how to split input in a Keras model

Asked Jul 27 '18 at 23:28

Active Jul 28 '18 at 07:21

Viewed 1,209 times

I'm trying to do something that's pretty simple in keras with no success. I have an input X with size (?, 1452, 1). All I want to do is split this input to a vector of 1450 and a vector of 2 and deal with them seperately in the network. I tried:

X1 = X[:, 1450:1452, :]

X2 = X[:, 0:1450, :]

and then do whatever I want. It compiles fine until it gets to the line where I create the model. I get an error saying that my Tensor object doesn't have any attribute called _keras_history even though it does. So i'm guessing keras converts X1 and X2 to a regular tensor. so, I tried using Lambda layers in the following way:

X1 = Lambda(lambda x: X[:, 0:1450, :], output_shape=(1450, 1))(X)

after using this, the network compiled and trained completely fine, and the only problem was saving the model to json/yaml. It gave me an error in the copy.deepcopy part saying:

TypeError: can't pickle _thread.lock objects. after researching online

I found out there's a problem in saving Lambda layers to json for some reason.

Question: Do you know of any way to split the input in a normal way that would allow me to save the model later? Thanks!

edited Jul 28 '18 at 07:21

aleksandrbel

1,422
3
20
38

asked Jul 27 '18 at 23:28

toxin9

Please provide more code with the place where an error occurs in the first step. And show stacktrace of an error if there is more information. – aleksandrbel Jul 27 '18 at 23:45
Can you use tf.slice? – sdcbr Jul 27 '18 at 23:58
Why not just have a model with two inputs? – Dr. Snoopy Jul 28 '18 at 07:45
tf.slice is the same as X[:, 1, :]... I wasn't aware there's an option to have a model with two inputs. I'll try. thanks! – toxin9 Jul 28 '18 at 16:00
After reading on feeding multiple inputs to my model and adjusting the code it seems to work. Thank you very much! had no idea the solution was so easy – toxin9 Jul 28 '18 at 16:08

how to split input in a Keras model

0 Answers0