Hyperparameters: Difference between revisions
| (One intermediate revision by the same user not shown) | |||
| Line 34: | Line 34: | ||
==CNNs== | ==CNNs== | ||
===Kernel Size=== | ===Kernel Size=== | ||
Historically, people used larger kernels like 27x27. | |||
Since VGGnet, people now use smaller kernels like 3x3 or 4x4 along with more layers. | |||
===Stride=== | ===Stride=== | ||
Strided convolution is good for decreasing the resolution and increasing the receptive field. | |||
Typically, you can choose between using a strided convolution or using pooling to drop the resolution of images or increase the receptive field. | |||
The alternative to strided convolutions is to use dilated convolutions, like in Atrous Spatial Pyramid Pooling (ASPP). | |||
==Activation functions== | ==Activation functions== | ||