How to freeze Keras or Tensorflow 1.X graph

What is the problem?

Keras is easy to save the model by only saving weights as model_weights.h5. What if you need model.pb?

When naively saving the model with

keras.callbacks.ModelCheckpoint(self.checkpoint_path,
verbose=0, save_weights_only=False),

Compare the difference to this

keras.callbacks.ModelCheckpoint(self.checkpoint_path,
verbose=0, save_weights_only=True),

The first one will save the whole model: architecture and the model weights

the second one will save only the weights.

You may get some error like this for the first one.

Traceback (most recent call last):
File "sunrgbd.py", line 740, in <module>
input_channels=INPUT_CHANNELS_)
File "/Users/jimmy/anaconda3/envs/faceana/lib/python3.7/site-packages/mask_rcnn-2.1-py3.7.egg/mrcnn/model.py", line 2522, in train
File "/Users/jimmy/anaconda3/envs/faceana/lib/python3.7/site-packages/keras/legacy/interfaces.py", line 91, in wrapper
return func(*args, **kwargs)
File "/Users/jimmy/anaconda3/envs/faceana/lib/python3.7/site-packages/keras/engine/training.py", line 1732, in fit_generator
initial_epoch=initial_epoch)
File "/Users/jimmy/anaconda3/envs/faceana/lib/python3.7/site-packages/keras/engine/training_generator.py", line 260, in fit_generator
callbacks.on_epoch_end(epoch, epoch_logs)
File "/Users/jimmy/anaconda3/envs/faceana/lib/python3.7/site-packages/keras/callbacks/callbacks.py", line 152, in on_epoch_end
callback.on_epoch_end(epoch, logs)
File "/Users/jimmy/anaconda3/envs/faceana/lib/python3.7/site-packages/keras/callbacks/callbacks.py", line 730, in on_epoch_end
self.model.save(filepath, overwrite=True)
File "/Users/jimmy/anaconda3/envs/faceana/lib/python3.7/site-packages/keras/engine/network.py", line 1152, in save
save_model(self, filepath, overwrite, include_optimizer)
File "/Users/jimmy/anaconda3/envs/faceana/lib/python3.7/site-packages/keras/engine/saving.py", line 449, in save_wrapper
save_function(obj, filepath, overwrite, *args, **kwargs)
File "/Users/jimmy/anaconda3/envs/faceana/lib/python3.7/site-packages/keras/engine/saving.py", line 541, in save_model
_serialize_model(model, h5dict, include_optimizer)
File "/Users/jimmy/anaconda3/envs/faceana/lib/python3.7/site-packages/keras/engine/saving.py", line 129, in _serialize_model
model_config['config'] = model.get_config()
File "/Users/jimmy/anaconda3/envs/faceana/lib/python3.7/site-packages/keras/engine/network.py", line 950, in get_config
return copy.deepcopy(config)
File "/Users/jimmy/anaconda3/envs/faceana/lib/python3.7/copy.py", line 150, in deepcopy
y = copier(x, memo)
File "/Users/jimmy/anaconda3/envs/faceana/lib/python3.7/copy.py", line 241, in _deepcopy_dict
y[deepcopy(key, memo)] = deepcopy(value, memo)
File "/Users/jimmy/anaconda3/envs/faceana/lib/python3.7/copy.py", line 150, in deepcopy
y = copier(x, memo)
File "/Users/jimmy/anaconda3/envs/faceana/lib/python3.7/copy.py", line 216, in _deepcopy_list
append(deepcopy(a, memo))
File "/Users/jimmy/anaconda3/envs/faceana/lib/python3.7/copy.py", line 150, in deepcopy
y = copier(x, memo)
File "/Users/jimmy/anaconda3/envs/faceana/lib/python3.7/copy.py", line 241, in _deepcopy_dict
y[deepcopy(key, memo)] = deepcopy(value, memo)
File "/Users/jimmy/anaconda3/envs/faceana/lib/python3.7/copy.py", line 150, in deepcopy
y = copier(x, memo)
File "/Users/jimmy/anaconda3/envs/faceana/lib/python3.7/copy.py", line 241, in _deepcopy_dict
y[deepcopy(key, memo)] = deepcopy(value, memo)
File "/Users/jimmy/anaconda3/envs/faceana/lib/python3.7/copy.py", line 150, in deepcopy
y = copier(x, memo)
File "/Users/jimmy/anaconda3/envs/faceana/lib/python3.7/copy.py", line 221, in _deepcopy_tuple
y = [deepcopy(a, memo) for a in x]
File "/Users/jimmy/anaconda3/envs/faceana/lib/python3.7/copy.py", line 221, in <listcomp>
y = [deepcopy(a, memo) for a in x]
File "/Users/jimmy/anaconda3/envs/faceana/lib/python3.7/copy.py", line 150, in deepcopy
y = copier(x, memo)
File "/Users/jimmy/anaconda3/envs/faceana/lib/python3.7/copy.py", line 221, in _deepcopy_tuple
y = [deepcopy(a, memo) for a in x]
File "/Users/jimmy/anaconda3/envs/faceana/lib/python3.7/copy.py", line 221, in <listcomp>
y = [deepcopy(a, memo) for a in x]
File "/Users/jimmy/anaconda3/envs/faceana/lib/python3.7/copy.py", line 180, in deepcopy
y = _reconstruct(x, memo, *rv)
File "/Users/jimmy/anaconda3/envs/faceana/lib/python3.7/copy.py", line 281, in _reconstruct
state = deepcopy(state, memo)
File "/Users/jimmy/anaconda3/envs/faceana/lib/python3.7/copy.py", line 150, in deepcopy
y = copier(x, memo)
File "/Users/jimmy/anaconda3/envs/faceana/lib/python3.7/copy.py", line 241, in _deepcopy_dict
y[deepcopy(key, memo)] = deepcopy(value, memo)
File "/Users/jimmy/anaconda3/envs/faceana/lib/python3.7/copy.py", line 180, in deepcopy
y = _reconstruct(x, memo, *rv)
File "/Users/jimmy/anaconda3/envs/faceana/lib/python3.7/copy.py", line 281, in _reconstruct
state = deepcopy(state, memo)
File "/Users/jimmy/anaconda3/envs/faceana/lib/python3.7/copy.py", line 150, in deepcopy
y = copier(x, memo)
File "/Users/jimmy/anaconda3/envs/faceana/lib/python3.7/copy.py", line 241, in _deepcopy_dict
y[deepcopy(key, memo)] = deepcopy(value, memo)
File "/Users/jimmy/anaconda3/envs/faceana/lib/python3.7/copy.py", line 180, in deepcopy
y = _reconstruct(x, memo, *rv)
File "/Users/jimmy/anaconda3/envs/faceana/lib/python3.7/copy.py", line 281, in _reconstruct
state = deepcopy(state, memo)
File "/Users/jimmy/anaconda3/envs/faceana/lib/python3.7/copy.py", line 150, in deepcopy
y = copier(x, memo)
File "/Users/jimmy/anaconda3/envs/faceana/lib/python3.7/copy.py", line 241, in _deepcopy_dict
y[deepcopy(key, memo)] = deepcopy(value, memo)
File "/Users/jimmy/anaconda3/envs/faceana/lib/python3.7/copy.py", line 180, in deepcopy
y = _reconstruct(x, memo, *rv)
File "/Users/jimmy/anaconda3/envs/faceana/lib/python3.7/copy.py", line 281, in _reconstruct
state = deepcopy(state, memo)
File "/Users/jimmy/anaconda3/envs/faceana/lib/python3.7/copy.py", line 150, in deepcopy
y = copier(x, memo)
File "/Users/jimmy/anaconda3/envs/faceana/lib/python3.7/copy.py", line 241, in _deepcopy_dict
y[deepcopy(key, memo)] = deepcopy(value, memo)
File "/Users/jimmy/anaconda3/envs/faceana/lib/python3.7/copy.py", line 169, in deepcopy
rv = reductor(4)
TypeError: can't pickle _thread.RLock objects

