analysis/failure_case_analysis/ov_coco/scripts/run_comparison.sh

#!/bin/bash

# OV-COCO Failure Case Analysis - 完整流程脚本
# 包含：推理 -> 转换 -> 对比分析 -> 可视化 -> 生成报告
#
# 重要：使用官方 F-ViT test.py 进行推理，确保结果与评估指标一致
#
# Usage:
#   bash run_comparison.sh [BACKBONE] [NUM_GPUS]
#
# Arguments:
#   BACKBONE: vitb16 或 vitl14 (默认: vitb16)
#   NUM_GPUS: 使用的GPU数量 (默认: 8)
#
# Examples:
#   bash run_comparison.sh vitb16 8    # 使用ViT-B/16，8卡
#   bash run_comparison.sh vitl14 8    # 使用ViT-L/14，8卡

set -e

# 参数解析
BACKBONE=${1:-vitb16}
NUM_GPUS=${2:-8}
PORT=${PORT:-12346}

# 路径设置
BASE_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
CLIPSELF_DIR="$(cd "$BASE_DIR/../../CLIPSelf/F-ViT" && pwd)"

# 数据集路径
ANN_FILE="$CLIPSELF_DIR/data/coco/zero-shot/instances_val2017_all_2.json"
IMG_DIR="$CLIPSELF_DIR/data/coco/val2017"
SEEN_CLASSES="$CLIPSELF_DIR/datasets/mscoco_seen_classes.json"

# 模型配置
declare -A DECLIP_CKPTS=(
    ["vitb16"]="/opt/tiger/xiaomoguhzz/declip2_ovcoco_detector/EVAB_dinov2B_epoch2.pth"
    ["vitl14"]="/opt/tiger/xiaomoguhzz/declip2_ovcoco_detector/EVAL_dinov2L_epoch3.pth"
)

declare -A CLIPSELF_CKPTS=(
    ["vitb16"]="/opt/tiger/xiaomoguhzz/fvit_eva_vitb16_ovcoco_clipself_proposals.pth"
    ["vitl14"]="/opt/tiger/xiaomoguhzz/fvit_eva_vitl14_ovcoco_clipself_proposals.pth"
)

declare -A CLIP_CKPTS=(
    ["vitb16"]="$CLIPSELF_DIR/work_dirs/clearclip_ovcoco/epoch_3.pth"
    ["vitl14"]="$CLIPSELF_DIR/work_dirs/clearclip_ovcoco_large/epoch_3.pth"
)

echo "========================================"
echo "OV-COCO Failure Case Analysis Pipeline"
echo "========================================"
echo "Backbone: $BACKBONE"
echo "Num GPUs: $NUM_GPUS"
echo "Base Dir: $BASE_DIR"
echo ""

# 检查必要文件
echo "[Checking] Required files..."
if [ ! -f "$ANN_FILE" ]; then
    echo "  Warning: Annotation file not found (may be a symlink on dev machine)"
    echo "  Path: $ANN_FILE"
fi
if [ ! -f "$SEEN_CLASSES" ]; then
    echo "  Warning: Seen classes file not found"
    echo "  Path: $SEEN_CLASSES"
fi

# 结果目录
DECLIP_PKL="$BASE_DIR/results/declip_${BACKBONE}/predictions.pkl"
CLIPSELF_PKL="$BASE_DIR/results/clipself_${BACKBONE}/predictions.pkl"
CLIP_PKL="$BASE_DIR/results/clearclip_${BACKBONE}/predictions.pkl"

DECLIP_PRED="$BASE_DIR/results/declip_${BACKBONE}/predictions.json"
CLIPSELF_PRED="$BASE_DIR/results/clipself_${BACKBONE}/predictions.json"
CLIP_PRED="$BASE_DIR/results/clearclip_${BACKBONE}/predictions.json"

COMPARISON_OUTPUT="$BASE_DIR/analysis_output/comparison_${BACKBONE}"
VISUALIZATION_OUTPUT="$BASE_DIR/analysis_output/visualizations_${BACKBONE}"
REPORT_OUTPUT="$BASE_DIR/analysis_output/paper_report_${BACKBONE}"

# ============================================================================
# Step 1: 使用官方 F-ViT test.py 推理 (保存pkl)
# ============================================================================
echo ""
echo "========================================"
echo "Step 1: Official F-ViT Inference"
echo "========================================"

run_official_inference() {
    local model_name=$1
    local config=$2
    local checkpoint=$3
    local output_pkl=$4

    if [ -f "$output_pkl" ]; then
        echo "  [$model_name] Results exist: $output_pkl, skipping..."
        return
    fi

    if [ ! -f "$checkpoint" ]; then
        echo "  [$model_name] Error: Checkpoint not found: $checkpoint"
        return
    fi

    echo "  [$model_name] Running official inference..."
    echo "    Config: $config"
    echo "    Checkpoint: $checkpoint"

    local work_dir="/opt/tiger/fvit_eval_${model_name}_${BACKBONE}"
    local tmpdir="/opt/tiger/mmdet_tmpdir_${model_name}_${BACKBONE}"
    mkdir -p "$work_dir" "$tmpdir" "$(dirname $output_pkl)"

    cd "$CLIPSELF_DIR"

    torchrun \
        --nnodes=1 \
        --node_rank=0 \
        --nproc_per_node=$NUM_GPUS \
        --master_addr=127.0.0.1 \
        --master_port=$PORT \
        "$CLIPSELF_DIR/test.py" \
        "$config" \
        "$checkpoint" \
        --launcher pytorch \
        --work-dir "$work_dir" \
        --out "$output_pkl" \
        --eval bbox \
        --tmpdir "$tmpdir" \
        --cfg-options data.test.ann_file="$ANN_FILE" data.test.img_prefix="$IMG_DIR"

    echo "  [$model_name] Done! Saved to $output_pkl"

