We deploy lot of our models from TF1 by saving them through graph freezing:
tf.train.write_graph(self.session.graph_def, some_path)
# get graph definitions with weights
output_graph_def = tf.graph_util.convert_variables_to_constants(
self.session, # The session is used to retrieve the weights
self.session.graph.as_graph_def(), # The graph_def is used to retrieve the nodes
output_nodes, # The output node names are used to select the usefull nodes
)
# optimize graph
if optimize:
output_graph_def = optimize_for_inference_lib.optimize_for_inference(
output_graph_def, input_nodes, output_nodes, tf.float32.as_datatype_enum
)
with open(path, "wb") as f:
f.write(output_graph_def.SerializeToString())
and then loading them through:
with tf.Graph().as_default() as graph:
with graph.device("/" + args[name].processing_unit):
tf.import_graph_def(graph_def, name="")
for key, value in inputs.items():
self.input[key] = graph.get_tensor_by_name(value + ":0")
We would like to save TF2 models in similar way. One protobuf file which will include graph and weights. How can I achieve this?
I know that there are some methods for saving:
keras.experimental.export_saved_model(model, 'path_to_saved_model')
Which is experimental and creates multiple files :(.
model.save('path_to_my_model.h5')
Which saves h5 format :(.
tf.saved_model.save(self.model, "test_x_model")
Which agains save multiple files :(.
I use TF2 to convert model like:
- pass
keras.callbacks.ModelCheckpoint(save_weights_only=True) to model.fit and save checkpoint while training;
- After training,
self.model.load_weights(self.checkpoint_path) load checkpoint, and convert to h5: self.model.save(h5_path, overwrite=True, include_optimizer=False);
- convert
h5 to pb:
import logging
import tensorflow as tf
from tensorflow.compat.v1 import graph_util
from tensorflow.python.keras import backend as K
from tensorflow import keras
# necessary !!!
tf.compat.v1.disable_eager_execution()
h5_path = '/path/to/model.h5'
model = keras.models.load_model(h5_path)
model.summary()
# save pb
with K.get_session() as sess:
output_names = [out.op.name for out in model.outputs]
input_graph_def = sess.graph.as_graph_def()
for node in input_graph_def.node:
node.device = ""
graph = graph_util.remove_training_nodes(input_graph_def)
graph_frozen = graph_util.convert_variables_to_constants(sess, graph, output_names)
tf.io.write_graph(graph_frozen, '/path/to/pb/model.pb', as_text=False)
logging.info("save pb successfully！")
The way I do it at the moment is TF2 -> SavedModel (via keras.experimental.export_saved_model) -> frozen_graph.pb (via the freeze_graph tools, which can take a SavedModel as input). I don't know if this is the "recommended" way to do this though.
Also, I still don't know how to load back the frozen model and run inference "the TF2 way" (aka no graphs, sessions, etc).
You may also take a look at keras.save_model('path', save_format='tf') which seems to produce checkpoint files (you still need to freeze them, though, so I personally think the saved model path is better)
the above code is a little old. when convert vgg16, it could succeed, but it failed when convert resnet_v2_50 model. my tf version is tf 2.2.0
finally, I found a useful code snippet:
import tensorflow as tf
from tensorflow import keras
from tensorflow.python.framework.convert_to_constants import convert_variables_to_constants_v2
import numpy as np
#set resnet50_v2 as a example
model = tf.keras.applications.ResNet50V2()
full_model = tf.function(lambda x: model(x))
full_model = full_model.get_concrete_function(
tf.TensorSpec(model.inputs[0].shape, model.inputs[0].dtype))
# Get frozen ConcreteFunction
frozen_func = convert_variables_to_constants_v2(full_model)
frozen_func.graph.as_graph_def()
layers = [op.name for op in frozen_func.graph.get_operations()]
print("-" * 50)
print("Frozen model layers: ")
for layer in layers:
print(layer)
print("-" * 50)
print("Frozen model inputs: ")
print(frozen_func.inputs)
print("Frozen model outputs: ")
print(frozen_func.outputs)
# Save frozen graph from frozen ConcreteFunction to hard drive
tf.io.write_graph(graph_or_graph_def=frozen_func.graph,
logdir="./frozen_models",
name="frozen_graph.pb",
as_text=False)
ref: https://github.com/leimao/Frozen_Graph_TensorFlow/blob/master/TensorFlow_v2/train.py
I encountered as similar issue and found a solution below, which is
- originally posted by dkurt@github at https://github.com/opencv/opencv/issues/16879
- written for a MLP MNIST classification problem
- this is for tensorflow 2.x
from tensorflow.python.framework.convert_to_constants import convert_variables_to_constants_v2
from tensorflow.python.tools import optimize_for_inference_lib
loaded = tf.saved_model.load('models/mnist_test')
infer = loaded.signatures['serving_default']
f = tf.function(infer).get_concrete_function(
flatten_input=tf.TensorSpec(shape=[None, 28, 28, 1],
dtype=tf.float32)) # change this line for your own inputs
f2 = convert_variables_to_constants_v2(f)
graph_def = f2.graph.as_graph_def()
if optimize :
# Remove NoOp nodes
for i in reversed(range(len(graph_def.node))):
if graph_def.node[i].op == 'NoOp':
del graph_def.node[i]
for node in graph_def.node:
for i in reversed(range(len(node.input))):
if node.input[i][0] == '^':
del node.input[i]
# Parse graph's inputs/outputs
graph_inputs = [x.name.rsplit(':')[0] for x in frozen_func.inputs]
graph_outputs = [x.name.rsplit(':')[0] for x in frozen_func.outputs]
graph_def = optimize_for_inference_lib.optimize_for_inference(graph_def,
graph_inputs,
graph_outputs,
tf.float32.as_datatype_enum)
# Export frozen graph
with tf.io.gfile.GFile('optimized_graph.pb', 'wb') as f:
f.write(graph_def.SerializeToString())
