rebuild

model/benchmark.py

@@ -0,0 +1,142 @@
+import torch
+import torch.nn as nn
+import time
+import numpy as np
+from resnet_attention import resnet18_cbam, resnet34_cbam, resnet50_cbam
+
+# 设置随机种子以确保结果可复现
+torch.manual_seed(42)
+np.random.seed(42)
+
+# 设备配置
+device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
+print(f"测试设备: {device}")
+
+# 测试参数
+batch_sizes = [1, 4, 8, 16]
+image_sizes = [224, 384, 512]
+num_runs = 100  # 每个配置运行的次数
+warmup_runs = 20  # 预热运行次数，排除启动开销
+
+# 模型配置
+model_configs = {
+    "resnet18": {
+        "base_model": lambda: resnet18_cbam(use_cbam=False),
+        "attention_model": lambda: resnet18_cbam(use_cbam=True)
+    },
+    "resnet34": {
+        "base_model": lambda: resnet34_cbam(use_cbam=False),
+        "attention_model": lambda: resnet34_cbam(use_cbam=True)
+    },
+    "resnet50": {
+        "base_model": lambda: resnet50_cbam(use_cbam=False),
+        "attention_model": lambda: resnet50_cbam(use_cbam=True)
+    }
+}
+
+
+# 基准测试函数
+def benchmark_model(model, input_size, batch_size, num_runs, warmup_runs):
+    """
+    测试模型的推理性能
+
+    参数:
+    - model: 待测试的模型
+    - input_size: 输入图像尺寸
+    - batch_size: 批次大小
+    - num_runs: 测试运行次数
+    - warmup_runs: 预热运行次数
+
+    返回:
+    - 平均推理时间(毫秒)
+    - 吞吐量(样本/秒)
+    """
+    # 设置为评估模式
+    model.eval()
+    model.to(device)
+
+    # 创建随机输入
+    input_tensor = torch.randn(batch_size, 3, input_size, input_size, device=device)
+
+    # 预热
+    with torch.no_grad():
+        for _ in range(warmup_runs):
+            _ = model(input_tensor)
+            if device.type == 'cuda':
+                torch.cuda.synchronize()  # 同步GPU操作
+
+    # 测量推理时间
+    start_time = time.time()
+    with torch.no_grad():
+        for _ in range(num_runs):
+            _ = model(input_tensor)
+            if device.type == 'cuda':
+                torch.cuda.synchronize()  # 同步GPU操作
+    end_time = time.time()
+
+    # 计算指标
+    total_time = end_time - start_time
+    avg_time_per_batch = total_time / num_runs * 1000  # 毫秒
+    throughput = batch_size * num_runs / total_time  # 样本/秒
+
+    return avg_time_per_batch, throughput
+
+
+# 运行测试
+results = {}
+
+for model_name, config in model_configs.items():
+    results[model_name] = {}
+
+    # 创建模型
+    base_model = config["base_model"]()
+    attention_model = config["attention_model"]()
+
+    # 计算参数量
+    base_params = sum(p.numel() for p in base_model.parameters() if p.requires_grad)
+    attention_params = sum(p.numel() for p in attention_model.parameters() if p.requires_grad)
+    param_increase = (attention_params - base_params) / base_params * 100
+
+    print(f"\n测试模型: {model_name}")
+    print(f"  基础参数量: {base_params / 1e6:.2f}M")
+    print(f"  带注意力参数量: {attention_params / 1e6:.2f}M")
+    print(f"  参数量增加: {param_increase:.2f}%")
+
+    for batch_size in batch_sizes:
+        for image_size in image_sizes:
+            key = f"batch_{batch_size}_size_{image_size}"
+            results[model_name][key] = {}
+
+            # 测试基础模型
+            base_time, base_throughput = benchmark_model(
+                base_model, image_size, batch_size, num_runs, warmup_runs
+            )
+
+            # 测试注意力模型
+            attention_time, attention_throughput = benchmark_model(
+                attention_model, image_size, batch_size, num_runs, warmup_runs
+            )
+
+            # 计算增加的百分比
+            time_increase = (attention_time - base_time) / base_time * 100
+            throughput_decrease = (base_throughput - attention_throughput) / base_throughput * 100
+
+            results[model_name][key]["base_time"] = base_time
+            results[model_name][key]["attention_time"] = attention_time
+            results[model_name][key]["time_increase"] = time_increase
+            results[model_name][key]["base_throughput"] = base_throughput
+            results[model_name][key]["attention_throughput"] = attention_throughput
+            results[model_name][key]["throughput_decrease"] = throughput_decrease
+
+            print(f"  配置: 批次大小={batch_size}, 图像尺寸={image_size}x{image_size}")
+            print(f"    基础模型: 平均时间={base_time:.2f}ms, 吞吐量={base_throughput:.2f}样本/秒")
+            print(f"    注意力模型: 平均时间={attention_time:.2f}ms, 吞吐量={attention_throughput:.2f}样本/秒")
+            print(f"    时间增加: {time_increase:.2f}%, 吞吐量下降: {throughput_decrease:.2f}%")
+
+# 保存结果
+import json
+
+with open('benchmark_results.json', 'w') as f:
+    json.dump(results, f, indent=2)
+
+print("\n测试完成，结果已保存到 benchmark_results.json")