Is there any way to import graphdef in Python and run it?
I am implementing a graph optimizer, similar to a grappler pass but is implemented in the Python frontend. Thus, the python optimizer would be:
graph_to_optimize.pbtxt -> Python graph optimizing -> optimized_graph.pbtxt
Thus, in the C++ grappler pass, the above process could be implemented as :
// Dump graph DumpGraphDefToFile("graph_to_optimize", item.graph); // Optimizing in Python // python_opt_process() // Read optimized graph GrapplerItem tempitem; ReadGraphDefFromFile("optimized_graph.pbtxt", &tempitem.graph); // Copy to optimized one CopyGraph(tempitem, optimized_graph);
The above code requires building TF from the source, I wish to avoid the building from the source. I want to ask if there is a pure Python way, to let users use the optimization process from stock TF. I.e., the user could directly import the graphdef from the file and replace the previous graph.
I found a way of
import_graph_def. If this is the answer, could you offer an example of how could I import graphDef and let it run the training task?
Thanks in advance!