search

DrSleep / light-weight-refinenet

Light-Weight RefineNet for Real-Time Semantic Segmentation
Other
739 stars 164 forks source link

Confusion about task_idx #42

Closed MihaiDogariu closed 5 years ago

MihaiDogariu commented 5 years ago

Hi! Great work with this repo!

I have a question regarding how you structured the training process. Why do you have 3-dimensional values in the config file?

I've seen that the task_idx value will iterate until it reaches the value of num_stages (which is assigned the depth of the num_classes variable, i.e. 3) and I am not sure of what it means or why it is used.

Thanks!

DrSleep commented 5 years ago

Essentially, we divide our training into 3 stages as written here: https://github.com/DrSleep/light-weight-refinenet#training-scripts

The idea is to reduce the learning rate at the end of each stage; optionally, you can choose not only the learning rate but also other variables for each stage, hence why almost all parameters are of length 3