Simple TensorFlow interface developed during MScR thesis, allows for definition of experiments using a config file (list of python dictionary entries). Multiple experiments can be run from the command line, or scheduled using a Grid Engine system.
Experiments are run by run.py using the run_task function which takes a config (python dictionary) as input.
Example config included at setup.py, two formats possible;
- Sequential computational graph (SimpleModel): one input, one output, sequential NN modules
- Customisable computational graph (GraphModel): requires handlers to be given names and an adjacency list to be defined in the config.
- Currently, GraphModel does not support concatenation of handlers, i.e. no multi-modal support.
- Splitting of handlers will work, i.e. multi-task learning.
Example data providers and handlers are available in the source directory.
Grid Engine run script included at run.sh, for the son of grid engine system used by our internal GPU cluster.
Plotting functions at viz.py, for learning curves and bar charts of multiple experiments.