# Keras import model gave different prediction results

**URL:** <https://community.konduit.ai/t/keras-import-model-gave-different-prediction-results/166>\
**Category:** DL4J\
**Created:** [February 24, 2020, 9:17pm UTC](https://community.konduit.ai/t/keras-import-model-gave-different-prediction-results/166 "2020-02-24T21:17:28Z")\
**Posts on this page:** 2\
**Page:** 1

<div class="post-metadata">

**Author:** ![gocome](https://avatars.discourse-cdn.com/v4/letter/g/46a35a/32.png) [@gocome](https://community.konduit.ai/u/gocome)\
**Post date:** [February 24, 2020, 9:17pm UTC](https://community.konduit.ai/t/keras-import-model-gave-different-prediction-results/166/1 "2020-02-24T21:17:28Z")

</div>

I am trying KerasModelImport.importKerasSequentialModelAndWeights() to load a Keras model generated by Python Keras, but cannot get the same result from Python and Java.

```auto
# Python code:
input_length = 3
model = Sequential()
model.add(Embedding(7, 4, input_length=input_length))
model.add(Flatten())
model.add(Dense(1, activation='relu'))
model.compile(optimizer=SGD(lr=0.01, momentum=0.9), loss='mean_squared_error')

# model.save('.\\data\\simple_mlp_4.h5')
model = load_model('.\\data\\simple_mlp_4.h5')

data = [[1, 2, 3], [4, 3, 6]]
data = np.asarray(data, dtype=np.int32)

output = model.predict(data)
print(output)

```

```auto
// Java code:
final String modelFile = new File(dataLocalPath,"simple_mlp_4.h5").getAbsolutePath();
MultiLayerNetwork model = KerasModelImport.importKerasSequentialModelAndWeights(modelFile, true);

INDArray input = Nd4j.create(new float[]{1,2,3,4,3,6},new int[]{2,3}).castTo(DataType.INT);
INDArray output = model.output(input);

System.out.println(output);

```

Python code prints out [[0.04062647] [0.01502968]], while Java code prints out [0.0604, 0.0385]. They loaded the same model and weights, but produce the different results. Anyone can help?

---

<div class="post-metadata">

**Author:** ![treo](https://yyz1.discourse-cdn.com/flex035/user_avatar/community.konduit.ai/treo/32/47_2.png) [@treo](https://community.konduit.ai/u/treo)\
**Post date:** [February 26, 2020, 10:54am UTC](https://community.konduit.ai/t/keras-import-model-gave-different-prediction-results/166/2 "2020-02-26T10:54:16Z")

</div>

Unfortunately, you’ve run into a known issue (see [DL4J: Add CNN1D & RNN 'channels last' support (Keras import) · Issue #8441 · eclipse/deeplearning4j · GitHub](https://github.com/eclipse/deeplearning4j/issues/8441))

Depending on the complexity of your actual model, you will be able to work around it with something like the following. It replaces how the Flatten is handled after the model is imported:

```java
 // Replace automatically created preprocessor with a more appropriate one for your case
model.getLayerWiseConfigurations().getInputPreProcessors()
     .put(1, new KerasFlattenPermutingingRnnPreprocessor(
             (KerasFlattenRnnPreprocessor) model.getLayerWiseConfigurations().getInputPreProcess(1)));

```

with `KerasFlattenPermutingingRnnPreprocessor` defined like this:

```java
public class KerasFlattenPermutingingRnnPreprocessor extends KerasFlattenRnnPreprocessor {
    public KerasFlattenPermutingingRnnPreprocessor(KerasFlattenRnnPreprocessor wrapped) {
        super(wrapped.getDepth(), wrapped.getTsLength());
    }

    @Override
    public INDArray preProcess(INDArray input, int miniBatchSize, LayerWorkspaceMgr workspaceMgr) {
        return super.preProcess(input.permute(0, 2, 1), miniBatchSize, workspaceMgr);
    }
}

```