colinmorris / sketch-rnn-experiments

Experiments with Sketch-RNN and the Quick Draw dataset

A playground for experiments with the Quick Draw dataset and Sketch-RNN. Sorry about the mess.

get_perplexities, the function in the iPython notebook for calculating the loss per sketch in the dataset relies on a one-line local modification I made to Magenta to make it possible to grab the pre-aggregated loss of a batch. I added the following line after the call to get_lossfunc in sketch_rnn/model.py: self.lossfunc = lossfunc.

If you don't want to bother building Magenta from source, you can use _get_perplexities with a model having hps.batch_size = 1, but it's an order of magnitude slower.

To run the notebook code, you'll need at least one pre-trained Sketch-RNN model, saved locally to models/ (see sketch_rnn_train.download_pretrained_models), and at least one slice of Quick Draw data as an npz file saved locally to data/. Links here. I used flamingo/owl models and datasets, plus a few datasets for unrelated categories.