Skip to main content

This is a module for Hyper-parameter tuning and rapid prototyping

Project description

autoprototype

This library is designed for optimization of hyper-parameters in Deep Learning and Machine Learning Models. This module aims at giving the user the best set of hyper-parameters for their DL/ML models. Along with suggesting the hyper-parameters, this module aims at returning to the user the best suggestive model framework/architecture based on the user input data. The user has to just provide the dataset and the module will return :

  1. The best model architecture based on the input data.
  2. The best hyper-parameters for the model architecture.

In this way, the whole process of experimenting with the architecture is narrowed down. This is easy and requires only few lines of codes. Apart from the input data, only few other parameters are required by the module. All, the processes in rapid prototyping is automated thereafter, through this module. The structure is based on some default values for parameters spaces required for the optimization. However, user would also have freedom to dynamically construct the search spaces for the hyperparameters.

This module is a wrap around the popular Hyper-parameter Optimization Tool called Optuna. Optuna is used for the optimization process using iterative trails. This module takes the data as the primary input and suggests the user the model based on this Optuna trials. Optuna enables efficient hyperparameter optimization by adopting state-of-the-art algorithms for sampling hyper-parameters and pruning efficiently unpromising trials.

Some key features of Optuna, that we used are:

1.Lightweight, versatile, and platform agnostic architecture

  1. Pythonic search spaces

  2. Efficient Optimization Algorithms

This module, includiong all examples, is entirely written on Python.

Installation Guide

Users can use pip install to install the packages to start working on the prototyping. Below is given the pip command.

pip install autoprototype

The usage or the API calls for each supported ML/DL libraries are shown in the next sections.

#Supported Integrations Libraries

##SKLearn At this point, the following models and their respective hyper-parameters are supported

a. Decision Tree

b. SVC

c. Linear Regression

d. Ridge and Lasso Regression

e. Random Forest

f. Logistic Regression

Starting the prototyping process consists of just two line of code

    from autoprototype.sklearn import sklearnopt
    hpo = sklearnopt(X_train,y_train)
    trial , params , value = hpo.get_best_params(n_trials=trials)

This would return to us, the best trail, the set of best parameters including the model and the best objective value based on which the optimization is done.

To run the example navigate to examples and use:

python iris.py

tf.keras

At this point this supports, the following models and their hyper-parameters are supported:

Artificial Neural Networks

This also requires few lines of codes as shown below.

    from autoprototype.tf_keras import kerasopt
    hpo = kerasopt(x_train,y_train,EPOCHS=10,classes=CLASSES)
    trial , params , value = hpo.get_best_params(n_trials=n_trials)

By default it would run the trials and suggest you:

a. n_layers : Number of hidden layers in the model.

b. units_i : Number of units in each layers

c. dropout_rate : the dropout rate

d. lr : Learning rate of optimizers

e. optimizer_name : Name of the best optimizer

The loss criterion and the maximum number of layers, is set to sparse_categorical_crossentropy and 5 respectively, by default. User can also provide any other loss function based on requirement as follows:

hpo = HyperparamOpt(x_train,y_train,EPOCHS=10,classes=CLASSES,loss="your_loss_function",n_ann_layers=number_of layers)

To run the ANN example, navigate to examples and run :

python ann_tfkeras.py

Convolution Neural Network

The API for CNN model is fairly the same as that of the above ANN. The user, must pass few other optional parameters to construct the suggestive CNN architecture. All other syntaxes are same.

    hpo = kerasopt(x_train,y_train,EPOCHS=10,classes=120,
                   max_units_fcl=400, max_conv_filters=1000,
                   arch="cnn",input_shape=(128,128,3),steps_per_epoch=10)

Two mandatory arguments to run CNN:

arch refers to the specification of architecture to 'cnn',

input_shape must be provided for any CNN models.

Other optional arguments

max_units_fcl : max units in the fully connected layers of CNN, default 1000

max_nconv : max number of convolution layers, default 5

max_fully_conn_layers : max number of fully connected layers. default 2

steps_per_epoch : Number of steps to be taken in an epoch, default 216

max_conv_filters : maximum filters in the convolution layers, default 256

To run the CNN example [tf_keras/examples] and run :

Download the data from :

https://www.kaggle.com/c/dog-breed-identification/data

Navigate to examples folder

cd examples Make an empty directory called data. mkdir data move the downloaded data into the data folder. Please check the path to data provided in the cnn_tfkeras.py script!

Run: python cnn_tfkeras.py

Project details


Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

autoprototype-1.2.2.tar.gz (7.5 kB view details)

Uploaded Source

Built Distribution

autoprototype-1.2.2-py3-none-any.whl (7.6 kB view details)

Uploaded Python 3

File details

Details for the file autoprototype-1.2.2.tar.gz.

File metadata

  • Download URL: autoprototype-1.2.2.tar.gz
  • Upload date:
  • Size: 7.5 kB
  • Tags: Source
  • Uploaded using Trusted Publishing? No
  • Uploaded via: twine/3.4.2 importlib_metadata/4.6.1 pkginfo/1.5.0.1 requests/2.22.0 requests-toolbelt/0.9.1 tqdm/4.42.1 CPython/3.7.6

File hashes

Hashes for autoprototype-1.2.2.tar.gz
Algorithm Hash digest
SHA256 61c4473dac8e7d586fc0187577c15532849871ca167f43dc6794fbc19dbaa5e3
MD5 cf1dc6fe88a71bc69816a92e4423b0c9
BLAKE2b-256 a5433730f5bc53c27b30aaa3f1843bfda477886b2ae38ed9c802ca9865b0f061

See more details on using hashes here.

File details

Details for the file autoprototype-1.2.2-py3-none-any.whl.

File metadata

  • Download URL: autoprototype-1.2.2-py3-none-any.whl
  • Upload date:
  • Size: 7.6 kB
  • Tags: Python 3
  • Uploaded using Trusted Publishing? No
  • Uploaded via: twine/3.4.2 importlib_metadata/4.6.1 pkginfo/1.5.0.1 requests/2.22.0 requests-toolbelt/0.9.1 tqdm/4.42.1 CPython/3.7.6

File hashes

Hashes for autoprototype-1.2.2-py3-none-any.whl
Algorithm Hash digest
SHA256 4055a63aec8e749c527135bae8c9f772a5a3972434e49acf376e3f93feda4e64
MD5 5f92db946f74c980bda517bf84f5e30d
BLAKE2b-256 d83130d0cfa0165d0c5828604d415bcdf267b1d2b63750b65090bd71e3610da0

See more details on using hashes here.

Supported by

AWS AWS Cloud computing and Security Sponsor Datadog Datadog Monitoring Fastly Fastly CDN Google Google Download Analytics Microsoft Microsoft PSF Sponsor Pingdom Pingdom Monitoring Sentry Sentry Error logging StatusPage StatusPage Status page