Published April 9, 2021 | Version 1.2.2

IchiruTake/Bit2Edge: AIPBDET

Authors/Creators

Description

Change relatively to version 1.2.1

This release has fixed some several issues as well as gaining some better performance received.

-> Project

  • [Minor Performance Boost]: Now it is faster to validate the data type by 15% - 25% by changing syntax (Syntax: type(a) is B or type(a) is C --> isinstance(a, (B, C))) but it is not significant. Test case show that for 50M loop, OLD syntax: 6.5515s <--> NEW syntax: 5.4085s with 10% memory cache lower. Apply for all Python file.

-> Preprocessing.py

  • [API Changed]: Function 'DatabaseMinimization' --> NonlabelCleaning
  • [New Feature]: Function 'IndexBasedCleaning' now has been introduced which shares some identical to 'NonlabelCleaning'
  • [API Changed]: Hyper-parameter are varied between "FileName" and "filePath" is now merged as "FileName".

-> Creator.py

  • [Bug Fixed]: Index-based Cleaning is now worked when "strict_cleaning=True". Now it has been passed to Preprocessing.py "IndexBasedCleaning()"
  • [Method Changed]: function ".getFeatureColumn()" now has extrahyper-parameter "number_of_input" which is directly attached with ".createData()" in the same class and return the original column file. That column file is saved into the object.

-> Learning.py

  • [Attribute Changed]: Some attribute's names has been changed to make better adaptation.
  • [Method Changed]: ".getModelRatio()" now has been converted to "prepare_model()". Add extra hyper-parameter "DataCleaning" and "strict_cleaning" which performs exactly in "DataGenerator.createData()" but it only works if 'retrainModel' is False. Note that using 'retrainModel' is typically dangerous as it did not pass much validation test case. This hyper-parameter is only used if when the training process is corrupted. This supplementation is applied after we have finished all of the value, and only be used if you want to apply Data Cleaning in the "training set" only (NOT on validation set or testing set).

-> Predict.py

  • [New Feature]: Function ".generate()" is now introduced to calculate some configuration that can be applied to some other function. This method is directly attached into ".createData()" (extra 4KB) but better design pattern for debugging. Reduced 2 - 3 seconds when need to find special position to locate instead of searching at everytime called. Extracted from ".process_available_data()"
  • [Optimization]: Function ".generate()" is now faster than older version by 15-20% (from 4.4-4.6 s to 3.5-3.7s) by using built-in Python method rather than Numpy.char; replace some while-loop which performs same behaviour as for-loop.

  • [Bug Fixed]: Function ".process_available_data()" now accept self.number_of_input=4

  • [Attribute Changed]: self.n_bondType_Saved is now corrected as self.bondType_Saved
  • [Documentation]: now PredictModel() will print out some extra line to separate method section at IDE Console // Terminal to make easier verification.

  • [NEW]: Function ".estimate_density()" is introduced to estimate the fingerprint status relationship versus BDE estimation or BDE Error. This model will return two DataFrame(s), one for observations and one for overview

    • "sparsity_mode": bool (default to False). If True, the density array would then be changed into "sparsity".
    • "percentage_mode": bool (default to False). If True, all of the data will be multiplied with 100.
    • "dataType": numpy.dtype (default to numpy.float32). Don't select integer as dataType as all the result will be turned all into zero. Suggestion: float, np.float16, np.float32, ... (np.float64 or TF.dtype or PyTorch.dtype is not recommended).
    • "output": str (default to None). Choose the path and name to create a .csv file. If the filename contain ".csv" at the end, it will create the csv of DataFrame only. Else, it will create both DataFrame and Series into two .csv file with extra name behind.
  • [API Changed]: Function ".visualize()" now has removed "axis-number" as hyper-parameter.

  • [Behaviour Changed]: Function ".visualize()" sometimes required user's key input at some specific "chosen_type".
  • [Design Pattern + Performance]: Function ".visualize()": Hyper-parameter "chosen_type" will be called and assigned the lower mode of "chosen_type" once.
  • [New Feature]: Function ".visualize()": Hyper-parameter "decomposeMethod" now use "umap" as default with n_neighbors is equal to the number of unique bond recorded in the input dataset multiplied by 2.
  • [Behaviour Changed]: Predict.py: "visualize" method now accept some extra "chosen_type". With "chosen_type = "last" --> Remain the same. With "chosen_type = "fingerprint" ("fingerprints", "multi-fingerprints") --> Remain the same but can select the coloring representation. --> Compute all fingerprint and merge as three alone feature With "chosen_type = "single" --> Compute each 'selected' environment fingerprint and return as one value only per each With "chosen_type = "meaning" ("nature", "attribute") --> Compute large (Full Structural) and smaller (Radicals) environments into two values for representation With "chosen_type = "density" or "sparsity" --> Visualize data density / sparsity that correlate to BDE or bond type. However, this method only create new density DataFrame only if it is not stored when calling ".estimate_density()". If you want to rebuilt new ones, use ".reset_density()"
  • [Adaptation]: Now ".predict()" don't raise mis-understanding warning when "get_last_layer=False"

Note: Next batch will introduce documentation to give better understanding for scaling and reproducing.

Files

IchiruTake/Bit2Edge-1.2.2.zip

Files (299.0 kB)

Name Size
md5:71f88cae4e6ff83cd51b358fa18f97b5
299.0 kB Preview Download

Additional details

Related works