如何估算模型训练T(FL)OPS efficiency

Naive方法

以Torch Vision ResNet50-v1.5为例。

• Step 1: 获取模型的前向理论需求MACs(Multiply–ACcumulate)
  可使用thop得到模型的前向MACS。使用如下代码可得Torch Vision ResNet50-v1.5的前向MACs为4.112G。

  from torchvision.models import resnet50
  from thop import profile, clever_format
  import torch
  model = resnet50()
  input = torch.randn(1, 3, 224, 224)
  macs, params = profile(model, inputs=(input,))
  print(clever_format([macs, params], "%.3f"))

   

• Step 2: 估算模型在某个实测性能下每秒需求的T(FL)OPS
  估算公式以OpenAI AI and Compute估算公式为基础：

  required_T(FL)OPS = (MACs per forward pass) * (2 (FL)OPs/MAC) * (3 for forward and backward pass) * (number of examples per second)

  再由实测性能数据：

  accelerator	data type	bs	IPS
  V100	FP16	256	1325
  V100	FP32	128	303.1

  
  以V100 FP16训练为例，有：
  MACs per forward pass = 4.112G
  number of examples per second = 1325
  required_(FL)OPS = 4.112G * 2 * 3 * 1325 = 32.69 T
  汇总结果为：

  accelerator	data type	bs	IPS	required T(FL)OPS
  V100	FP16	256	1325	32.69
  V100	FP32	128	303.1	7.478

• Step 3: 估算模型理论峰值算力利用率

  • 理论峰值算力
    

  • 理论峰值算力利用率

    required_T(FL)OPS / peak_T(FL)OPS

    accelerator	data type	bs	IPS	required TF(L)OPS	peak ratio
    V100	FP16	256	1325	32.69	29.2%
    V100	FP32	128	303.1	7.478	53%

References

1. NV Training Performance Benchmark

2. thop

3. OpenAI AI and Compute