The goal of this project is to build a model that is able to generate different film synopsys from a set of predefined genres. The model is LSTM one to many
For training the net we will be using a dataset of >100K pairs of <genres,synopsis> (data is in spanish).
Lets put the code in the src
folder and all the input and output data in the data
folder (without pushing any data to the repo).
.
├── data
│ ├── others
│ │ └── predictions
│ ├── tensorboard_logs
│ └── weights
├── notebooks
├── src
│
└── tensorboard_logs
- RNN basics: The Unreasonable Effectiveness of Recurrent Neural Networks
- LSTM basics: Understanding LSTM Networks
- At prediction time we will use Beam Search to generate various synopsis from the same set of genres: Beam Search video
- Repo that uses LSTM + beamsearch to generate headlines of paragraphs: headlines
- Repo that uses a keras one-to-many model as well (for image captioning specifically): caption_generator