How to solve the problem?

get a frozen graph and then save it.[1]

def freeze_session(session, keep_var_names=None, output_names=None, clear_devices=True):
    """
    Freezes the state of a session into a pruned computation graph.

    Creates a new computation graph where variable nodes are replaced by
    constants taking their current value in the session. The new graph will be
    pruned so subgraphs that are not necessary to compute the requested
    outputs are removed.
    @param session The TensorFlow session to be frozen.
    @param keep_var_names A list of variable names that should not be frozen,
                          or None to freeze all the variables in the graph.
    @param output_names Names of the relevant graph outputs.
    @param clear_devices Remove the device directives from the graph for better portability.
    @return The frozen graph definition.
    """
    graph = session.graph
    with graph.as_default():
        freeze_var_names = list(set(v.op.name for v in tf.global_variables()).difference(keep_var_names or []))
        output_names = output_names or []
        output_names += [v.op.name for v in tf.global_variables()]
        input_graph_def = graph.as_graph_def()
        if clear_devices:
            for node in input_graph_def.node:
                node.device = ""
        frozen_graph = tf.graph_util.convert_variables_to_constants(
            session, input_graph_def, output_names, freeze_var_names)
        return frozen_graph

Then [1]

from keras import backend as K

# Create, compile and train model...

frozen_graph = freeze_session(K.get_session(),
                              output_names=[out.op.name for out in model.outputs])
tf.train.write_graph(frozen_graph, "some_directory", "my_model.pb", as_text=False)

You are done. I did it by following [1] and it works pretty well for TensorFlow 1.14.

Reference

[1] https://stackoverflow.com/questions/45466020/how-to-export-keras-h5-to-tensorflow-pb