run first example returns probability > 1

[brucedlg]

Tried the first example: - training first model, then image classification. Get following result with probability > 1 or < 0:

        class: "0", probability: 12.16896
        class: "2", probability: 2.88387
        class: "6", probability: 2.50789
        class: "8", probability: -2.6e+00
        class: "1", probability: -2.7e+00

Tried on master, v0.16.0 and v0.15.0. Same results.
It's run on Red Hat 8.3 with JDK 11.

What did I do wrong? Any pointers? Thanks!

Here is the complete log:

examples]$ ./gradlew jar   

BUILD SUCCESSFUL in 2s
6 actionable tasks: 1 executed, 5 up-to-date

 examples]$ ./gradlew run -Dmain=ai.djl.examples.training.TrainMnist

> Task :examples:run
[INFO ] - Training on: cpu().
[INFO ] - Load MXNet Engine Version 1.8.0 in 0.075 ms.
Training:    100% |\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588| Accuracy: 0.93, SoftmaxCrossEntropyLoss: 0.24, speed: 1165.17 items/sec
Validating:  100% |\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588|
[INFO ] - Epoch 1 finished.NG [55s]
[INFO ] - Train: Accuracy: 0.93, SoftmaxCrossEntropyLoss: 0.24
[INFO ] - Validate: Accuracy: 0.96, SoftmaxCrossEntropyLoss: 0.13
Training:    100% |\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588| Accuracy: 0.97, SoftmaxCrossEntropyLoss: 0.10, speed: 1700.95 items/sec
Validating:  100% |\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588|
[INFO ] - Epoch 2 finished.NG [1m 27s]
[INFO ] - Train: Accuracy: 0.97, SoftmaxCrossEntropyLoss: 0.10
[INFO ] - Validate: Accuracy: 0.97, SoftmaxCrossEntropyLoss: 0.10
[INFO ] - train P50: 19.250 ms, P90: 26.651 ms
[INFO ] - forward P50: 0.444 ms, P90: 0.728 ms
[INFO ] - training-metrics P50: 0.016 ms, P90: 0.025 ms
[INFO ] - backward P50: 0.739 ms, P90: 1.092 ms
[INFO ] - step P50: 1.827 ms, P90: 2.418 ms
[INFO ] - epoch P50: 52.175 s, P90: 52.175 s

examples]$ ./gradlew run -Dmain=ai.djl.examples.inference.ImageClassification

> Task :examples:run
[INFO ] - [
        class: "0", probability: 12.16896
        class: "2", probability: 2.88387
        class: "6", probability: 2.50789
        class: "8", probability: -2.6e+00
        class: "1", probability: -2.7e+00
]

BUILD SUCCESSFUL in 3s
37 actionable tasks: 2 executed, 35 up-to-date

[frankfliu]

I think we didn't apply softmax in the ImageClassificationTranslator.

The default value of applySoftmax was true in earlier version, and now it's false.

We have to explicitly set to true in the example.

[brucedlg]

It works as charm!
After set optApplySoftmax(true), the result makes sense now. Thanks a bunch!

[INFO ] - [
        class: "0", probability: 0.99984
        class: "2", probability: 0.00009
        class: "6", probability: 0.00006
        class: "8", probability: 3.7e-07
        class: "1", probability: 3.5e-07
]