Utility functions for deployment

The following functions will help you to deploy your models.

Freeze graph

As you probably may know, TensorFlow saves basically four files when training a model:

• checkpoint
• model.ckpt.data-00000-of-00001
• model.ckpt.index
• model.ckpt.meta

It may be awkward to deal with all of these files for deployment and it is preferrible to deal with only one that contains both graph and weights.

You can do that by freezing the graph with freeze_graph.py. What it basically does is to transform all the graph variables to constants and dump a single .pb file.

You can both embed this function to your code or use it as an utility after your training procedure

Example

python freeze_graph.py --model_dir /path/to/ckpt/ --output_tensors tensor_list --output_pb /path/to/model.pb

Visualize graph

If you want to visualize your graph using Tensorboard, you first need to generate the events files.

For this, use pb_viewer.py

You can both embed this function to your code or use it as an utility after freezing your model.

Example

python pb_viewer.py --input_graph_pb /path/to/model.pb --output_events_file /path/to/event_files