Datasets:

xiaomoguhzz
/

DeCLIP-TPAMI

Tags:

File size: 11,344 Bytes

d8c26e5

#!/bin/bash
# TRT 加速检测器性能对比实验 - 一键运行脚本
# 对比 DeCLIP (csa模式) vs CLIP (vanilla模式) 在 TRT 加速前后的速度和精度

set -e

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

# ============================================================
# 配置
# ============================================================
DECLIP_ROOT="$(dirname "$SCRIPT_DIR")"
CHECKPOINT_DIR="/mnt/SSD8T/home/wjj/code/my_CLIPSelf/checkpoints"
COCO_ROOT="/mnt/SSD8T/home/wjj/dataset/standard_coco"

# 模型检查点
EVACLIP_B16="${CHECKPOINT_DIR}/EVA02_CLIP_B_psz16_s8B.pt"
EVACLIP_L14="${CHECKPOINT_DIR}/EVA02_CLIP_L_psz14_336_s8B.pt"

# 测试配置
IMAGE_SIZE=560
BATCH_SIZE=1
WARMUP=10
ITERATIONS=100

# 输出目录
OUTPUT_DIR="results"
ENGINE_DIR="engines"

# ============================================================
# 函数
# ============================================================
log() {
    echo "[$(date '+%Y-%m-%d %H:%M:%S')] $1"
}

check_dependencies() {
    log "Checking dependencies..."
    
    python3 -c "import torch; print(f'PyTorch: {torch.__version__}')"
    python3 -c "import tensorrt; print(f'TensorRT: {tensorrt.__version__}')" 2>/dev/null || echo "Warning: TensorRT not installed"
    python3 -c "import onnx; print(f'ONNX: {onnx.__version__}')" 2>/dev/null || echo "Warning: ONNX not installed"
    python3 -c "import onnxruntime; print(f'ONNX Runtime: {onnxruntime.__version__}')" 2>/dev/null || echo "Warning: ONNX Runtime not installed"
    
    # 检查 CUDA
    if command -v nvidia-smi &> /dev/null; then
        nvidia-smi --query-gpu=name,memory.total,driver_version --format=csv,noheader | head -1
    fi
}

convert_models() {
    log "Converting models to TensorRT..."
    
    mkdir -p "$ENGINE_DIR"
    
    # CLIP (vanilla mode)
    log "Converting CLIP (vanilla mode)..."
    python3 convert_model.py \
        --model-type clip \
        --mode vanilla \
        --model-name EVA02-CLIP-B-16 \
        --checkpoint "$EVACLIP_B16" \
        --image-size $IMAGE_SIZE \
        --output-dir "$ENGINE_DIR" \
        --output-name clip_vanilla_${IMAGE_SIZE} \
        --fp16 \
        --dynamic \
        --verify || log "Warning: CLIP conversion failed"
    
    # DeCLIP (csa mode)
    log "Converting DeCLIP (csa mode)..."
    python3 convert_model.py \
        --model-type declip \
        --mode csa \
        --model-name EVA02-CLIP-B-16 \
        --checkpoint "$EVACLIP_B16" \
        --image-size $IMAGE_SIZE \
        --output-dir "$ENGINE_DIR" \
        --output-name declip_csa_${IMAGE_SIZE} \
        --fp16 \
        --dynamic \
        --verify || log "Warning: DeCLIP conversion failed"
}

benchmark_speed() {
    log "Running speed benchmarks..."
    
    mkdir -p "$OUTPUT_DIR"
    
    # PyTorch benchmarks
    log "Benchmarking PyTorch (vanilla mode)..."
    python3 benchmark_speed.py \
        --pytorch-model "$EVACLIP_B16" \
        --model-name EVA02-CLIP-B-16 \
        --mode vanilla \
        --image-size $IMAGE_SIZE $IMAGE_SIZE \
        --batch-size $BATCH_SIZE \
        --warmup $WARMUP \
        --iterations $ITERATIONS \
        --output "$OUTPUT_DIR/speed_pytorch_vanilla.json" || true
    
    log "Benchmarking PyTorch (csa mode)..."
    python3 benchmark_speed.py \
        --pytorch-model "$EVACLIP_B16" \
        --model-name EVA02-CLIP-B-16 \
        --mode csa \
        --image-size $IMAGE_SIZE $IMAGE_SIZE \
        --batch-size $BATCH_SIZE \
        --warmup $WARMUP \
        --iterations $ITERATIONS \
        --output "$OUTPUT_DIR/speed_pytorch_csa.json" || true
    
