run_eval_retrieval_cpu.sh

#!/bin/bash
# ============================================================================
# ASR 评估脚本（GPU/CPU + 原文库检索/生成结合）
#
# 目标：
# - 直接在服务器跑 evaluate_model.py（默认 CUDA）
# - 默认开启原文库检索 + 局部 span 抽取 + 检索/生成结合
# - 路径可通过环境变量覆盖
#
# 常用环境变量：
#   ASR_BASE_MODEL
#   ASR_LORA_PATH
#   ASR_TEST_FILE
#   ASR_DEVICE (auto/cuda/cpu)
#   ASR_CUDA_VISIBLE_DEVICES
#   ASR_CONDA_ENV (默认: verl；设为 none 可跳过 conda 激活)
#   ASR_CONDA_SH  (可选: conda.sh 绝对路径)
#   ASR_PYTHON    (可选: 指定 python 可执行文件)
#   ASR_FORCE_REPLACE (默认: 0；设 1 强制检索直替换)
#   ASR_RETRIEVAL_MATCH_MODE (默认: doc_span)
#   ASR_RETRIEVAL_SOURCE_FILES (默认: 补充小库 + Classical-Modern 原文源 + chinese-poetry)
#   ASR_RETRIEVAL_CORPUS (默认: none；candidate 模式下可指向预构建语料)
#   ASR_REBUILD_RETRIEVAL_CORPUS (默认: 0；仅 candidate 模式下生效)
#   ASR_RETRIEVAL_DOC_MAX_LEN (默认: 512)
#   ASR_RETRIEVAL_DOC_TOP_K (默认: 8)
#   ASR_RETRIEVAL_LOCAL_CANDIDATE_K (默认: 12)
#   ASR_RETRIEVAL_ENABLE_PATCH (默认: 0；默认关闭 patch，只保留更稳的 full 候选接管)
#   ASR_RETRIEVAL_MIN_FULL_SPAN_RATIO (默认: 0.90；短于该比例的 full span 直接丢弃)
#   ASR_RETRIEVAL_PREFER_FULL_MIN_SCORE (默认: 0.45)
#   ASR_RETRIEVAL_FULL_MIN_SPAN_RATIO / ASR_RETRIEVAL_FULL_MAX_SPAN_RATIO
#   ASR_RETRIEVAL_SHORT_QUERY_MAX_LEN / ASR_RETRIEVAL_SHORT_QUERY_MIN_LOCAL_SCORE
#   ASR_RETRIEVAL_PATCH_MIN_SCORE / ASR_RETRIEVAL_PATCH_USE_ALIGN_SCORE / ASR_RETRIEVAL_PATCH_MARGIN / ASR_RETRIEVAL_PATCH_MAX_EDIT_RATIO
#   ASR_RETRIEVAL_MAX_CANDIDATES (默认: 1000000)
#   ASR_RETRIEVAL_MIN_SCORE (默认: 0.45)
#   ASR_RETRIEVAL_MARGIN (默认: 0.10)
#   ASR_RETRIEVAL_MAX_EDIT_RATIO (默认: 0.50)
#   ASR_EXCLUDE_EVAL_TARGETS (默认: 0；如需开启再设 1)
#   ASR_SAVE_ROOT / ASR_OUTPUT_ROOT / ASR_LOG_DIR
#   ASR_MAX_SAMPLES
# ============================================================================

set -euo pipefail
set -x

echo "=========================================="
echo "ASR 评估（默认CUDA + 原文库局部匹配）"
echo "=========================================="

SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
BASE_DIR="${ASR_BASE_DIR:-${SCRIPT_DIR}}"

if [ ! -f "${BASE_DIR}/evaluate_model.py" ]; then
    echo "ERROR: evaluate_model.py not found under ${BASE_DIR}"
    exit 1
fi

