keras-image-models icon indicating copy to clipboard operation
keras-image-models copied to clipboard

Published 20 hours ago • verified publisher icon james77777778

feature extraction stages

Open jawi289o opened this issue 11 months ago • 1 comments
trafficstars

I have a question about understanding the structure and functionality of backbone stages. I want to know how to identify which layers belong to each stage. For example, MobileNetV3 has 5 feature extraction stages (or downsampling stages), and I’d like to understand what layers or operations are included in each. Stage 1: What kind of layers (e.g., convolutions, activations) are included? Stage 2, Stage 3, etc.: What layers or operations are present here? I want to better understand the breakdown of these stages to analyze how features are processed and extracted.

jawi289o avatar Dec 20 '24 15:12 jawi289o

Hi @jawi289o Thanks for asking! You can check the variable features in all models to identify which tensor belongs to the extraction stage.

For example:

  • Initialize features: https://github.com/james77777778/keras-image-models/blob/main/kimm/_src/models/mobilenet_v3.py#L141-L142
  • "STEM": https://github.com/james77777778/keras-image-models/blob/main/kimm/_src/models/mobilenet_v3.py#L158
  • "BLOCK": https://github.com/james77777778/keras-image-models/blob/main/kimm/_src/models/mobilenet_v3.py#L229

james77777778 avatar Dec 24 '24 06:12 james77777778