EMG_NORMALIZATION_PARTICIPANT}" != "None" ]] && PYTHON_ARGS+=("--emg_normalization_participant" "${EMG_NORMALIZATION_PARTICIPANT}")

	# LiLoRA 参数
	PYTHON_ARGS+=("--lora_rank" "${LORA_RANK}")
	PYTHON_ARGS+=("--lora_down_dim" "${LORA_DOWN_DIM}")
	PYTHON_ARGS+=("--lora_up_dim" "${LORA_UP_DIM}")
	[[ -n "${LORA_WEIGHT}" && "${LORA_WEIGHT}" != "None" ]] && PYTHON_ARGS+=("--lora_weight" "${LORA_WEIGHT}")

	# 视频参数
	PYTHON_ARGS+=("--video_length" "${VIDEO_LENGTH}")
	PYTHON_ARGS+=("--fps" "${FPS}")
	[[ -n "${PARTIAL_VIDEO_LENGTH}" && "${PARTIAL_VIDEO_LENGTH}" != "None" ]] && PYTHON_ARGS+=("--partial_video_length" "${PARTIAL_VIDEO_LENGTH}")
	PYTHON_ARGS+=("--overlap_video_length" "${OVERLAP_VIDEO_LENGTH}")

	# 权重类型
	PYTHON_ARGS+=("--weight_dtype" "${WEIGHT_DTYPE}")

	# 推理参数 (单样本时使用，批量时从 JSON 读取)
	if [[ -z "${TEST_JSON_PATH}" \|\| "${TEST_JSON_PATH}" == "None" ]]; then
	[[ -n "${FIRST_FRAME_PATH}" && "${FIRST_FRAME_PATH}" != "None" ]] && PYTHON_ARGS+=("--first_frame_path" "${FIRST_FRAME_PATH}")
	[[ -n "${REF_VIDEO_PATH}" && "${REF_VIDEO_PATH}" != "None" ]] && PYTHON_ARGS+=("--ref_video_path" "${REF_VIDEO_PATH}")
	[[ -n "${PROMPT}" && "${PROMPT}" != "None" ]] && PYTHON_ARGS+=("--prompt" "${PROMPT}")
	PYTHON_ARGS+=("--negative_prompt" "${NEGATIVE_PROMPT}")
	PYTHON_ARGS+=("--guidance_scale" "${GUIDANCE_SCALE}")
	PYTHON_ARGS+=("--seed" "${SEED}")
	PYTHON_ARGS+=("--num_inference_steps" "${NUM_INFERENCE_STEPS}")
	else
	PYTHON_ARGS+=("--test_json_path" "${TEST_JSON_PATH}")
	fi

	# 输出路径
	PYTHON_ARGS+=("--save_path" "${SAVE_PATH}")

	# =============================================================================
	# 运行推理脚本 (Run Inference Script)
	# =============================================================================
	echo "========================================"
	echo "CogVideoX-Fun I2V Inference"
	echo "========================================"
	echo "Model: ${MODEL_NAME}"
	echo "LiLoRA: ${LILORA_PATH}"
	echo "HyperNetwork: ${HYPERNETWORK_PATH}"
	echo "Multimodal Ref Encoder: ${MULTIMODAL_REF_ENCODER_CKPT_PATH}"
	echo "Video Length: ${VIDEO_LENGTH}"
	if [[ -n "${TEST_JSON_PATH}" && "${TEST_JSON_PATH}" != "None" ]]; then
	echo "Test Set: ${TEST_JSON_PATH}"
	else
	echo "First Frame: ${FIRST_FRAME_PATH}"
	echo "Prompt: ${PROMPT}"
	fi
	echo "========================================"

	torchrun --nproc-per-node=1 --master_port=29501\
	/home/xhm001/Desktop/Code/VideoX-Fun-main/examples/cogvideox_fun/predict_i2v.py \
	"${PYTHON_ARGS[@]}"


	echo ""
	echo "Inference completed!"