Skip to content

ziedbouf/waternn

 
 

Repository files navigation

Fully-linear DenseNet

Implementation of a Fully-linear DenseNet for pipe busrt location of a water distribution network.

Some cmd arguments:

  • --arch choose a kind of architecture to use
  • --out the file name of predict result
  • --lr learning rate
  • --epoch training epochs
  • --kind the name of the log file recording information during running
  • --lr_decay learning rate decay
  • --data_dir the path of dataset
  • --save_path the path of saving model (need to change code)
  • --num_init_features an argument of arch waterdsnetf_self_define
  • --growth_rate an argument of arch waterdsnetf_self_define

example cmd 1 (default Anytown dataset) :

python train.py main --arch waterdsnetf --lr 0.6 --epoch 120 --kind Anytown_P10C10_B0105_4M --data_dir ~/water/Anytown_P10C10_B0105_4M --save_path './water/modelparams' --out Anytown_P10C10_B0105_4M

example cmd 2 (dataset Anytown with Duration changed) :

python train.py main --arch waterdsnetf_self_define --num_init_features 320 --growth_rate 32 --lr 0.6 --epoch 120 --kind Anytown_P10C10_B1030_Duration5 --data_dir ~/water/Anytown_P10C10_B1030_Duration5 --save_path './water/modelparams' --out Anytown_P10C10_B1030_Duration5

example cmd 3 (Mudu dataset) :

CUDA_VISIBLE_DEVICES=0 python train.py main --arch waterdsnetf_in4_out58 --lr 0.6 --epoch 120 --kind Mudu_P10C10_B1030_4M --data_dir ~/water/Mudu_P10C10_B1030_4M --save_path './water/modelparams' --out Mudu_P10C10_B1030_4M

Tricks during training

  • diminish learning rate with traning epochs increasing
You can directly changde the codes to set the periods of diminishing.

other super-parameters:

  • batch_size
Default batch_size is 128, you can change it in the train.py
  • weight_decay
Default weight_decay is 0.00005, you can change it in config.py

About

predict broken waterpipe

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Jupyter Notebook 74.3%
  • Python 25.7%