Metadata-Version: 2.1
Name: text-gen
Version: 1.1.0
Summary: build a text generation model
Home-page: https://github.com/Emekaborisama/textgen
Author: Emeka Boris Ama
Author-email: borisphilosophy@gmail.com
License: MIT
Description: 
        <h1 align="center">
        <p>Text Gen :goat:</p>
        
        <p align="center">
        
        
        <img alt="python" src="https://img.shields.io/badge/python-%3E%3D3.6-blue?logo=python">
        <img alt="tensorflow" src="https://img.shields.io/badge/tensorflow-%3E%3D2.3.0-orange?logo=tensorflow">
        <a href="https://pypi.org/project/text-gen/">
          <img alt="PyPI" src="https://img.shields.io/pypi/v/text-gen?color=%234285F4&label=release&logo=pypi&logoColor=%234285F4">
        </a>
        </p>
        </h1>
        <h2 align="center">
        <p>Almost State-of-the-art Text Generation library</p>
        </h2>
        
        <p align="center">
        Text gen is a python library that allow you build a custom text generation model with ease :smile:
         Something sweet built with Tensorflow and Pytorch(coming soon) - This is the brain of Rosalove ai (https://rosalove.xyz/)</p>
        
        
        
        ## How to use it
        Install text-gen
        ```bash
        pip install -U text-gen
        ```
        import the library
        ```python
        from text_gen import ten_textgen as ttg
        ```
        
        
        Load your data. your data must be in a text format.
        
        Download the example data from the [example folder](https://github.com/Emekaborisama/textgen/tree/master/example)
        #### load data
        ```python
        data = 'rl.csv'
        text = ttg.loaddata(data)
        ```
        
        
        #### parameters
        You can choose to write this argument internally in the train model method
        ```python 
        activation = 'softmax'
        lstmlayer = 128
        padding_method = 'pre'
        
        loss='categorical_crossentropy'
        optimizer='adam'
        metrics='accuracy'
        epochs=500
        verbose = 0
        patience = 10
        batch = 300
        dropout = 0.25
        
        ```
        
        
        ```python
        pipeline = ttg.tentext(text)
        seq_text = pipeline.sequence(padding_method)
        configg = pipeline.configmodel(seq_text, lstmlayer, activation, dropout = dropout)
        
        ```
        
        
        #### train model
        ```python
        model_history = pipeline.fit(loss = loss, optimizer = optimizer, batch = batch, metrics = metrics, epochs = epochs, verbose = 0, patience = patience)
        
        ```
        
        
        #### generate text using the phrase
        ```python
        pipeline.predict('hello love', word_length = 200)
        ```
        
        #### sentence sengment the output
        ```python
        output = pipeline.predict('hello love', word_length = 200)
        pipeline.segment(output)
        ```
        
        
        #### plot loss and accuracy
        ```python
        pipeline.plot_loss_accuracy()
        ```
        
        #### Hyper parameter optimization
        Tune your model to know the best optimizer, activation method to use.
        ```python
        pipeline.hyper_params(epochs = 500)
        ```
        
        <h1 align="center">
        <span> Give us a star :star: </span> 🐉
        </h1>
        
        If you want to contribute, take a look at the issues and the [Futurework.md](https://github.com/Emekaborisama/textgen/blob/master/futurework.md) file 
        
        
        Contributors 
        
        - [Emeka Boris Ama](https://twitter.com/emeka_boris)
        - [King Zikie](https://twitter.com/kingzikie)
        - [David](https://twitter.com/iEphods)
        
        
Platform: UNKNOWN
Description-Content-Type: text/markdown