# 默认自动激活 verl；设 ASR_CONDA_ENV=none 可关闭
CONDA_ENV="${ASR_CONDA_ENV:-verl}"
if [ "${CONDA_ENV}" != "none" ]; then
    CONDA_SH=""
    if [ -n "${ASR_CONDA_SH:-}" ] && [ -f "${ASR_CONDA_SH}" ]; then
        CONDA_SH="${ASR_CONDA_SH}"
    else
        for c in \
            "/root/miniconda3/etc/profile.d/conda.sh" \
            "${HOME}/miniconda3/etc/profile.d/conda.sh" \
            "/opt/conda/etc/profile.d/conda.sh"; do
            if [ -f "${c}" ]; then
                CONDA_SH="${c}"
                break
            fi
        done
    fi

    if [ -n "${CONDA_SH}" ]; then
        source "${CONDA_SH}"
        if ! conda activate "${CONDA_ENV}"; then
            echo "ERROR: conda activate failed for env: ${CONDA_ENV}"
            exit 1
        fi
    else
        echo "WARNING: conda.sh not found, skip conda activation."
        echo "         You can set ASR_CONDA_SH=/path/to/conda.sh"
    fi
fi

PYTHON_BIN="${ASR_PYTHON:-python}"
FORCE_REPLACE="${ASR_FORCE_REPLACE:-0}"
RETRIEVAL_MATCH_MODE="${ASR_RETRIEVAL_MATCH_MODE:-doc_span}"
DEFAULT_RETRIEVAL_SOURCE_FILES="${BASE_DIR}/retrieval_extra/liezi.jsonl,${BASE_DIR}/retrieval_extra/textbook_prose_manual.jsonl,${BASE_DIR}/retrieval_extra/textbook_poetry_manual.jsonl,${BASE_DIR}/retrieval_extra/yuefu_history_manual.jsonl,${BASE_DIR}/retrieval_extra/yuanqu_dialogue_manual.jsonl,${BASE_DIR}/retrieval_extra/classical_modern_originals.jsonl,${BASE_DIR}/chinese-poetry"
RETRIEVAL_SOURCE_FILES="${ASR_RETRIEVAL_SOURCE_FILES:-${DEFAULT_RETRIEVAL_SOURCE_FILES}}"
RETRIEVAL_CORPUS="${ASR_RETRIEVAL_CORPUS:-none}"
REBUILD_RETRIEVAL_CORPUS="${ASR_REBUILD_RETRIEVAL_CORPUS:-0}"
RETRIEVAL_DOC_MAX_LEN="${ASR_RETRIEVAL_DOC_MAX_LEN:-512}"
RETRIEVAL_DOC_TOP_K="${ASR_RETRIEVAL_DOC_TOP_K:-8}"
RETRIEVAL_LOCAL_CANDIDATE_K="${ASR_RETRIEVAL_LOCAL_CANDIDATE_K:-12}"
RETRIEVAL_ENABLE_PATCH="${ASR_RETRIEVAL_ENABLE_PATCH:-0}"
RETRIEVAL_MIN_FULL_SPAN_RATIO="${ASR_RETRIEVAL_MIN_FULL_SPAN_RATIO:-0.90}"
RETRIEVAL_PREFER_FULL_MIN_SCORE="${ASR_RETRIEVAL_PREFER_FULL_MIN_SCORE:-0.45}"
RETRIEVAL_FULL_MIN_SPAN_RATIO="${ASR_RETRIEVAL_FULL_MIN_SPAN_RATIO:-0.90}"
RETRIEVAL_FULL_MAX_SPAN_RATIO="${ASR_RETRIEVAL_FULL_MAX_SPAN_RATIO:-1.25}"
RETRIEVAL_SHORT_QUERY_MAX_LEN="${ASR_RETRIEVAL_SHORT_QUERY_MAX_LEN:-8}"
RETRIEVAL_SHORT_QUERY_MIN_LOCAL_SCORE="${ASR_RETRIEVAL_SHORT_QUERY_MIN_LOCAL_SCORE:-0.52}"
RETRIEVAL_PATCH_MIN_SCORE="${ASR_RETRIEVAL_PATCH_MIN_SCORE:-0.65}"
RETRIEVAL_PATCH_USE_ALIGN_SCORE="${ASR_RETRIEVAL_PATCH_USE_ALIGN_SCORE:-0.80}"
RETRIEVAL_PATCH_MARGIN="${ASR_RETRIEVAL_PATCH_MARGIN:-1.00}"
RETRIEVAL_PATCH_MAX_EDIT_RATIO="${ASR_RETRIEVAL_PATCH_MAX_EDIT_RATIO:-0.20}"
RETRIEVAL_MAX_CANDIDATES="${ASR_RETRIEVAL_MAX_CANDIDATES:-1000000}"
RETRIEVAL_MIN_SCORE="${ASR_RETRIEVAL_MIN_SCORE:-0.45}"
RETRIEVAL_MARGIN="${ASR_RETRIEVAL_MARGIN:-0.10}"
RETRIEVAL_MAX_EDIT_RATIO="${ASR_RETRIEVAL_MAX_EDIT_RATIO:-0.50}"
EXCLUDE_EVAL_TARGETS="${ASR_EXCLUDE_EVAL_TARGETS:-0}"
BASE_MODEL="${ASR_BASE_MODEL:-${BASE_DIR}/ChineseErrorCorrector3-4B}"
LORA_PATH="${ASR_LORA_PATH:-./asr/check/asr_poetry_lora_20260308_191410}"
DEVICE="${ASR_DEVICE:-cuda}"

