ImageCaptioning
This repo contains an implementation of a ImageCaptioning model. It was implemented as a part of 4 ECTS course Deep Learning of the Data Science Bachelor at the FHNW.
Architecture
The architecture is basically as follows:
- A pretrained CNN-model (e.g.
ResNet50) is used to generate features from the images. - With the help of an embedding, the dimension is adapted to the vocab size and the embedding dimension is selected based on available computing resources. Technically, a higher dimension should be better but it takes longer to train and requires more resources.
- This vector is then passed as the first
hidden statein a LSTM.
Futher details
Please have a look at main.ipynb