TensorFlow Keras returning multiple predictions while expecting one

Question

I'm learning TensorFlow and LSTM and I'm wondering why my prediction output has multiple values when I'm training it to return one. My goal is to get a single value between 0 and 1 after training it with arrays for sentiment analysis.

The training input data looks like:

[[59, 21, ... 118, 194], ... [12, 110, ... 231, 127]]

All input arrays are of the same length padded with 0. The training target data looks like:

[1.0, 0.5, 0.0, 1.0, 0.0 ...]

Model:

model = Sequential()
model.add(Embedding(input_length, 64, mask_zero=True))
model.add(LSTM(100))
model.add(Dense(1, activation=tf.nn.sigmoid))

Why does the prediction seem to evaluate each individual value at a time instead of an array as a whole?

model.predict([192])
# Returns [[0.5491102]]
model.predict([192, 25])
# Returns [[0.5491102, 0.4923803]]
model.predict([192, 25, 651])
# Returns [[0.5491102, 0.4923803, 0.53853387]]

I don't want to take an average of the output because the relationships between the values in the input arrays matter for sentiment analysis. If I'm training to predict a single value I'm not understanding why a single value isn't output. I'm new to TensorFlow, Keras, and layered neural networks, so I'm sure I'm missing something obvious.

Have you tried model.predict([[192, 25]])?

pouyan
– pouyan

2018-12-30 16:15:14 +00:00
Commented Dec 30, 2018 at 16:15 — pouyan
– pouyan, Commented Dec 30, 2018 at 16:15
@pooyan Yes, same output.

Matt S
– Matt S

2018-12-30 16:18:21 +00:00
Commented Dec 30, 2018 at 16:18 — Matt S
– Matt S, Commented Dec 30, 2018 at 16:18

today · Accepted Answer · 2018-12-30 16:29:16Z

8

When you write:

model.predict([192, 25, 651])

it is if you are giving the model three input samples and therefore in return you would get three outputs, one for each input sample. Instead, if by [192, 25, 651] you really mean one input sample, then you wrap it in two lists:

model.predict([[[192, 25, 651]]])

The reason: the most outer list corresponds to the list of all the input data for all the input layers of the model, which is one here. The second list corresponds to the data for the first (and only) input layer and the third list corresponds the one input sample. That's the case with list inputs, since multi-input (and multi-output) Keras models should take a list of input arrays as input. One better way is to use a numpy array instead:

model.predict(np.array([[192, 25, 651]]))

np.array([[192, 25, 651]]) has a shape of (1,3) which means one sample of length 3.

edited Dec 30, 2018 at 16:29

answered Dec 30, 2018 at 16:19

today

33.6k8 gold badges100 silver badges122 bronze badges

Sign up to request clarification or add additional context in comments.

4 Comments

Matt S Over a year ago

I get the same results.

today Over a year ago

@MattS Put it in another list. I am editing my answer.

today Over a year ago

@MattS Edited my answer.

Matt S Over a year ago

WOW that was it! Makes sense. Thank you!

Collectives™ on Stack Overflow

TensorFlow Keras returning multiple predictions while expecting one

1 Answer 1

4 Comments

Your Answer

Linked

Hot Network Questions

Collectives™ on Stack Overflow

1 Answer 1

4 Comments

Your Answer

Sign up or log in

Post as a guest

Linked

Related