From 6d626fa7c7a775da2e4887184e3ac1a7ef618349 Mon Sep 17 00:00:00 2001
From: wangna11BD <79366697+wangna11BD@users.noreply.github.com>
Date: Wed, 29 Sep 2021 18:55:14 +0800
Subject: [PATCH] add BasicVSR Benchmark (#440)

* add Benchmark

* add BasicVSR Benchmark
---
 benchmark/README.md      | 37 +++++++++++++++++++++++++++++++++++++
 benchmark/benchmark.yaml | 10 +++++++++-
 benchmark/run_all.sh     | 16 +++++++++++++---
 3 files changed, 59 insertions(+), 4 deletions(-)
 create mode 100644 benchmark/README.md

diff --git a/benchmark/README.md b/benchmark/README.md
new file mode 100644
index 0000000..f8eea34
--- /dev/null
+++ b/benchmark/README.md
@@ -0,0 +1,37 @@
+# PaddGAN模型性能复现
+## 目录
+
+```
+├── README.md               # 说明文档
+├── benchmark.yaml          # 配置文件，设置测试模型及模型参数
+├── run_all.sh              # 执行入口，测试并获取所有生成对抗模型的训练性能
+└── run_benchmark.sh        # 执行实体，测试单个分割模型的训练性能  
+```
+
+## 环境介绍
+### 物理机环境
+- 单机（单卡、8卡）
+  - 系统：CentOS release 7.5 (Final)
+  - GPU：Tesla V100-SXM2-32GB * 8
+  - CPU：Intel(R) Xeon(R) Gold 6271C CPU @ 2.60GHz * 80
+  - CUDA、cudnn Version: cuda10.2-cudnn7
+
+#### 备注
+BasicVSR模型因竞品torch模型只能测4卡，故这里也测4卡。
+
+### Docker 镜像
+
+- **镜像版本**: `registry.baidubce.com/paddlepaddle/paddle:2.1.2-gpu-cuda10.2-cudnn7`
+- **paddle 版本**: `2.1.2`
+- **CUDA 版本**: `10.2`
+- **cuDnn 版本**: `7`
+
+## 测试步骤
+
+```bash
+bash benchmark/run_all.sh  
+```
+
+## 输出
+
+执行完成后，在PaddleGAN目录会产出模型训练性能数据的文件，比如`esrgan_mp_bs32_fp32_8`等文件。
diff --git a/benchmark/benchmark.yaml b/benchmark/benchmark.yaml
index e9389d0..4a5c9e6 100644
--- a/benchmark/benchmark.yaml
+++ b/benchmark/benchmark.yaml
@@ -23,9 +23,17 @@ esrgan:
   log_interval: 10
 
 edvr:
-  dataset_web: https://paddlegan.bj.bcebos.com/datasets/REDS.tar
+  dataset: data/REDS
   config: configs/edvr_m_wo_tsa.yaml
   fp_item: fp32
   bs_item: 4 64
   total_iters: 300
   log_interval: 10
+
+basicvsr:
+  dataset: data/REDS
+  config: configs/basicvsr_reds.yaml
+  fp_item: fp32
+  bs_item: 2 4
+  total_iters: 300
+  log_interval: 10
diff --git a/benchmark/run_all.sh b/benchmark/run_all.sh
index 35597f2..90e5284 100755
--- a/benchmark/run_all.sh
+++ b/benchmark/run_all.sh
@@ -32,9 +32,14 @@ for model_mode in ${model_mode_list[@]}; do
       eval total_iters='$'"${model_mode}_total_iters"
       eval epochs='$'"${model_mode}_epochs"
       eval dataset_web='$'"${model_mode}_dataset_web"
+      eval dataset='$'"${model_mode}_dataset"
       eval log_interval='$'"${model_mode}_log_interval"
-      wget ${dataset_web} -O data/${model_mode}.tar
-      tar -vxf data/${model_mode}.tar -C data/
+      if [ -n "$dataset" ]; then
+          cp -r ${dataset} data/
+      else
+          wget ${dataset_web} -O data/${model_mode}.tar
+          tar -vxf data/${model_mode}.tar -C data/
+      fi
       if [ -n "$total_iters" ]; then
           mode="total_iters"
           max_iter=$total_iters
@@ -52,7 +57,12 @@ for model_mode in ${model_mode_list[@]}; do
             sleep 60
             echo "index is speed, 8gpus, run_mode is multi_process, begin, ${model_name}"
             run_mode=mp
-            CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 bash benchmark/run_benchmark.sh ${run_mode} ${bs_item} ${fp_item} ${mode} ${max_iter} ${model_mode} ${config} ${log_interval}
+            basicvsr_name=basicvsr
+            if [ ${model_mode} = ${basicvsr_name} ]; then
+                CUDA_VISIBLE_DEVICES=0,1,2,3 bash benchmark/run_benchmark.sh ${run_mode} ${bs_item} ${fp_item} ${mode} ${max_iter} ${model_mode} ${config} ${log_interval}
+            else
+                CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 bash benchmark/run_benchmark.sh ${run_mode} ${bs_item} ${fp_item} ${mode} ${max_iter} ${model_mode} ${config} ${log_interval}
+            fi
             sleep 60
             done
       done
-- 
GitLab