Train Utility

It is still in development stage, i.e. name and content of classes and functions in this library may be changed or removed

Train Utility is a simple utility for training in Pytorch.

Overview

Class

ModelInfo
TrainResultInfo
TrainProcess
TrainProcess_DictLoad
TrainProcess_DictSave
TrainProcess_Folder
TrainProcess_Hook
TrainProcess_PythonFile
TrainProcess_ResultGraph
TrainProcess_ResultRecord
Interface_CodePath
INterface_DictData

Function

plotConfusionMatrix
plotAccuracy
plotLoss

Example

Create ModelInfo

ModelInfo store all the information about the model and training. The primary usage of ModelInfo is to record all the object and data that will be used in training so we can just pass this single object instead of many parameters into training function.
Below will demonstrate creating ModelInfo and filling the minimum data into the ModelInfo.

# create TrainProcess
info = ModelInfo()

# fill data
info.model      = # model (nn.Module)
info.optimizer  = # optimizer (optim)

info.epoch          = 100
info.batch_size     = 2
info.learning_rate  = 1e-4

Create TrainProcess and Signaling

TrainProcess will be executed when a specific stage is reached. TrainProcess is the abstract class and should not be created directly. Below is the list of some children class of TrainProcess.

• DictLoad
• DictSave
• Folder
• Hook
• PythonFile
• ResultGraph
• ResultRecord

Example will use TrainProcess_Folder, which will create a folder and normally needed to be called after the training is finished and some training result need a space to store.

# create TrainProcess_Folder
process = TrainProcess_Folder()

# TrainProcess_Folder will use the data in ModelInfo to create the folder
# here, it is assumed that ModelInfo (model) is created and minimum data is filled
# we need to further fill more information for TrainProcess_Folder
info.save_path      = "./Result"
info.save_folder    = "Result_xxx"

# add TrainProcess_Folder to ModelInfo
info.process_list.append(process)

To trigger the process, we need to call the ModelInfo.executeProcess(stage, info). There are some default stages in ModelInfo.Stage.

• TRAIN_START
• ITERATION_TRAIN_START
• ITERATION_TRAIN_END
• ITERATION_VAL_START
• ITERATION_VAL_END
• TRAIN_END

Here we need to trigger TrainProcess_Folder at TRAIN_END stage.

# ----- init -----
# acutally we don't need to set this
# ModelInfo.Stage.TRAIN is added to TrainProcess_Folder.stage in __init___
process.stage = [ModelInfo.Stage.TRAIN_END]

# ----- train end ------
# data: Dict
# what inside data is depended on different stage and user is free to add more into it
info.executeProcess(ModelInfo.Stage.TRAIN_END, data)

Save Epoch Result

TODO: Not yet completed

Plot Graph

TODO: Not yet completed

Repo that Use this Library

• https://github.com/Jack-GVDL/FastRCNN