sketch

Project https://github.com/udibr/sketch

Use RNN to model handwriting using Theano blocks. I am trying to reproduce handwriting model given by Alex Graves, see also his demo

Dependencies

• Blocks follow the install instructions. This will install all the other dependencies for you (Theano, Fuel, etc.). See this note if using python 3.4
• Download handwriting dataset using this notebook

Running code

Run parameteres are used to build a unique name which is used to create a directory where the model is saved and also a sample image sketch.png

For example:

python sketch.py --dim 900 --depth 1 --bs 56 --lr 1e-3 -G 10
   Running experiment handwriting-1X900m20d5r13b56e15G10
   ...
   Epoch 1, step 105
   test_sequence_log_likelihood: -631.519287109
python sketch.py --dim 900 --depth 1 --bs 56 --lr 3e-4 -G 10 --model handwriting-1X900m20d5r13b56e15G10
   Running experiment handwriting-1X900m20d5r34b56e15G10
   ...
   Epoch 5, step 25
   test_sequence_log_likelihood: -950.866210938
python sketch.py --dim 900 --depth 1 --bs 56 --lr 1e-4 -G 10 --model handwriting-1X900m20d5r34b56e15G10
   Running experiment handwriting-1X900m20d5r14b56e15G10
   ...
   Epoch 153, step 165
   test_sequence_log_likelihood: -1482.28918457
python sketch.py --dim 900 --depth 1 --bs 56 --lr 1e-5  -G 10 --model handwriting-1X900m20d5r14b56e15G10
   Epoch 11, step 84
   test_sequence_log_likelihood: -1607.07531738

The result directory can be downloaded here and after opening it you can generate samples with:

python sketch.py --dim 900 --depth 1 --model handwriting-1X900m20d5r15b56e15G10 --sample