    log "Benchmarking PyTorch FP16 (vanilla mode)..."
    python3 benchmark_speed.py \
        --pytorch-model "$EVACLIP_B16" \
        --model-name EVA02-CLIP-B-16 \
        --mode vanilla \
        --image-size $IMAGE_SIZE $IMAGE_SIZE \
        --batch-size $BATCH_SIZE \
        --warmup $WARMUP \
        --iterations $ITERATIONS \
        --fp16 \
        --output "$OUTPUT_DIR/speed_pytorch_vanilla_fp16.json" || true
    
    log "Benchmarking PyTorch FP16 (csa mode)..."
    python3 benchmark_speed.py \
        --pytorch-model "$EVACLIP_B16" \
        --model-name EVA02-CLIP-B-16 \
        --mode csa \
        --image-size $IMAGE_SIZE $IMAGE_SIZE \
        --batch-size $BATCH_SIZE \
        --warmup $WARMUP \
        --iterations $ITERATIONS \
        --fp16 \
        --output "$OUTPUT_DIR/speed_pytorch_csa_fp16.json" || true
    
    # TensorRT benchmarks
    if [ -f "$ENGINE_DIR/clip_vanilla_${IMAGE_SIZE}_fp16.engine" ]; then
        log "Benchmarking TensorRT (vanilla mode)..."
        python3 benchmark_speed.py \
            --trt-engine "$ENGINE_DIR/clip_vanilla_${IMAGE_SIZE}_fp16.engine" \
            --mode vanilla \
            --image-size $IMAGE_SIZE $IMAGE_SIZE \
            --batch-size $BATCH_SIZE \
            --warmup $WARMUP \
            --iterations $ITERATIONS \
            --output "$OUTPUT_DIR/speed_trt_vanilla_fp16.json" || true
    fi
    
    if [ -f "$ENGINE_DIR/declip_csa_${IMAGE_SIZE}_fp16.engine" ]; then
        log "Benchmarking TensorRT (csa mode)..."
        python3 benchmark_speed.py \
            --trt-engine "$ENGINE_DIR/declip_csa_${IMAGE_SIZE}_fp16.engine" \
            --mode csa \
            --image-size $IMAGE_SIZE $IMAGE_SIZE \
            --batch-size $BATCH_SIZE \
            --warmup $WARMUP \
            --iterations $ITERATIONS \
            --output "$OUTPUT_DIR/speed_trt_csa_fp16.json" || true
    fi
}

eval_accuracy() {
    log "Running accuracy evaluations..."
    
    mkdir -p "$OUTPUT_DIR"
    
    # PyTorch evaluations
    log "Evaluating PyTorch (vanilla mode)..."
    python3 eval_panoptic.py \
        --backend pytorch \
        --checkpoint "$EVACLIP_B16" \
        --model-name EVA02-CLIP-B-16 \
        --mode vanilla \
        --coco-root "$COCO_ROOT" \
        --image-size $IMAGE_SIZE \
        --output "$OUTPUT_DIR/accuracy_pytorch_vanilla.json" || true
    
    log "Evaluating PyTorch (csa mode)..."
    python3 eval_panoptic.py \
        --backend pytorch \
        --checkpoint "$EVACLIP_B16" \
        --model-name EVA02-CLIP-B-16 \
        --mode csa \
        --coco-root "$COCO_ROOT" \
        --image-size $IMAGE_SIZE \
        --output "$OUTPUT_DIR/accuracy_pytorch_csa.json" || true
    
    # TensorRT evaluations
    if [ -f "$ENGINE_DIR/clip_vanilla_${IMAGE_SIZE}_fp16.engine" ]; then
        log "Evaluating TensorRT (vanilla mode)..."
        python3 eval_panoptic.py \
            --backend tensorrt \
            --engine "$ENGINE_DIR/clip_vanilla_${IMAGE_SIZE}_fp16.engine" \
            --mode vanilla \
            --coco-root "$COCO_ROOT" \
            --image-size $IMAGE_SIZE \
            --output "$OUTPUT_DIR/accuracy_trt_vanilla.json" || true
    fi
    
    if [ -f "$ENGINE_DIR/declip_csa_${IMAGE_SIZE}_fp16.engine" ]; then
        log "Evaluating TensorRT (csa mode)..."
        python3 eval_panoptic.py \
            --backend tensorrt \
            --engine "$ENGINE_DIR/declip_csa_${IMAGE_SIZE}_fp16.engine" \
            --mode csa \
            --coco-root "$COCO_ROOT" \
            --image-size $IMAGE_SIZE \
            --output "$OUTPUT_DIR/accuracy_trt_csa.json" || true
    fi
}

summarize_results() {
    log "Summarizing results..."
    