if [ -n "${ASR_CUDA_VISIBLE_DEVICES:-}" ]; then
    export CUDA_VISIBLE_DEVICES="${ASR_CUDA_VISIBLE_DEVICES}"
fi

if [ -n "${ASR_TEST_FILE:-}" ]; then
    TEST_FILE="${ASR_TEST_FILE}"
elif [ -f "${BASE_DIR}/train_data_v4/test_real_asr.jsonl" ]; then
    TEST_FILE="${BASE_DIR}/train_data_v4/test_real_asr.jsonl"
else
    TEST_FILE="${BASE_DIR}/train_data_v3/test_real_asr.jsonl"
fi

TIMESTAMP="$(date +%Y%m%d_%H%M%S)"
SAVE_ROOT="${ASR_SAVE_ROOT:-./asr/check}"
OUTPUT_ROOT="${ASR_OUTPUT_ROOT:-${SAVE_ROOT}}"
LOG_DIR="${ASR_LOG_DIR:-${OUTPUT_ROOT}/logs}"
DEVICE_TAG="${DEVICE//[^a-zA-Z0-9_-]/_}"
OUTPUT_DIR="${ASR_EVAL_OUTPUT_DIR:-${OUTPUT_ROOT}/asr_eval_retrieval_${DEVICE_TAG}_${TIMESTAMP}}"
OUTPUT_FILE="${ASR_OUTPUT_FILE:-${OUTPUT_DIR}/evaluation_results.jsonl}"
LOG_FILE="${LOG_DIR}/eval_retrieval_${DEVICE_TAG}_${TIMESTAMP}.log"

mkdir -p "${OUTPUT_DIR}" "${LOG_DIR}"

if [ ! -d "${BASE_MODEL}" ]; then
    echo "ERROR: base model not found: ${BASE_MODEL}"
    exit 1
fi

if [ ! -d "${LORA_PATH}" ]; then
    echo "ERROR: LoRA path not found: ${LORA_PATH}"
    exit 1
fi

if [ ! -f "${TEST_FILE}" ]; then
    echo "ERROR: test file not found: ${TEST_FILE}"
    exit 1
fi

if ! "${PYTHON_BIN}" -c "import torch, transformers, peft" >/dev/null 2>&1; then
    echo "ERROR: missing python deps (torch/transformers/peft)."
    echo "       tried python: ${PYTHON_BIN}"
    echo "       set ASR_CONDA_ENV / ASR_CONDA_SH, or ASR_PYTHON manually."
    exit 1
fi

