Lecture 7_CNN.md

Basic concepts

CNNs architecture

• CONV

  • filter size F: 3 by 3, 5 by 5
  • number of filters K: power of 2, e.g., 32, 64, 128
  • stride S: 1, 2
  • zero padding P: 0, 1, 2

• POOLING

  • filter size F: 2 by 2
  • stride S: 2
  • DOES not change the number of depth
  • Not common to use zero-padding

• FC

CNNs properties

• Local connectivity: the depth of filter is equals to the depth of the input.
• Parameter sharing.

Example

Please refer to tf_5_CNN_text_classification.ipynb for more details about CNNs in text classification.