[FIXED] I am getting "Requested tensor connection from unknown node: "keras_layer_input:0"". error while loading keras model

Issue

I had Saved model using

tf.keras.experimental.export_saved_model(model, export_path)

This model have custom layers and loss function.

Loading model using

import tensorflow as tf
import tensorflow_hub as hub
import keras
class training_model:
    def __init__(self):
        path_bce="D:\\nsfw\\training_model\\models\\bce_20210120_153631"
        path2="D:\\nsfw\\training_model\\models\\soft-f1_20210120_153631"
        self.graph = tf.Graph()
        with self.graph.as_default():
            self.session = tf.Session()
            with self.session.as_default() :
                self.reloaded =tf.keras.experimental.load_from_saved_model(path2, custom_objects={'KerasLayer':hub.KerasLayer})
                 
training_model=training_model()
img = keras.preprocessing.image.load_img(
    "0drqz7883ox51.jpg", target_size=(224, 224)
)

img_array = keras.preprocessing.image.img_to_array(img)
img_array = tf.expand_dims(img_array, 0) # Create a batch
with training_model.graph.as_default():
    with training_model.session.as_default():
        print(training_model.reloaded.predict(img_array,steps=1))

It is working fine if i remove Graph and Session but I want to serve this model with API.

Solution

You can try something like this.

        with self.graph1.as_default():
            self.face_graph = tf.compat.v1.GraphDef()
            fid = tf.io.gfile.GFile(self.facenet_model, "rb")
            serialized_graph = fid.read()
            self.face_graph.ParseFromString(serialized_graph)
            tf.import_graph_def(self.face_graph, name="")
            self.facenet_sess = tf.compat.v1.Session(graph=self.graph1)
            self.images_placeholder = self.graph1.get_tensor_by_name("input:0")
            self.embeddings = self.graph1.get_tensor_by_name("embeddings:0")

Answered By - Neel Gajjar