DeCLIP-TPAMI / code /detection_trt /benchmark_speed.py
xiaomoguhzz's picture
Upload code/detection_trt/benchmark_speed.py with huggingface_hub
4468391 verified
#!/usr/bin/env python3
"""
速度测试脚本 - 对比 PyTorch vs TensorRT 推理性能
测试 DeCLIP (csa模式) 和 CLIP (vanilla模式) 的延迟和吞吐量
Usage:
python benchmark_speed.py --pytorch-model <path> --trt-engine <path>
python benchmark_speed.py --all # 运行所有配置的测试
"""
import os
import sys
import argparse
import time
import json
from pathlib import Path
from dataclasses import dataclass, asdict
from typing import Optional, List, Dict, Any
from datetime import datetime
import torch
import torch.cuda
import numpy as np
# 添加项目根目录到路径
SCRIPT_DIR = Path(__file__).parent
DECLIP_ROOT = SCRIPT_DIR.parent
sys.path.insert(0, str(DECLIP_ROOT))
sys.path.insert(0, str(DECLIP_ROOT / 'src'))
@dataclass
class BenchmarkResult:
"""Benchmark 结果"""
model_name: str
mode: str # vanilla or csa
backend: str # pytorch, onnx, tensorrt
precision: str # fp32, fp16, int8
image_size: tuple
batch_size: int
warmup_rounds: int
test_rounds: int
latency_mean_ms: float
latency_std_ms: float
latency_min_ms: float
latency_max_ms: float
throughput_fps: float
memory_mb: float
timestamp: str
class PyTorchBenchmark:
"""PyTorch 模型 benchmark"""
def __init__(self, model, device='cuda:0', mode='csa'):
self.model = model
self.device = torch.device(device)
self.mode = mode
self.model.to(self.device)
self.model.eval()
@torch.no_grad()
def run(self, input_tensor: torch.Tensor) -> torch.Tensor:
"""运行单次推理"""
if hasattr(self.model, 'visual'):
return self.model.visual.encode_dense(
input_tensor, keep_shape=True, mode=self.mode
)
else:
return self.model.encode_dense(
input_tensor, keep_shape=True, mode=self.mode
)
def benchmark(
self,
image_size: tuple = (560, 560),
batch_size: int = 1,
warmup: int = 10,
iterations: int = 100,
fp16: bool = False
) -> BenchmarkResult:
"""运行 benchmark"""
H, W = image_size
dtype = torch.float16 if fp16 else torch.float32
# 创建输入
input_tensor = torch.randn(
batch_size, 3, H, W,
device=self.device,
dtype=dtype
)
if fp16:
self.model.half()
# Warmup
for _ in range(warmup):
_ = self.run(input_tensor)
torch.cuda.synchronize()
# 测试
latencies = []
torch.cuda.reset_peak_memory_stats()
for _ in range(iterations):
torch.cuda.synchronize()
start = time.perf_counter()
_ = self.run(input_tensor)
torch.cuda.synchronize()
end = time.perf_counter()
latencies.append((end - start) * 1000) # ms
memory_mb = torch.cuda.max_memory_allocated() / (1024 * 1024)
latencies = np.array(latencies)
return BenchmarkResult(
model_name=type(self.model).__name__,
mode=self.mode,
backend='pytorch',
precision='fp16' if fp16 else 'fp32',
image_size=image_size,
batch_size=batch_size,
warmup_rounds=warmup,
test_rounds=iterations,
latency_mean_ms=float(latencies.mean()),
latency_std_ms=float(latencies.std()),
latency_min_ms=float(latencies.min()),
latency_max_ms=float(latencies.max()),
throughput_fps=float(batch_size * 1000 / latencies.mean()),
memory_mb=float(memory_mb),
timestamp=datetime.now().isoformat()
)
class TensorRTBenchmark:
"""TensorRT 引擎 benchmark"""
def __init__(self, engine_path: str, device='cuda:0'):
self.engine_path = engine_path
self.device = device
self.engine = None
self.context = None
self.stream = None
self._load_engine()
def _load_engine(self):
"""加载 TRT 引擎"""
try:
import tensorrt as trt
import pycuda.driver as cuda
import pycuda.autoinit
self.cuda = cuda
self.trt = trt
logger = trt.Logger(trt.Logger.WARNING)
