Serialize native Keras model for Tensorflow Serving

Code adapted from discussion about this in Tensorflow Serving Issue 310, specifically the recipe suggested by @tspthomas.



In [1]:

    
from tensorflow.python.saved_model import builder as saved_model_builder
from tensorflow.python.saved_model import utils, tag_constants, signature_constants
from tensorflow.python.saved_model.signature_def_utils import build_signature_def, predict_signature_def
from tensorflow.contrib.session_bundle import exporter

import keras.backend as K
from keras.models import load_model
import os
import shutil

K.set_learning_phase(0)









    



Using TensorFlow backend.

Load model

The model we will use is the best model produced by this Keras model.



In [2]:

    
DATA_DIR = "../../data"
EXPORT_DIR = os.path.join(DATA_DIR, "tf-export")

MODEL_NAME = "keras-mnist-fcn"
MODEL_VERSION = 1

MODEL_BIN = os.path.join(DATA_DIR, "{:s}-best.h5".format(MODEL_NAME))
EXPORT_PATH = os.path.join(EXPORT_DIR, MODEL_NAME)



In [3]:

    
model = load_model(MODEL_BIN)

Export model

Resulting exported model should be as follows, under the export directory given by EXPORT_DIR.

.
└── keras-mnist-fcn
    └── 1
        ├── saved_model.pb
        └── variables
            ├── variables.data-00000-of-00001
            └── variables.index



In [4]:

    
shutil.rmtree(EXPORT_PATH, True)



In [5]:

    
full_export_path = os.path.join(EXPORT_PATH, str(MODEL_VERSION))
builder = saved_model_builder.SavedModelBuilder(full_export_path)
signature = predict_signature_def(inputs={"images": model.input},
                                  outputs={"scores": model.output})

with K.get_session() as sess:
    builder.add_meta_graph_and_variables(sess=sess,
                                         tags=[tag_constants.SERVING],
                                         signature_def_map={"predict": signature})
    builder.save()









    



INFO:tensorflow:No assets to save.
INFO:tensorflow:No assets to write.
INFO:tensorflow:SavedModel written to: ../../data/tf-export/keras-mnist-fcn/1/saved_model.pb



In [ ]: