GenSynth Documentation

Preparing your Model

To run GenSynth, you need a model metafile and a folder containing a checkpoint file, and any *.data and *.index files. GenSynth will load the data specified by the model_checkpoint_path in the checkpoint file. These files are generated when saving a model through the tf.train.Saver.

  • The metafile is the saved MetaGraphDef file, describing all of the tensors and network operations.

  • The checkpoint contains variable values.

For example, create the meta and checkpoint files with Tensorflow's tf.train.Saver.save(..., write_meta_graph=True) function.

If you select the model from the GenSynth resources, the README file for each model has information about the inputs and outputs of the model. It is worth looking at a couple of these examples.

Place the network metafile and checkpoint files in an appropriate folder within your workspace so that all workers can read the files.

Once you do that, GenSynth requires some information about your model.

What are the names and shapes of the input and label nodes? Often these nodes will be the get_next operators obtained from dataset iterators. In other cases they might be placeholders.

GenSynth may remove portions of your network that do not reside between the inputs and outputs. If your network has preprocessing, be sure to identify the inputs to the earliest processing that you want in the deployed network.

Warning

GenSynth replaces whatever data input pipeline there may be in the saved model with a custom pipeline in order to feed examples and data through the network.

In order to run GenSynth, you will need to create a dataset interface that can feed data to these inputs. See Preparing your Data for more information.

Note

You will need to prepare both your model and your data before you can create a new model in GenSynth.