In normal pooling operations we have to mention pool size for pooling operation, like for 2D pooling operations we mention (2,2); however, in global pooling operation it is not required. So is it the same size as input? I am working on Keras. Here one author mentions that the pooling size is the same as input size or input size-filter size+1. Which one is correct?
Asked
Active
Viewed 846 times
1
2 Answers
3
In case of 1D pooling, as mentioned in Keras docs, it takes as input an array of shape (batch_size, steps, features) and its output shape is (batch_size, features). So the pool size is equal to steps.
In case of 2D pooling, as mentioned in Keras docs, it takes as input an array of shape (batch_size, rows, cols, channels) and its output shape is (batch_size, channels). So the pool size is equal to (rows, cols).
In both cases, the pool size is consistent with the intuition behind it: taking maximum value over the whole data axes (i.e. global).
today
- 32,602
- 8
- 95
- 115
1
If the input shape is (None, rows, cols, filters), then global pooling uses pool_size of (rows, cols).
Kota Mori
- 6,510
- 1
- 21
- 25