Keras model freezer

Complete exemplar usage

I provide complete command sequence to create static, readonly tensorflow graph from model created in keras framework.

Following commands to those things:

1. Permute convolutional kernel dimensions from channels_first to channels_last (meaning is explained in keras docs) and shuffle dense layers weights accordingly.
2. Create ckpt files that are usable in raw tensorflow. Additionaly, all op names are dumped (they will be used in later commands). pbtxt graph from session is also dumped.
3. Freeze model using ckpt and pbtxt files to create static, readonly graph. Input node names and output node names are provided in op_names.txt file*.
4. Optimize model for inference ( :) ). This means, few ops are fused to be executed in one step, dropouts are deleted etc.

python th_to_tf_converter.py -w <path_to_weights.h5> -j <not_required_json_architecture.json> -c -o <output_directory_path>
python save_graph_and_checkpoint.py -w <path_to_weights.h5> -j <not_required_json_architecture.json> -o <output_directory_path>
python freeze_graph.py --input_checkpoint <input_model_checkpoint_path.ckpt> --output_graph <output_graph_path.pb> --output_node_names <comma_separated_output_nodes_names> --input_graph <input_graph_path.pbtxt>
python optimize_for_inference.py --input <input_graph_path.pb> --output <output_graph_path.pb> --frozen_graph=True --output_names <comma_separated_output_nodes_names> --input_names <comma_separated_input_nodes_names>

*You are required to focus and choose wisely those names. Default name for input tensor is input_1 for single input network and Sigmoid, Softmax... for output nodes. It can be different in your case (especially if you customized layer names or something like this). More info on op names are below.

Guidelines to find node names:

1. Input nodes:

   1. They usually are first in op_names.txt file. If you customized layer names, they can be different than input_1.

2. Output nodes:

   1. They usually have name provided by activation function name with capital letter (ex. Sigmoid)
   2. If modified name, they can be found before following sequence of names in op_names.txt:

   ...
   Sigmoid
   Placeholder
   Assign
   Placeholder_1
   Assign_1
   Placeholder_2
   Assign_2
   Placeholder_3
   Assign_3
   Placeholder_4
   Assign_4
   Placeholder_5
   Assign_5
   Placeholder_6
   Assign_6
   ...
   

   In this case, Sigmoid is our output node name

3. Naming nodes:

• if you finally found your nodes' names, then to get nodes themselves by name you have to call method on tf.Graph object get_tensor_by_name(<name>) where <name> should be written as import/<found_node_name>:0