runtime = trt.Runtime(logger)
with open(self.engine_path, 'rb') as f:
self.engine = runtime.deserialize_cuda_engine(f.read())
self.context = self.engine.create_execution_context()
self.stream = cuda.Stream()
# 获取输入输出信息
self.input_name = self.engine.get_tensor_name(0)
self.output_name = self.engine.get_tensor_name(1)
except ImportError as e:
raise ImportError(f"TensorRT/PyCUDA not installed: {e}")
def run(self, input_np: np.ndarray) -> np.ndarray:
"""运行单次推理"""
cuda = self.cuda
# 设置输入形状
self.context.set_input_shape(self.input_name, input_np.shape)
# 获取输出形状
output_shape = self.context.get_tensor_shape(self.output_name)
output_np = np.empty(output_shape, dtype=np.float32)
# 分配 GPU 内存
d_input = cuda.mem_alloc(input_np.nbytes)
d_output = cuda.mem_alloc(output_np.nbytes)
# 复制输入
cuda.memcpy_htod_async(d_input, input_np.astype(np.float32), self.stream)
# 设置绑定
self.context.set_tensor_address(self.input_name, int(d_input))
self.context.set_tensor_address(self.output_name, int(d_output))
# 推理
self.context.execute_async_v3(self.stream.handle)
# 复制输出
cuda.memcpy_dtoh_async(output_np, d_output, self.stream)
self.stream.synchronize()
return output_np
def benchmark(
self,
image_size: tuple = (560, 560),
batch_size: int = 1,
warmup: int = 10,
iterations: int = 100
) -> BenchmarkResult:
"""运行 benchmark"""
cuda = self.cuda
H, W = image_size
# 创建输入
input_np = np.random.randn(batch_size, 3, H, W).astype(np.float32)
# 设置输入形状
self.context.set_input_shape(self.input_name, input_np.shape)
output_shape = self.context.get_tensor_shape(self.output_name)
output_np = np.empty(output_shape, dtype=np.float32)
# 预分配 GPU 内存
d_input = cuda.mem_alloc(input_np.nbytes)
d_output = cuda.mem_alloc(output_np.nbytes)
# Warmup
for _ in range(warmup):
cuda.memcpy_htod(d_input, input_np)
self.context.set_tensor_address(self.input_name, int(d_input))
self.context.set_tensor_address(self.output_name, int(d_output))
self.context.execute_async_v3(self.stream.handle)
self.stream.synchronize()
# 测试
latencies = []
for _ in range(iterations):
cuda.memcpy_htod(d_input, input_np)
start = time.perf_counter()
self.context.set_tensor_address(self.input_name, int(d_input))
self.context.set_tensor_address(self.output_name, int(d_output))
self.context.execute_async_v3(self.stream.handle)
self.stream.synchronize()
end = time.perf_counter()
latencies.append((end - start) * 1000) # ms
latencies = np.array(latencies)
# 从引擎路径解析精度
precision = 'fp16' if 'fp16' in self.engine_path else 'fp32'
if 'int8' in self.engine_path:
precision = 'int8'
return BenchmarkResult(
model_name=Path(self.engine_path).stem,
mode='unknown', # 从文件名解析
backend='tensorrt',
precision=precision,
image_size=image_size,
batch_size=batch_size,
warmup_rounds=warmup,
test_rounds=iterations,
latency_mean_ms=float(latencies.mean()),
latency_std_ms=float(latencies.std()),
latency_min_ms=float(latencies.min()),
latency_max_ms=float(latencies.max()),
throughput_fps=float(batch_size * 1000 / latencies.mean()),
memory_mb=0.0, # TRT 内存统计需要额外处理
timestamp=datetime.now().isoformat()
)
class ONNXBenchmark:
"""ONNX Runtime benchmark"""
def __init__(self, onnx_path: str, device='cuda'):
self.onnx_path = onnx_path
self.device = device
import onnxruntime as ort
providers = ['CUDAExecutionProvider'] if device == 'cuda' else ['CPUExecutionProvider']
self.session = ort.InferenceSession(onnx_path, providers=providers)
self.input_name = self.session.get_inputs()[0].name
def run(self, input_np: np.ndarray) -> np.ndarray:
"""运行单次推理"""
