Stuck on an issue?

Lightrun Answers was designed to reduce the constant googling that comes with debugging 3rd party libraries. It collects links to all the places you might be looking at while hunting down a tough bug.

And, if you’re still stuck at the end, we’re happy to hop on a call to see how we can help out.

How to export model for Tensorflow Serving?

See original GitHub issue

Follow some tutorials of exporting model for Tensorflow Serving, I’ve come up below trial:

cfg_from_file('ctpn/text.yml')
config = tf.ConfigProto(allow_soft_placement=True)
with tf.Session(config=config) as sess:
        net = get_network("VGGnet_test")

        saver = tf.train.Saver()
        try:
            ckpt = tf.train.get_checkpoint_state(cfg.TEST.checkpoints_path)
            saver.restore(sess, ckpt.model_checkpoint_path)
        except:
            raise 'Missing pre-trained model: {}'.format(ckpt.model_checkpoint_path)

       # The main export trial is here
       #############################
        export_path = os.path.join(
            tf.compat.as_bytes('/tmp/ctpn'),
            tf.compat.as_bytes(str(1)))
        builder = tf.saved_model.builder.SavedModelBuilder(export_path)

        freezing_graph = sess.graph
        prediction_signature = tf.saved_model.signature_def_utils.predict_signature_def(
            inputs={'input': freezing_graph.get_tensor_by_name('Placeholder:0')},
            outputs={'output': freezing_graph.get_tensor_by_name('Placeholder_1:0')}
        )

        builder.add_meta_graph_and_variables(
            sess,
            [tf.saved_model.tag_constants.SERVING],
            signature_def_map={
                tf.saved_model.signature_constants.DEFAULT_SERVING_SIGNATURE_DEF_KEY: prediction_signature
            },
            clear_devices=True)

        builder.save()
        print('[INFO] Export SavedModel into {}'.format(export_path))
        #############################

With output of freezing_graph is:

Tensor("Placeholder:0", shape=(?, ?, ?, 3), dtype=float32)
Tensor("conv5_3/conv5_3:0", shape=(?, ?, ?, 512), dtype=float32)
Tensor("rpn_conv/3x3/rpn_conv/3x3:0", shape=(?, ?, ?, 512), dtype=float32)
Tensor("lstm_o/Reshape_2:0", shape=(?, ?, ?, 512), dtype=float32)
Tensor("lstm_o/Reshape_2:0", shape=(?, ?, ?, 512), dtype=float32)
Tensor("rpn_cls_score/Reshape_1:0", shape=(?, ?, ?, 20), dtype=float32)
Tensor("rpn_cls_prob:0", shape=(?, ?, ?, ?), dtype=float32)
Tensor("Reshape_2:0", shape=(?, ?, ?, 20), dtype=float32)
Tensor("rpn_bbox_pred/Reshape_1:0", shape=(?, ?, ?, 40), dtype=float32)
Tensor("Placeholder_1:0", shape=(?, 3), dtype=float32)

After I got /tmp/ctpn/1 exported model, I try to load into Tensorflow Serving server:

tensorflow_model_server --port=9000 --model_name=ctpn --model_base_path=/tmp/ctpn

But it came up an error:

Loading servable: {name: detector version: 1} failed: Not found: Op type not registered 'PyFunc' in binary running on [...]. Make sure the Op and Kernel are registered in the binary running in this process.

So there are 2 questions: