hwr-experiments

Experiments for handwriting recognition

Notice

This repository is mainly for internal usage. If you want to get the data, please have a view at write-math.com/data

Update an existing classifier

If you have new data, but don't want to change the classifier (hence: no difference in preprocessing, features or model and no new symbols), you should do the following:

1. $ backup.py: 30 minutes
2. Update hwrt/misc/latex2writemathindex.csv
3. Update preprocessing data source to new file which was just created by backup.py
4. hwrt filter_dataset -s ~/GitHub/hwrt/hwrt/misc/symbols.yml -r ~/GitHub/hwr-experiments/raw-datasets/2015-11-14-17-06-handwriting_datasets-all-raw.pickle -d ~/GitHub/hwr-experiments/raw-datasets/2015-11-15-filtered-raw.pickle
5. Delete files in preprocessing, features and model folder
6. $ train.py: 30 minutes (eventually multiple times for SLP)

Add new symbols

See "update existing classifier", but with modifications:

1. Adjust the model files (number of output neurons)
2. Adjust the data source (look in hwr-experiments/raw-datasets)

After new training

1. Run $ test.py -n 3 (should be less than 5%)