return self.session.run(None, {self.input_name: input_np})[0]
def benchmark(
self,
image_size: tuple = (560, 560),
batch_size: int = 1,
warmup: int = 10,
iterations: int = 100
) -> BenchmarkResult:
"""运行 benchmark"""
H, W = image_size
input_np = np.random.randn(batch_size, 3, H, W).astype(np.float32)
# Warmup
for _ in range(warmup):
_ = self.run(input_np)
# 测试
latencies = []
for _ in range(iterations):
start = time.perf_counter()
_ = self.run(input_np)
end = time.perf_counter()
latencies.append((end - start) * 1000)
latencies = np.array(latencies)
return BenchmarkResult(
model_name=Path(self.onnx_path).stem,
mode='unknown',
backend='onnxruntime',
precision='fp32',
image_size=image_size,
batch_size=batch_size,
warmup_rounds=warmup,
test_rounds=iterations,
latency_mean_ms=float(latencies.mean()),
latency_std_ms=float(latencies.std()),
latency_min_ms=float(latencies.min()),
latency_max_ms=float(latencies.max()),
throughput_fps=float(batch_size * 1000 / latencies.mean()),
memory_mb=0.0,
timestamp=datetime.now().isoformat()
)
def print_result(result: BenchmarkResult):
"""打印 benchmark 结果"""
print(f"\n{'='*60}")
print(f"Benchmark Results: {result.model_name}")
print(f"{'='*60}")
print(f"Backend: {result.backend}")
print(f"Mode: {result.mode}")
print(f"Precision: {result.precision}")
print(f"Image Size: {result.image_size}")
print(f"Batch Size: {result.batch_size}")
print(f"Warmup: {result.warmup_rounds} rounds")
print(f"Test: {result.test_rounds} rounds")
print(f"-" * 40)
print(f"Latency Mean: {result.latency_mean_ms:.2f} ms")
print(f"Latency Std: {result.latency_std_ms:.2f} ms")
print(f"Latency Min: {result.latency_min_ms:.2f} ms")
print(f"Latency Max: {result.latency_max_ms:.2f} ms")
print(f"Throughput: {result.throughput_fps:.1f} FPS")
if result.memory_mb > 0:
print(f"GPU Memory: {result.memory_mb:.1f} MB")
print(f"{'='*60}")
def compare_results(results: List[BenchmarkResult]):
"""对比多个结果"""
if len(results) < 2:
return
print(f"\n{'='*60}")
print("Comparison Summary")
print(f"{'='*60}")
# 按 backend 排序
results = sorted(results, key=lambda x: (x.mode, x.backend))
# 打印表格头
headers = ["Model", "Mode", "Backend", "Precision", "Latency (ms)", "FPS", "Speedup"]
row_format = "{:<20} {:<8} {:<12} {:<10} {:>12} {:>8} {:>8}"
print(row_format.format(*headers))
print("-" * 80)
# 基准 (第一个 PyTorch 结果)
baseline = next((r for r in results if r.backend == 'pytorch'), results[0])
for result in results:
speedup = baseline.latency_mean_ms / result.latency_mean_ms
speedup_str = f"{speedup:.2f}x" if speedup != 1.0 else "-"
print(row_format.format(
result.model_name[:20],
result.mode,
result.backend,
result.precision,
f"{result.latency_mean_ms:.2f}",
f"{result.throughput_fps:.1f}",
speedup_str
))
def save_results(results: List[BenchmarkResult], output_path: str):
"""保存结果到 JSON"""
data = [asdict(r) for r in results]
with open(output_path, 'w') as f:
json.dump(data, f, indent=2)
print(f"\nResults saved to: {output_path}")
def parse_args():
parser = argparse.ArgumentParser(description='速度测试')
parser.add_argument('--pytorch-model', type=str,
help='PyTorch 模型检查点路径')
parser.add_argument('--trt-engine', type=str,
help='TensorRT 引擎路径')
parser.add_argument('--onnx-model', type=str,
help='ONNX 模型路径')
parser.add_argument('--model-name', type=str, default='EVA02-CLIP-B-16',
help='模型名称')
parser.add_argument('--mode', type=str, default='csa',
choices=['vanilla', 'csa'],
help='特征模式')
parser.add_argument('--image-size', type=int, nargs=2, default=[560, 560],