    # 更新端口避免冲突
    PORT=$((PORT + 1))
}

# 运行三个模型的推理
run_official_inference "declip" \
    "$BASE_DIR/configs/declip_${BACKBONE}_ovcoco.py" \
    "${DECLIP_CKPTS[$BACKBONE]}" \
    "$DECLIP_PKL"

run_official_inference "clipself" \
    "$BASE_DIR/configs/clipself_${BACKBONE}_ovcoco.py" \
    "${CLIPSELF_CKPTS[$BACKBONE]}" \
    "$CLIPSELF_PKL"

run_official_inference "clearclip" \
    "$BASE_DIR/configs/clearclip_${BACKBONE}_ovcoco.py" \
    "${CLIP_CKPTS[$BACKBONE]}" \
    "$CLIP_PKL"

# ============================================================================
# Step 1.5: 转换 pkl 到 json
# ============================================================================
echo ""
echo "========================================"
echo "Step 1.5: Convert pkl to json"
echo "========================================"

cd "$BASE_DIR"
python convert_pkl_to_json.py --backbone "$BACKBONE"

# ============================================================================
# Step 2: 跨模型对比分析
# ============================================================================
echo ""
echo "========================================"
echo "Step 2: Cross-Model Comparison Analysis"
echo "========================================"

if [ ! -f "$DECLIP_PRED" ] || [ ! -f "$CLIPSELF_PRED" ] || [ ! -f "$CLIP_PRED" ]; then
    echo "Error: Some prediction files are missing. Please run inference first."
    echo "  DeCLIP: $DECLIP_PRED"
    echo "  CLIPSelf: $CLIPSELF_PRED"
    echo "  CLIP: $CLIP_PRED"
    exit 1
fi

cd "$BASE_DIR"
python compare_models_v2.py \
    --declip-pred "$DECLIP_PRED" \
    --clipself-pred "$CLIPSELF_PRED" \
    --clip-pred "$CLIP_PRED" \
    --ann-file "$ANN_FILE" \
    --seen-classes "$SEEN_CLASSES" \
    --output "$COMPARISON_OUTPUT" \
    --top-k 50

echo "Comparison analysis saved to: $COMPARISON_OUTPUT"

# ============================================================================
# Step 3: 可视化（优势 + 未解决）
# ============================================================================
echo ""
echo "========================================"
echo "Step 3: Visualization (Advantage + Unsolved)"
echo "========================================"

python visualize_comparison.py \
    --cases "$COMPARISON_OUTPUT/top_advantage_cases.json" \
    --ann-file "$ANN_FILE" \
    --img-dir "$IMG_DIR" \
    --output "$VISUALIZATION_OUTPUT/advantage" \
    --case-type advantage \
    --dpi 150

python visualize_comparison.py \
    --cases "$COMPARISON_OUTPUT/unsolved_small_novel_cases.json" \
    --ann-file "$ANN_FILE" \
    --img-dir "$IMG_DIR" \
    --output "$VISUALIZATION_OUTPUT/unsolved" \
    --case-type unsolved \
    --dpi 150

echo "Visualizations saved to:"
echo "  - $VISUALIZATION_OUTPUT/advantage"
echo "  - $VISUALIZATION_OUTPUT/unsolved"

# ============================================================================
# Step 4: 生成论文报告
# ============================================================================
echo ""
echo "========================================"
echo "Step 4: Generate Paper Report"
echo "========================================"

python generate_paper_report.py \
    --statistics "$COMPARISON_OUTPUT/statistics.json" \
    --output "$REPORT_OUTPUT" \
    --dpi 300

echo "Report saved to: $REPORT_OUTPUT"

# ============================================================================
# 完成
# ============================================================================
echo ""
echo "========================================"
echo "Pipeline Complete!"
echo "========================================"
echo ""
echo "Output directories:"
echo "  - Comparison analysis: $COMPARISON_OUTPUT"
echo "  - Visualizations: $VISUALIZATION_OUTPUT"
echo "  - Paper report: $REPORT_OUTPUT"
echo ""
echo "Key files:"
echo "  - Top 50 advantage cases: $COMPARISON_OUTPUT/top_advantage_cases.json"
echo "  - Advantage index: $VISUALIZATION_OUTPUT/advantage/index.html"
echo "  - Unsolved index: $VISUALIZATION_OUTPUT/unsolved/index.html"
echo "  - LaTeX tables: $REPORT_OUTPUT/tables.tex"
echo "  - Classification accuracy chart: $REPORT_OUTPUT/cls_acc_comparison.pdf"
echo ""