解读tensorflow benchmark工具的结果

Interpreting results of tensorflow benchmark tool

Tensorflow 的基准测试工具很少:

对于.pb model and for .tflite model

我对 .pb 基准测试工具的参数有几个问题:

  1. num_threads与单线程实验的并行运行数有关,还是与tensorflow使用的内部线程有关?
  2. 在为桌面构建工具时是否可以使用 GPU,即不适用于移动设备?如果是这样,如何确保不使用GPU?

还有一些关于结果解释的问题:

  1. 结果输出中的count是什么? Timings (microseconds): count=--max_num_runs 参数有什么关系?

示例:

Run --num_threads=-1 --max_num_runs=1000:
    2019-03-20 14:30:33.253584: I tensorflow/core/util/stat_summarizer.cc:85] Timings (microseconds): count=1000 first=3608 curr=3873 min=3566 max=8009 avg=3766.49 std=202
    2019-03-20 14:30:33.253584: I tensorflow/core/util/stat_summarizer.cc:85] Memory (bytes): count=1000 curr=3301344(all same)
    2019-03-20 14:30:33.253591: I tensorflow/core/util/stat_summarizer.cc:85] 207 nodes observed
    2019-03-20 14:30:33.253597: I tensorflow/core/util/stat_summarizer.cc:85]
    2019-03-20 14:30:33.378352: I tensorflow/tools/benchmark/benchmark_model.cc:636] FLOPs estimate: 116.65M
    2019-03-20 14:30:33.378390: I tensorflow/tools/benchmark/benchmark_model.cc:638] FLOPs/second: 46.30B

Run --num_threads=1 --max_num_runs=1000:
    2019-03-20 14:32:25.591915: I tensorflow/core/util/stat_summarizer.cc:85] Timings (microseconds): count=1000 first=7502 curr=7543 min=7495 max=7716 avg=7607.22 std=34
    2019-03-20 14:32:25.591934: I tensorflow/core/util/stat_summarizer.cc:85] Memory (bytes): count=1000 curr=3301344(all same)
    2019-03-20 14:32:25.591952: I tensorflow/core/util/stat_summarizer.cc:85] 207 nodes observed
    2019-03-20 14:32:25.591970: I tensorflow/core/util/stat_summarizer.cc:85]
    2019-03-20 14:32:25.805970: I tensorflow/tools/benchmark/benchmark_model.cc:636] FLOPs estimate: 116.65M
    2019-03-20 14:32:25.806007: I tensorflow/tools/benchmark/benchmark_model.cc:638] FLOPs/second: 15.46B

Run --num_threads=-1 --max_num_runs=10000:
    2019-03-20 14:38:48.045824: I tensorflow/core/util/stat_summarizer.cc:85] Timings (microseconds): count=3570 first=3961 curr=3899 min=3558 max=6997 avg=3841.2 std=175
    2019-03-20 14:38:48.045829: I tensorflow/core/util/stat_summarizer.cc:85] Memory (bytes): count=3570 curr=3301344(all same)
    2019-03-20 14:38:48.045833: I tensorflow/core/util/stat_summarizer.cc:85] 207 nodes observed
    2019-03-20 14:38:48.045837: I tensorflow/core/util/stat_summarizer.cc:85]
    2019-03-20 14:38:48.169368: I tensorflow/tools/benchmark/benchmark_model.cc:636] FLOPs estimate: 116.65M
    2019-03-20 14:38:48.169412: I tensorflow/tools/benchmark/benchmark_model.cc:638] FLOPs/second: 48.66B

Run --num_threads=1 --max_num_runs=10000:
    2019-03-20 14:35:50.826722: I tensorflow/core/util/stat_summarizer.cc:85] Timings (microseconds): count=1254 first=7496 curr=7518 min=7475 max=7838 avg=7577.23 std=50
    2019-03-20 14:35:50.826735: I tensorflow/core/util/stat_summarizer.cc:85] Memory (bytes): count=1254 curr=3301344(all same)
    2019-03-20 14:35:50.826746: I tensorflow/core/util/stat_summarizer.cc:85] 207 nodes observed
    2019-03-20 14:35:50.826757: I tensorflow/core/util/stat_summarizer.cc:85]
    2019-03-20 14:35:51.053143: I tensorflow/tools/benchmark/benchmark_model.cc:636] FLOPs estimate: 116.65M
    2019-03-20 14:35:51.053180: I tensorflow/tools/benchmark/benchmark_model.cc:638] FLOPs/second: 15.55B

即当使用 --max_num_runs=10000 计数时 count=3570count=1254 是什么意思?

对于 .tflite 基准工具:

--num_threads=1 --num_runs=10000
    Initialized session in 0.682ms
    Running benchmark for at least 1 iterations and at least 0.5 seconds
    count=54 first=23463 curr=8019 min=7911 max=23463 avg=9268.5 std=2995
    Running benchmark for at least 1000 iterations and at least 1 seconds
    count=1000 first=8022 curr=6703 min=6613 max=10333 avg=6766.23 std=337
    Average inference timings in us: Warmup: 9268.5, Init: 682, no stats: 6766.23

no stats: 6766.23 是什么意思?

深入研究代码后,我发现了以下内容(所有时间均以微秒为单位):

  • count:实际运行s
  • 个数
  • first: 第一次迭代的时间
  • curr: 上次迭代的时间
  • min: 迭代花费的最短时间
  • max: 迭代花费的最长时间
  • avg:一次迭代的平均时间
  • std:所有 运行s
  • 时间的标准偏差
  • Warmup:热身运行平均
  • Init:启动时间(应始终与 Initialized session in 相同)
  • no stats: 是非常不好命名的平均运行时间(匹配上一行的avg=
  • num_threads:这用于设置intra_op_parallelism_threadsinter_op_parallelism_threads(更多信息here

相关文件(链接到正确的行)是:

我不太确定使用 GPU 还是不使用 GPU。如果您使用 freeze_graph 导出 .pb 文件,那么它将存储图中每个节点的设备。您可以在导出前使用设备放置来执行此操作。如果您需要更改它,可以尝试设置环境变量CUDA_VISIBLE_DEVICES=""以确保未使用GPU。