keras: Will Keras output non-binary (continuous) predictions for a classification problem?
My current model is compiled as
model.compile(loss='categorical_crossentropy', optimizer=sgd, metrics=['accuracy'])
and I get predictions like
[[ 1. 0. 0. 0.]]
[[ 0. 1. 0. 0.]]
Instead I’d like my class predictions to show a percentage
[[ 0.8 0.1 0.1 0.4]]
Elsewhere I was advised to use loss='mse', but this was discouraged by another user.
This issue does not seem addressed in the docs (I’d be glad to add to docs any insight you can give).
Any advice on what is best practice using Keras?
About this issue
- Original URL
- State: closed
- Created 8 years ago
- Reactions: 8
- Comments: 16 (1 by maintainers)
@mptorr I have the same problem.
model.predict()returns predictions (zeros and ones) and not the probability (although every document on Keras states the opposite). Evenmodel.predict_proba()is returning predictions.If you’re using
predict()to generate your predictions, you should already get probabilities (provided your last layer is a softmax activation), nothing to do with the loss function. Are you sure you’re usingpredict(), notpredict_classes()?Just change the data types to float32 and normalize it the following way then the results will give the probabilities Data = X_train.astype(‘float32’) Data /=255 … model.predict_proba(Data, verbose=1)
try normalization on your test set and you may get the probabilities.
In my case, model.predict() and model.predict_proba() both give probabilities. model.predict_classes() gives binary responses
@mbollmann I check the predict_classes.
I don’t use the function of
predict_classesand use thepredictfunction Then i don’t know what does thepredict_classesreturn. Does it be the format,like this.I also don’t know when use the function of
predict_classes? Thanks.