    python3 << 'EOF'
import json
import os
from pathlib import Path

results_dir = Path("results")
if not results_dir.exists():
    print("No results found")
    exit(0)

print("\n" + "="*80)
print("EXPERIMENT SUMMARY")
print("="*80)

# Speed results
print("\n## Speed Benchmark Results")
print("-"*80)
print(f"{'Model':<30} {'Backend':<12} {'Precision':<10} {'Latency (ms)':<15} {'FPS':<10}")
print("-"*80)

for json_file in sorted(results_dir.glob("speed_*.json")):
    try:
        with open(json_file) as f:
            data = json.load(f)
            if isinstance(data, list):
                data = data[0]
            print(f"{data.get('model_name', 'N/A')[:30]:<30} "
                  f"{data.get('backend', 'N/A'):<12} "
                  f"{data.get('precision', 'N/A'):<10} "
                  f"{data.get('latency_mean_ms', 0):.2f}±{data.get('latency_std_ms', 0):.2f}{'':>4} "
                  f"{data.get('throughput_fps', 0):.1f}")
    except Exception as e:
        print(f"Error reading {json_file}: {e}")

# Accuracy results
print("\n## Accuracy Evaluation Results (COCO Panoptic)")
print("-"*80)
print(f"{'Model':<30} {'Backend':<12} {'Mode':<8} {'Thing mAcc@1':<12} {'Stuff mAcc@1':<12}")
print("-"*80)

for json_file in sorted(results_dir.glob("accuracy_*.json")):
    try:
        with open(json_file) as f:
            data = json.load(f)
            print(f"{data.get('model_name', 'N/A')[:30]:<30} "
                  f"{data.get('backend', 'N/A'):<12} "
                  f"{data.get('mode', 'N/A'):<8} "
                  f"{data.get('rois_thing_macc1', 0)*100:.2f}%{'':>6} "
                  f"{data.get('rois_stuff_macc1', 0)*100:.2f}%")
    except Exception as e:
        print(f"Error reading {json_file}: {e}")

print("="*80)
EOF
}

# ============================================================
# 主流程
# ============================================================
main() {
    log "Starting TRT Benchmark Experiment"
    log "=============================="
    
    # 解析参数
    SKIP_CONVERT=false
    SKIP_SPEED=false
    SKIP_ACCURACY=false
    
    while [[ $# -gt 0 ]]; do
        case $1 in
            --skip-convert)
                SKIP_CONVERT=true
                shift
                ;;
            --skip-speed)
                SKIP_SPEED=true
                shift
                ;;
            --skip-accuracy)
                SKIP_ACCURACY=true
                shift
                ;;
            --convert-only)
                SKIP_SPEED=true
                SKIP_ACCURACY=true
                shift
                ;;
            --speed-only)
                SKIP_CONVERT=true
                SKIP_ACCURACY=true
                shift
                ;;
            --accuracy-only)
                SKIP_CONVERT=true
                SKIP_SPEED=true
                shift
                ;;
            --help)
                echo "Usage: $0 [options]"
                echo ""
                echo "Options:"
                echo "  --skip-convert    Skip model conversion"
                echo "  --skip-speed      Skip speed benchmark"
                echo "  --skip-accuracy   Skip accuracy evaluation"
                echo "  --convert-only    Only run model conversion"
                echo "  --speed-only      Only run speed benchmark"
                echo "  --accuracy-only   Only run accuracy evaluation"
                echo "  --help            Show this help"
                exit 0
                ;;
            *)
                echo "Unknown option: $1"
                exit 1
                ;;
        esac
    done
    
    # 检查依赖
    check_dependencies
    
    # 执行步骤
    if [ "$SKIP_CONVERT" = false ]; then
        convert_models
    fi
    
    if [ "$SKIP_SPEED" = false ]; then
        benchmark_speed
    fi
    
    if [ "$SKIP_ACCURACY" = false ]; then
        eval_accuracy
    fi
    
    # 汇总结果
    summarize_results
    
    log "Experiment completed!"
    log "Results saved to: $OUTPUT_DIR/"
}

# 运行
main "$@"