All-in-one 的Serving分析

export_func.export(model, sess, signature_name=mission, version=fold + 1)

def export(model, sess, signature_name, export_path=root_path + '/all_in_one/demo/exported_models/', version=1):
    # export path
    export_path = os.path.join(os.path.realpath(export_path), signature_name, str(version))
    print('Exporting trained model to {} ...'.format(export_path))

    builder = tf.saved_model.builder.SavedModelBuilder(export_path)
    # Build the signature_def_map.
    classification_w = tf.saved_model.utils.build_tensor_info(model.w)
    # classification_is_training = tf.saved_model.utils.build_tensor_info(model.is_training)
    classification_dropout_keep_prob_mlp = tf.saved_model.utils.build_tensor_info(
        model.dropout_keep_prob_mlp)
    # score
    classification_outputs_scores = tf.saved_model.utils.build_tensor_info(model.y)

    classification_signature = tf.saved_model.signature_def_utils.build_signature_def(
        inputs={tf.saved_model.signature_constants.CLASSIFY_INPUTS: classification_w},
        outputs={
            tf.saved_model.signature_constants.CLASSIFY_OUTPUT_SCORES:
            classification_outputs_scores
        },
        method_name=tf.saved_model.signature_constants.CLASSIFY_METHOD_NAME)  # 'tensorflow/serving/classify'

    prediction_signature = tf.saved_model.signature_def_utils.build_signature_def(
        inputs={'input_plh': classification_w, 'dropout_keep_prob_mlp':
                classification_dropout_keep_prob_mlp,
                # 'is_training': classification_is_training
                },
        outputs={'scores': classification_outputs_scores},
        method_name=tf.saved_model.signature_constants.PREDICT_METHOD_NAME)  # 'tensorflow/serving/predict'
    builder.add_meta_graph_and_variables(
        sess, [tf.saved_model.tag_constants.SERVING],
        signature_def_map={
            signature_name: prediction_signature,
            tf.saved_model.signature_constants.DEFAULT_SERVING_SIGNATURE_DEF_KEY: classification_signature,
        })
    builder.save()

在signature_def_map中定义了两个，一个是自己设计的别名，一个是默认的。

定义一个解析类。

model_name 是启动服务时明确的model_name

signature_name是在signature_def_map中自己设计的别名对应的输入输出之类的。

def classify(self,  sents):
        self.sents=self.sents2id(sents)
        hostport = '192.168.31.186:6000'
        # grpc
        host, port = hostport.split(':')
        channel = implementations.insecure_channel(host, int(port))
        stub = prediction_service_pb2.beta_create_PredictionService_stub(channel)
        # build request
        request = predict_pb2.PredictRequest()
        request.model_spec.name = self.model_name
        request.model_spec.signature_name = self.signature_name
        request.inputs['input_plh'].CopyFrom(
            tf.contrib.util.make_tensor_proto(self.sents, dtype=tf.int32))
        request.inputs['dropout_keep_prob_mlp'].CopyFrom(
            tf.contrib.util.make_tensor_proto(1.0, dtype=tf.float32))
        model_result = stub.Predict(request, 60.0)
        model_result = np.array(model_result.outputs['scores'].float_val)
        model_result = [model_result.tolist()][0]
        index, _ =max(enumerate(model_result), key=operator.itemgetter(1))
        if index>0:
            label = self.label_dict[index-1]
        else:
            label = ""
        # print("index:{}\tlabel:{}".format(index, label))
        if self.encode == "part" :
            if label:
                label=self.part[label]
            else:
                label = "凌晨"
        if self.encode == "type" :
            if label:
                label=self.type[label]
            else:
                label = "录像"
        if self.encode == "door" and label:
            label = self.gate[label]

        return label