Param in the paper

The following figure comes from ECA-Net (CVPR2020, paper)

GitHub link ：https://github.com/BangguWu/ECANet

As soon as I saw Params Of M At the end, I thought of the storage capacity unit , It's not true , This M Refers to ten thousand , It means parameter quantity , As can be seen from the figure above Resnet50 The parameters of the model are 24.37 ten thousand individual ,ResNet101 The parameter quantity of the model is 42.49 ten thousand individual . To calculate the storage capacity, you need to convert the following formula ：

1 G = 1000 MB
1 MB = 1000 KB
1 KB= 1000 Byte
1 Byte = 8 bit

The model is generally saved with 32bit Of Double precision floating point , The storage capacity corresponding to a parameter is 4Byte, That is, about equal to 0.004KB, That is to say 0.000004MB.

Key formula ：

Memory of the model = Parameters of the model x 0.000004MB（ This is just an estimate , There may be some deviation in the specific model memory ）

There are two reasons for my analysis , One is the node name of the model , for example resnet_block、conv 、bn And so on. , Another is that the parameter storage format of the volume layer is different from that of other layers , It means The parameters of the whole model are not uniform double precision floating-point type

Try to calculate the following figure LeNet Memory of the model ：

Memory of the model （245.95KB） = Parameters of the model （61495） x 0.004KB

The picture below is 240.21KB, It's a little biased

Try to calculate the memory of the model in the following figure ：

Memory of the model （31.915MB） = Parameters of the model （7978856） x 0.000004MB

The picture below is 30.44MB, It's a little biased

It is recommended to read reference articles

Reference article ：CNN The computation of the model 、 Parameters 、 Memory occupation