How to convert model from HDF5 format to pb?

Question

I am using TensorFlow v2.3.0. I would like to convert a model I saved as HDF5 with a .h5 extension to protocol buffers (.pb file). There are existing solutions that I have tried but do not work directly because they were written for TensorFlow v1.x.x. So I tried editing the code to make them compatible with TensorFlow v2.3.0 and I ended up with this code:

import tensorflow as tf
from tensorflow.keras.models import load_model
from tensorflow.compat.v1.keras.backend import get_session
from tensorflow.python.platform import gfile
from tensorflow.compat.v1 import global_variables
from tensorflow.compat.v1.graph_util import convert_variables_to_constants as c_to_c

model = load_model('models/model-v2.h5')

# print(model.summary())
Model: "functional_1"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
input_1 (InputLayer)         [(None, 150, 150, 3)]     0         
_________________________________________________________________
conv2d (Conv2D)              (None, 148, 148, 16)      448       
_________________________________________________________________
max_pooling2d (MaxPooling2D) (None, 74, 74, 16)        0         
_________________________________________________________________
conv2d_1 (Conv2D)            (None, 72, 72, 32)        4640      
_________________________________________________________________
max_pooling2d_1 (MaxPooling2 (None, 36, 36, 32)        0         
_________________________________________________________________
conv2d_2 (Conv2D)            (None, 34, 34, 64)        18496     
_________________________________________________________________
max_pooling2d_2 (MaxPooling2 (None, 17, 17, 64)        0         
_________________________________________________________________
flatten (Flatten)            (None, 18496)             0         
_________________________________________________________________
dense (Dense)                (None, 512)               9470464   
_________________________________________________________________
dropout (Dropout)            (None, 512)               0         
_________________________________________________________________
dense_1 (Dense)              (None, 1)                 513       
=================================================================
Total params: 9,494,561
Trainable params: 9,494,561
Non-trainable params: 0

# print(model.outputs)
# [<tf.Tensor 'dense_1/Sigmoid:0' shape=(None, 1) dtype=float32>]
# print(model.inputs)
# [<tf.Tensor 'input_1:0' shape=(None, 150, 150, 3) dtype=float32>]


def freeze_session(session, keep_var_names=None, output_names=None, clear_devices=True):
    graph = session.graph
    with graph.as_default():
        freeze_var_names = list(set(v.op.name for v in global_variables()).difference(keep_var_names or []))
        output_names = output_names or []
        output_names += [v.op.name for v in global_variables()]
        # Graph -> GraphDef ProtoBuf
        input_graph_def = graph.as_graph_def()
        if clear_devices:
            print(input_graph_def.node)
            for node in input_graph_def.node:
                print('Node',node)
                node.device = ""
        frozen_graph = c_to_c(session, input_graph_def, output_names, freeze_var_names)
        return frozen_graph

frozen_graph = freeze_session(tf.compat.v1.Session(),
                            output_names=[out.op.name for out in model.outputs])

# Save to model/model.pb
tf.io.write_graph(frozen_graph, "models", "model_v2.pb", as_text=False)

I am however encountering this error: AssertionError: dense_1/Sigmoid is not in graph. Apparently the graph_def is returning an empty list because nothing is being printed as I loop through the nodes. I would like to know how to correct this code or if there is a better alternative to converting a .h5 model to .pb

Answer 1

After a lot of googling, I found the correct(compatible with TensorFlow 2.x.x) code here: https://leimao.github.io/blog/Save-Load-Inference-From-TF2-Frozen-Graph/

# Convert Keras model to ConcreteFunction
full_model = tf.function(lambda x: model(x))
full_model = full_model.get_concrete_function(
    x=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="simple_frozen_graph.pb",
                  as_text=False)

The code used if from this file: https://github.com/leimao/Frozen_Graph_TensorFlow/blob/master/TensorFlow_v2/example_1.py

Answer 2

Try to use tf.saved_model.save in compatibility mode:

tf.keras.Model instances constructed from inputs and outputs already have a signature and so do not require a @tf.function decorator or a signatures argument. If neither are specified, the model's forward pass is exported.

x = input_layer.Input((4,), name="x")
y = core.Dense(5, name="out")(x)
model = training.Model(x, y)
tf.compat.v1.saved_model.save(model, '/tmp/saved_model/')
# The exported SavedModel takes "x" with shape [None, 4] and returns "out"
# with shape [None, 5]

P.S. Have not opportunity to test, let me know about result please if you going to try it.