keras -> dl4j

dl4j.scala

  val model = KerasModelImport.importKerasSequentialModelAndWeights(
    "keras4j.json",
    "keras4j.h5")

  val (trackLength, vectorDim, groupSize) = (30, 2, 8)

  val o = Array.fill(trackLength * vectorDim * groupSize)(1.0)

  val ones = Nd4j.create(o, Array(1, groupSize, vectorDim, trackLength))

  val oneOut = model.output(ones)
  // Output is [0.23, 0.69, 0.08]

keras-model.py

groupSize = 8
vectorDim = 2
trackLength = 30

model = Sequential()
model.add(Convolution2D(32, 10, 2, border_mode='same', dim_ordering='tf', input_shape=(trackLength, vectorDim, groupSize)))
model.add(Activation('relu'))

model.add(Convolution2D(16, 5, 2, border_mode='same', dim_ordering='tf'))
model.add(Activation('relu'))
model.add(Flatten())
model.add(Dropout(0.2))

model.add(Dense(128))
model.add(Dropout(0.2))

model.add(Dense(3))
model.add(Activation('softmax'))

optim = RMSprop()
model.compile(loss='categorical_crossentropy', optimizer=optim)
model_json = model.to_json()

with open("keras4j.json", "w") as json_file:
    json_file.write(model_json)
# serialize weights to HDF5
model.save_weights("keras4j.h5")
print("Saved model to disk")

model.predict_proba(np.ones((1, trackLength, vectorDim, groupSize)))
# output is [ 0.51176351,  0.43943113,  0.04880537]