Predictions always equal 1.0


#1

Hi,

I try this tutorial : Diabetes Prediction with Deep Learning Studio: A Different approach towards Deep Learning

The problem is the prediction are always 1.0 :

The training plots are always flat :

not like the one from the blog.

I use the database and the model provided here : github

The only thing I change is to specifiy that the data named ‘outcome’ is an outputPort0 and not an InputPort0

How can I solve this problem?

I try with both Windows and Linux version (no GPU). I’ve the last DLS version (I’ve just downloaded it).

Best regards,

Philippe


#2

Hi as it can be clearly seen from the graph that your model is not learning. Can you share the model that you have used as in the blog you can see that I have build a very basic model in which at the last layer I have used sigmoid function so your value should be in float. So please share the snapshot of the model used and the hyperpameters along with the Data tab.

Regards
Rajat


#3

Hi Rajat,

Thank you for your reply.

I used the model from your Github repository (https://github.com/Rajat2712/Deep-Learning-Studio/tree/master/Pima-Indians-Diabetes) and I realize that the model uses a Softmax activation function instead of a Sigmoïd one.

Here’s your config.yaml file :

. - args: {activation: softmax, output_dim: ‘1’}
class: Dense
name: Dense_55
x: 397
y: 297

If I change Softmax to Sigmoid, it works. I get the following results :

image

Regards,

Philippe


#4

Ok if you want to use softmax function instead of sigmoid then it must be a classification problem. Softmax function converts the probability sum to 1 since you have only 1 value as output its value is 1.
If you want to use softmax then change the datatype of the output variable to categorical and the output dimension of the last layer to 2 from 1.

Regards
Rajat