USE_RETRIEVAL_CORPUS=1
if [ "${RETRIEVAL_CORPUS}" = "none" ]; then
    USE_RETRIEVAL_CORPUS=0
fi

if [ "${USE_RETRIEVAL_CORPUS}" = "1" ]; then
    if [ "${RETRIEVAL_MATCH_MODE}" != "candidate" ]; then
        if [ ! -f "${RETRIEVAL_CORPUS}" ]; then
            echo "WARNING: doc_span 模式不自动构建规范化语料，改为直接读取原文源。"
            USE_RETRIEVAL_CORPUS=0
        fi
    elif [[ "${RETRIEVAL_SOURCE_FILES}" == *,* ]] || [ ! -d "${RETRIEVAL_SOURCE_FILES}" ]; then
        echo "WARNING: retrieval source is not a single directory, skip normalized corpus build."
        USE_RETRIEVAL_CORPUS=0
    elif [ "${REBUILD_RETRIEVAL_CORPUS}" = "1" ] || [ ! -f "${RETRIEVAL_CORPUS}" ]; then
        if [ ! -f "${BASE_DIR}/build_poetry_retrieval_corpus.py" ]; then
            echo "ERROR: build_poetry_retrieval_corpus.py not found under ${BASE_DIR}"
            exit 1
        fi
        "${PYTHON_BIN}" "${BASE_DIR}/build_poetry_retrieval_corpus.py" \
            --poetry_dir "${RETRIEVAL_SOURCE_FILES}" \
            --output_file "${RETRIEVAL_CORPUS}"
    fi
fi

echo "Base dir:    ${BASE_DIR}"
echo "Base model:  ${BASE_MODEL}"
echo "LoRA path:   ${LORA_PATH}"
echo "Device:      ${DEVICE}"
echo "Python:      ${PYTHON_BIN}"
echo "ForceReplace:${FORCE_REPLACE}"
echo "RetrMode:    ${RETRIEVAL_MATCH_MODE}"
echo "RetrSource:  ${RETRIEVAL_SOURCE_FILES}"
echo "RetrCorpus:  ${RETRIEVAL_CORPUS}"
echo "UseCorpus:   ${USE_RETRIEVAL_CORPUS}"
echo "DocMaxLen:   ${RETRIEVAL_DOC_MAX_LEN}"
echo "DocTopK:     ${RETRIEVAL_DOC_TOP_K}"
echo "LocalCandK:  ${RETRIEVAL_LOCAL_CANDIDATE_K}"
echo "Patch:       ${RETRIEVAL_ENABLE_PATCH}"
echo "MinFullRat:  ${RETRIEVAL_MIN_FULL_SPAN_RATIO}"
echo "PreferFull:  ${RETRIEVAL_PREFER_FULL_MIN_SCORE}"
echo "FullSpanRat: ${RETRIEVAL_FULL_MIN_SPAN_RATIO}-${RETRIEVAL_FULL_MAX_SPAN_RATIO}"
echo "ShortQuery:  ${RETRIEVAL_SHORT_QUERY_MAX_LEN}/${RETRIEVAL_SHORT_QUERY_MIN_LOCAL_SCORE}"
echo "PatchScore:  ${RETRIEVAL_PATCH_MIN_SCORE}"
echo "PatchAlign:  ${RETRIEVAL_PATCH_USE_ALIGN_SCORE}"
echo "PatchMargin: ${RETRIEVAL_PATCH_MARGIN}"
echo "PatchEdit:   ${RETRIEVAL_PATCH_MAX_EDIT_RATIO}"
echo "MaxCand:     ${RETRIEVAL_MAX_CANDIDATES}"
echo "MinScore:    ${RETRIEVAL_MIN_SCORE}"
echo "Margin:      ${RETRIEVAL_MARGIN}"
echo "MaxEditRatio:${RETRIEVAL_MAX_EDIT_RATIO}"
echo "ExcludeEval: ${EXCLUDE_EVAL_TARGETS}"
echo "Test file:   ${TEST_FILE}"
echo "Output file: ${OUTPUT_FILE}"
echo "Log file:    ${LOG_FILE}"

if command -v nvidia-smi >/dev/null 2>&1; then
    echo "GPU:"
    nvidia-smi --query-gpu=name,memory.total --format=csv,noheader || true
fi

