Repository files navigation

Update 2019/06/24: A model trained on 10% of the Shepard-Metzler dataset has been added, the following notebook explains the main features of this model:nbviewer

This is a PyTorch implementation of the Generative Query Network (GQN)described in the DeepMind paper "Neural scene representation andrendering" by Eslami et al. For an introduction to the model and problemdescribed in the paper look at the article byDeepMind.

The current implementation generalises to any of the datasets describedin the paper. However, currently,only the Shepard-Metzler dataset hasbeen implemented. To use this dataset you can use the provided script in

sh scripts/data.sh data-dir batch-size

The model can be trained in full by in accordance to the paper by running thefilerun-gqn.py or by using the provided training script

sh scripts/gpu.sh data-dir

The implementation shown in this repository consists of all of therepresentation architectures described in the paper along with thegenerative model that is similar to the one described in"Towards conceptual compression" by Gregor et al.

Additionally, this repository also contains implementations of theDRAWmodel and the ConvolutionalDRAW model both described by Gregor et al.