# TIPC Linux端Benchmark测试文档 该文档为Benchmark测试说明,Benchmark预测功能测试的主程序为`benchmark_train.sh`,用于验证监控模型训练的性能。 # 1. 测试流程 ## 1.1 准备数据和环境安装 运行`test_tipc/prepare.sh`,完成训练数据准备和安装环境流程。 ```shell # 运行格式:bash test_tipc/prepare.sh train_benchmark.txt mode bash test_tipc/prepare.sh test_tipc/configs/det_mv3_db_v2.0/train_benchmark.txt benchmark_train ``` ## 1.2 功能测试 执行`test_tipc/benchmark_train.sh`,完成模型训练和日志解析 ```shell # 运行格式:bash test_tipc/benchmark_train.sh train_benchmark.txt mode params bash test_tipc/benchmark_train.sh test_tipc/configs/det_mv3_db_v2.0/train_benchmark.txt benchmark_train dynamic_bs8_null_SingleP_DP_N1C1 ``` params为test_tipc/benchmark_train.sh传入的参数,包含:模型类型、batchsize、fp精度、进程类型、运行模式以及分布式等信息。 `${modeltype}_${batch_size}_${fp_item}_${run_process_type}_${run_mode}_${device_num}` ## 2. 日志输出 运行后将输出模型的训练日志和日志解析日志,使用 `test_tipc/configs/det_mv3_db_v2.0/train_benchmark.txt` 参数文件的训练日志解析结果是: ``` {"model_branch": "dygaph", "model_commit": "7c39a1996b19087737c05d883fd346d2f39dbcc0", "model_name": "det_mv3_db_v2.0_bs8_fp32_SingleP_DP", "batch_size": 8, "fp_item": "fp32", "run_process_type": "SingleP", "run_mode": "DP", "convergence_value": "5.413110", "convergence_key": "loss:", "ips": 19.333, "speed_unit": "images/s", "device_num": "N1C1", "model_run_time": "0", "frame_commit": "8cc09552473b842c651ead3b9848d41827a3dbab", "frame_version": "0.0.0"} ``` 训练日志和日志解析结果保存在benchmark_log目录下,文件组织格式如下: ``` benchmark_log/ ├── index │   ├── PaddleOCR_det_mv3_db_v2.0_bs8_fp32_SingleP_DP_N1C1_speed │   └── PaddleOCR_det_mv3_db_v2.0_bs8_fp32_SingleP_DP_N1C4_speed ├── profiling_log │   └── PaddleOCR_det_mv3_db_v2.0_bs8_fp32_SingleP_DP_N1C1_profiling └── train_log ├── PaddleOCR_det_mv3_db_v2.0_bs8_fp32_SingleP_DP_N1C1_log └── PaddleOCR_det_mv3_db_v2.0_bs8_fp32_SingleP_DP_N1C4_log ```