CMD=(
    "${PYTHON_BIN}" "${BASE_DIR}/evaluate_model.py"
    --device "${DEVICE}"
    --enable_retrieval
    --retrieval_match_mode "${RETRIEVAL_MATCH_MODE}"
    --retrieval_source_files "${RETRIEVAL_SOURCE_FILES}"
    --retrieval_doc_max_len "${RETRIEVAL_DOC_MAX_LEN}"
    --retrieval_doc_top_k "${RETRIEVAL_DOC_TOP_K}"
    --retrieval_local_candidate_k "${RETRIEVAL_LOCAL_CANDIDATE_K}"
    --retrieval_min_full_span_ratio "${RETRIEVAL_MIN_FULL_SPAN_RATIO}"
    --retrieval_prefer_full_candidate_min_score "${RETRIEVAL_PREFER_FULL_MIN_SCORE}"
    --retrieval_full_min_span_ratio "${RETRIEVAL_FULL_MIN_SPAN_RATIO}"
    --retrieval_full_max_span_ratio "${RETRIEVAL_FULL_MAX_SPAN_RATIO}"
    --retrieval_short_query_max_len "${RETRIEVAL_SHORT_QUERY_MAX_LEN}"
    --retrieval_short_query_min_local_score "${RETRIEVAL_SHORT_QUERY_MIN_LOCAL_SCORE}"
    --retrieval_patch_min_score "${RETRIEVAL_PATCH_MIN_SCORE}"
    --retrieval_patch_use_align_score "${RETRIEVAL_PATCH_USE_ALIGN_SCORE}"
    --retrieval_patch_margin "${RETRIEVAL_PATCH_MARGIN}"
    --retrieval_patch_max_edit_ratio "${RETRIEVAL_PATCH_MAX_EDIT_RATIO}"
    --retrieval_max_candidates "${RETRIEVAL_MAX_CANDIDATES}"
    --retrieval_min_score "${RETRIEVAL_MIN_SCORE}"
    --retrieval_margin "${RETRIEVAL_MARGIN}"
    --retrieval_max_edit_ratio "${RETRIEVAL_MAX_EDIT_RATIO}"
    --base_model "${BASE_MODEL}"
    --lora_path "${LORA_PATH}"
    --test_file "${TEST_FILE}"
    --output_file "${OUTPUT_FILE}"
)

if [ "${USE_RETRIEVAL_CORPUS}" = "1" ]; then
    CMD+=(--retrieval_corpus "${RETRIEVAL_CORPUS}")
fi

if [ "${RETRIEVAL_ENABLE_PATCH}" = "1" ] || [ "${RETRIEVAL_ENABLE_PATCH}" = "true" ] || [ "${RETRIEVAL_ENABLE_PATCH}" = "TRUE" ]; then
    CMD+=(--retrieval_enable_patch)
fi

if [ "${FORCE_REPLACE}" = "1" ] || [ "${FORCE_REPLACE}" = "true" ] || [ "${FORCE_REPLACE}" = "TRUE" ]; then
    CMD+=(--retrieval_force_replace)
fi

if [ "${EXCLUDE_EVAL_TARGETS}" = "1" ] || [ "${EXCLUDE_EVAL_TARGETS}" = "true" ] || [ "${EXCLUDE_EVAL_TARGETS}" = "TRUE" ]; then
    CMD+=(--exclude_eval_targets_from_retrieval)
fi

if [ -n "${ASR_MAX_SAMPLES:-}" ]; then
    CMD+=(--max_samples "${ASR_MAX_SAMPLES}")
fi

# 允许在脚本末尾追加 evaluate_model.py 参数
# 例如：bash run_eval_retrieval_cpu.sh --retrieval_max_candidates 50000
if [ "$#" -gt 0 ]; then
    CMD+=("$@")
fi

set +e
"${CMD[@]}" 2>&1 | tee "${LOG_FILE}"
EXIT_CODE=${PIPESTATUS[0]}
set -e

echo ""
echo "=========================================="
echo "评估完成"
echo "Exit code:   ${EXIT_CODE}"
echo "Result file: ${OUTPUT_FILE}"
echo "Log file:    ${LOG_FILE}"
echo "=========================================="

exit "${EXIT_CODE}"