help='图像尺寸 [H, W]')
parser.add_argument('--batch-size', type=int, default=1,
help='批处理大小')
parser.add_argument('--warmup', type=int, default=10,
help='预热轮数')
parser.add_argument('--iterations', type=int, default=100,
help='测试轮数')
parser.add_argument('--fp16', action='store_true',
help='使用 FP16 (仅 PyTorch)')
parser.add_argument('--all', action='store_true',
help='运行所有配置的测试')
parser.add_argument('--output', type=str, default='results/benchmark_results.json',
help='输出文件路径')
parser.add_argument('--device', type=str, default='cuda:0',
help='设备')
return parser.parse_args()
def run_all_benchmarks(args):
"""运行所有配置的测试"""
from configs.fvit_tensorrt_fp16 import model_paths, benchmark_config
results = []
# 测试配置
modes = ['vanilla', 'csa']
image_sizes = benchmark_config['input_sizes']
batch_sizes = benchmark_config['batch_sizes']
# 检查引擎目录
engine_dir = SCRIPT_DIR / 'engines'
for mode in modes:
model_type = 'clip' if mode == 'vanilla' else 'declip'
# 查找对应的引擎
engine_pattern = f"{model_type}_{mode}_*_fp16.engine"
engines = list(engine_dir.glob(engine_pattern))
if not engines:
print(f"Warning: No TRT engine found for {model_type} {mode}")
continue
engine_path = engines[0]
for image_size in image_sizes:
for batch_size in batch_sizes:
# TensorRT benchmark
try:
benchmark = TensorRTBenchmark(str(engine_path), args.device)
result = benchmark.benchmark(
image_size=image_size,
batch_size=batch_size,
warmup=args.warmup,
iterations=args.iterations
)
result.mode = mode
results.append(result)
print_result(result)
except Exception as e:
print(f"Error benchmarking TRT {mode}: {e}")
return results
def main():
args = parse_args()
results = []
print(f"\n{'='*60}")
print("DeCLIP/CLIP Speed Benchmark")
print(f"{'='*60}")
print(f"Device: {args.device}")
print(f"Image Size: {args.image_size}")
print(f"Batch Size: {args.batch_size}")
print(f"Warmup: {args.warmup}")
print(f"Iterations: {args.iterations}")
if args.all:
results = run_all_benchmarks(args)
else:
image_size = tuple(args.image_size)
# PyTorch benchmark
if args.pytorch_model:
from open_clip import create_model
print(f"\nLoading PyTorch model: {args.pytorch_model}")
model = create_model(
args.model_name,
pretrained='eva',
device=args.device,
precision='fp32',
output_dict=True,
cache_dir=args.pytorch_model
)
benchmark = PyTorchBenchmark(model, args.device, args.mode)
result = benchmark.benchmark(
image_size=image_size,
batch_size=args.batch_size,
warmup=args.warmup,
iterations=args.iterations,
fp16=args.fp16
)
results.append(result)
print_result(result)
# ONNX benchmark
if args.onnx_model:
print(f"\nLoading ONNX model: {args.onnx_model}")
benchmark = ONNXBenchmark(args.onnx_model, 'cuda')
result = benchmark.benchmark(
image_size=image_size,
batch_size=args.batch_size,
warmup=args.warmup,
iterations=args.iterations
)
result.mode = args.mode
results.append(result)
print_result(result)
# TensorRT benchmark
if args.trt_engine:
print(f"\nLoading TensorRT engine: {args.trt_engine}")
benchmark = TensorRTBenchmark(args.trt_engine, args.device)
result = benchmark.benchmark(
image_size=image_size,
batch_size=args.batch_size,
warmup=args.warmup,
iterations=args.iterations
)
result.mode = args.mode
results.append(result)
print_result(result)
# 对比结果
if len(results) > 1:
compare_results(results)
# 保存结果
if results:
output_path = Path(args.output)
output_path.parent.mkdir(parents=True, exist_ok=True)
save_results(results, str(output_path))
if __name__ == '__main__':
main()