Dinosaur Name Generator

Introduction

A dinosaurs name generator project implemented in Tensorflow.

Dataset

Training dataset is downloaded from this Kaggle Notebook.

Due to the shortage of data in training dataset, model will be pretrained on another dataset consisting Latin name of animals belong to 2 classes of Vertebrate: Aves and Mammalia from this repo.

Getting started

Training

Remind: Since a previously trained model weight has been saved in weights folder, a training step is optional.

By default, simply run the following command to train the model.

python train.py

Here is a description on the command-line arguments:

--training: Path to the training data file. Default: ./data/dino.txt
--pretrain: Path to the pretraining data directory. Default: ./data/pretrain
--embedding_size: Size of the embedding layer. Default: 256
--hidden_units: Number of hidden units in the model. Default: 128
--epochs: Number of epochs for training. Default: 20
--pretrain_epochs: Number of epochs for pretraining. Default: 20
--batch_size: Batch size for training. Default: 64
--learning_rate: Learning rate for the optimizer. Default: 0.001
--save_path: Path to save the trained model. Default: ./weight/model_v2.h5

Prediction

The command below requires an argument start as the starting of the name. start can be a character or a string.

python predict.py --start ___

Here is a description on the command-line arguments:

--start: starting of the name. Required.
--embedding_size: Size of the embedding layer. Default: 256
--hidden_units: Number of hidden units in the model. Default: 128
--save_path: Path to save the trained model. Default: ./weight/model_v2.h5