Deep Learning Model-Saving Helper

Deep Learning Model-Saving Helper is a simple python framework which allows you to save your deep learning model's middle parameters (for visualization), deep learning model, and config file automatically without explicitly changing your code.

Introduction

For deep learning researcher, a proposed model usually need many experiments with different parameters. The process of recording the parameters need explicit code modification.

For example, changing you code from

tensorboard_log_dir = log/test_1  
batch_size = 32  
learning_rate = 1e-3  

to

tensorboard_log_dir = log/test_2   
batch_size = 40  
learning_rate = 1e-4    

The above manually changing method is time-wasting and error-prone.

Our framework is mainly focused on solving the problem and it allows you to arrange your experiment's results in a more convenient way.

Advantage

• Automatically increasing the experiment number in the dir name, once you define the path and dir name in the config file.

  E.g. you can define your middle results path(tensorboard_dir), model results(model_dir), note path(note_dir) in args.yaml in a yaml format.

  log_dir_params:
      tensorboard_dir: "./logs/tf_logs/resnet/log4tensorboard"
      model_dir: "./logs/tf_logs/resnet/log4model"
      note_dir: "./logs/tf_logs/resnet/note"
      self_increasing_mode: True       # False for overwriting the last dir
      pattern: "experiment"            # The dir name you want to use
      default_pattern: "defalut_name"
  

• Help you to manage your model's parameters and pass them to your model easily.

  E.g. you can define your parameters in args.yaml file under the "running_params" item.

  running_params:
      batch_size: 64
      gpu_num: 1
      classes_num: 2
      
      conv_weight_decay: 0.0001
      bn_is_training: True
      bn_decay: 0.997
      bn_epsilon: 1e-5
      bn_scale: Truei
  
      learning_rate: 1e-3
      decay_steps: 1000
      decay_rate: 0.7
  

• Copy the args.yaml file (which contains and manages you model's parameters) to the note path every time you run the program. If you wanna overwrite the last dir, change the self_increasing_mode item to "False".

  

Installation

Your don't need tensorflow or other deep learning framework to run.

• Download this repository:

  git clone https://github.com/lihaoliu-cambridge/deep-learning-model-saving-helper.git
  
  

Running

• Modify the args.yaml, add the parameters your deep learning model need under the "running_params" item .

• Pass the running_params (a python dict which contains the running parameters) to you own model.

• Finish you model, and run it:

  cd deep-learning-model-saving-helper
  python main.py
  

Question

Please open an issue or email 'lhliu1994@gmail.com' for any question.