【训练时间（Training runtime）、参数量(Params)、计算量(FLOPs)】

本文链接：https://blog.csdn.net/weixin_47244593/article/details/138809089

1. 运行时间

    start_time = time.time()
    trainer.train()
    end_time = time.time()
    
    runtime = end_time - start_time
    # 将秒转换为小时
    hours = time.strftime("%H:%M:%S", time.gmtime(runtime))

    # 打印结果
    print("Training runtime:", hours)   # Training runtime: 01:07:17
    print('Training runtime: %.2f h' % (runtime/3600))     # Training runtime: 1.12 h
    print("Training runtime: ", str(datetime.timedelta(seconds=runtime)))  # Training runtime: 1:07:17.291430

2. 计算量(FLOPs)、参数量(Params)

import torch
from thop import profile
from models.HRA_Net.HRA_Net import HRANet


MODEL = HRANet(n_channels=1, n_classes=2)    # 定义好的网络模型
input = torch.randn(1, 1, 256, 256)   # B C W H
flops, params = profile(MODEL, (input,))

# print('flops: ', flops, 'params: ', params)   # flops:  96817342592.0 params:  48597090.0
# print('flops: %.2f M, params: %.2f M' % (flops / 1000000.0, params / 1000000.0))   # flops: 96817.34 M, params: 48.60 M
print('flops: %.2f G, params: %.2f M' % ((flops / 1e6)/1024, params / 1e6))   # flops: 94.55 G, params: 48.60 M

注：

输入input的批量(batch size)，批量的大小不会影响参数量，但是会影响计算量，计算量是batch_size=1的倍数；图像的尺寸也会影响计算量；
- 输入数据：(1, 1, 256, 256)，一张256*256的灰度图像，结果是：flops: 94.55 G, params: 48.60 M
- 输入数据：(2, 1, 256, 256)，两张256*256的灰度图像，结果是：flops: 189.10 G, params:48.60 M
- 输入数据：(1, 1, 512, 512)，一张512*512的灰度图像，结果是：flops: 388.25 G, params: 48.60 M
- 输入数据：(1, 3, 512, 512)，一张512*512的RGB图像，结果是：flops: 388.55 G, params: 48.60 M
profile(MODEL, (input,))的 (input,)中必须加上逗号，否则会报错。