What does the number after a machine learning model name mean?
-
09-12-2020 - |
Question
I'm not sure if this is off-topic, but I'm posting here anyway.
So I saw lots of machine learning models have like an ID after their names, for example, resnet101
, resnet152
, densenet201
etc. What exactly do those numbers 101
, 152
and 201
mean? And how it's determined?
Solution
As @Icrmorin said the naming conventions may vary but for the examples you gave, ResNet and DenseNet, the numbers in the name correspond to the number of layers:
DenseNet
Table 1 in the Densenet paper provides an overview:
As you can see, for example, in the DenseNet-121 column this network has $1+6*2+1+12*2+1+24*2+1+16*2 + 1 = 121$ layers and that is where the name is derived from.
ResNet
The ResNet paper provides a similar overview:
Again, you can see how the names are derived: for example ResNet-18 has $1+2*2+2*2+2*2+2*2+1=18$ layers.
Note that in both papers only conv. and dense layers are counted but not the pooling layers.
OTHER TIPS
Sometimes it refers to a version (like windows 10), sometimes it refers to a size of the parameters, like the size of a layer, the number of parameters (like for GPT models), or the size of parameters in memory. They are just names so we can know what we are talking about. I don't think there is a general convention.