2017-09-22 14:20
浏览 183


I'm trying to setup a classical MNIST challenge model with keras, then save the tensorflow graph and subsequently load it in Go and evaluate with some input. I've been following this article which supplies full code on github. Nils is using just tensorflow to setup the comp.graph but I would like to use keras. I managd to save the model the same way as he does


   model = Sequential()
model.add(Conv2D(32, kernel_size=(3, 3),
                 input_shape=(28,28,1), name="inputNode"))
model.add(Conv2D(64, (3, 3), activation='relu'))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Dense(128, activation='relu'))
model.add(Dense(10, activation='softmax', name="inferNode"))

which runs ok, trains and evaluates and then saving as posted above:

builder = tf.saved_model.builder.SavedModelBuilder("mnistmodel_my")
# GOLANG note that we must tag our model so that we can retrieve it at inference-time
builder.add_meta_graph_and_variables(sess, ["serve"])

Which I then try to evaluate as :

result, runErr := model.Session.Run(
            model.Graph.Operation("inputNode").Output(0): tensor,

In Go I follow the example but when evaluating, I get:

    panic: nil-Operation. If the Output was created with a Scope object, see Scope.Err() for details.

goroutine 1 [running]:, 0x0, 0x0, 0x0)
    /Users/air/go/src/ +0xbb, 0xc420047ef0, 0x1, 0x1, 0x0, 0x0, 0x0, 0xc4200723c8)
    /Users/air/go/src/ +0x22d*Session).Run(0xc420078060, 0xc42006e210, 0xc420047ef0, 0x1, 0x1, 0x0, 0x0, 0x0, 0x0, 0x0, ...)
    /Users/air/go/src/ +0x153
    /Users/air/PycharmProjects/GoTensor/custom.go:36 +0x341
exit status 2

Since it says nil-Operation I think I might have incorrectly labelled the nodes. But I don't know which other nodes should I then label?

Many thanks!!!

  • 写回答
  • 好问题 提建议
  • 追加酬金
  • 关注问题
  • 收藏
  • 邀请回答

1条回答 默认 最新

  • dpoppu4300 2017-09-22 16:23

    Your code should work fine. You're right about the cause of the nil-operation.

    You just have to find the complete node name of your "inputNode".

    From python, after your model definition, you can loop over the graph nodes and look for the complete name, in that way:

    for n in sess.graph.as_graph_def().node:
        if "inputNode" in

    Once you got the complete name, you can use it in your Go program.

    Also, I suggest you to use a more complete and easy to use wrapper around the tensorflow API: tfgo

    解决 无用
    打赏 举报

相关推荐 更多相似问题