Python Guide


We hereby introduce a new set of Keras-Style API based on Keras 1.2.2 in BigDL for the sake of user-friendliness. Users, especially those familiar with Keras, are recommended to use the new API to create a BigDL model and train, evaluate or tune it in a distributed fashion.

To define a model in Python using the Keras-Style API, now one just need to import the following packages:

from bigdl.nn.keras.topology import *
from bigdl.nn.keras.layer import *

One of the highlighted features with regard to the new API is shape inference. Users only need to specify the input shape (a shape tuple excluding batch dimension, for example, input_shape=(3, 4) for 3D input) for the first layer of a model and for the remaining layers, the input dimension will be automatically inferred.

Define a model

You can define a model either using Sequential API or Functional API. Remember to specify the input shape for the first layer.

After creating a model, you can call the following methods:


See here on how to train, predict or evaluate a defined model.

Sequential API

The model is described as a linear stack of layers in the Sequential API. Layers can be added into the Sequential container one by one and the order of the layers in the model will be the same as the insertion order.

To create a sequential container:


Example code to create a sequential model:

from bigdl.nn.keras.topology import Sequential
from bigdl.nn.keras.layer import Dense, Activation

model = Sequential()
model.add(Dense(32, input_shape=(128, )))

Functional API

The model is described as a graph in the Functional API. It is more convenient than the Sequential API when defining some complex model (for example, a model with multiple outputs).

To create an input node:

Input(shape=None, name=None)


To create a graph container:

Model(input, output)


To merge a list of input nodes (NOT layers), following some merge mode in the Functional API:

merge(inputs, mode="sum", concat_axis=-1) # This will return an output NODE.


Example code to create a graph model:

from bigdl.nn.keras.topology import Model
from bigdl.nn.keras.layer import Input, Dense, merge

# instantiate input nodes
input1 = Input(shape=(8, )) 
input2 = Input(shape=(6, ))
# pass an input node into a layer and get an output node
dense1 = Dense(10)(input1)
dense2 = Dense(10)(input2)
# merge two nodes following some merge mode
output = merge([dense1, dense2], mode="sum")
# create a graph container
model = Model([input1, input2], output)


See here for all the available layers for the new set of Keras-Style API.

To set the name of a layer, you can either call set_name(name) or alternatively specify the argument name in the constructor when creating a layer.

LeNet Example

Here we adopt our Keras-Style API to define a LeNet CNN model to be trained on the MNIST dataset:

from bigdl.nn.keras.topology import Sequential
from bigdl.nn.keras.layer import *

model = Sequential()
model.add(Reshape((1, 28, 28), input_shape=(28, 28, 1)))
model.add(Convolution2D(6, 5, 5, activation="tanh", name="conv1_5x5"))
model.add(Convolution2D(12, 5, 5, activation="tanh", name="conv2_5x5"))
model.add(Dense(100, activation="tanh", name="fc1"))
model.add(Dense(10, activation="softmax", name="fc2"))

model.get_input_shape() # (None, 28, 28, 1)
model.get_output_shape() # (None, 10)

See here for detailed introduction of LeNet, the full example code and running instructions.

Keras Code Support

If you have an existing piece of Keras code for a model definition, without installing Keras, you can directly migrate the code to construct a BigDL model by just replacing Keras import lines with:

from bigdl.nn.keras.topology import *
from bigdl.nn.keras.layer import *

and making modifications subject to the following limitations:

  1. The Keras version we support and test is Keras 1.2.2 with TensorFlow backend.

  2. There exist some arguments supported in Keras layers but not supported in BigDL for now. See here for the full list of unsupported layer arguments. Also, currently we haven't supported self-defined Keras layers or Lambda layers.

  3. The default dim_ordering in BigDL is th (Channel First, channel_axis=1).

  4. Keras backend related code needs to be deleted or refactored appropriately.

  5. Code involving Keras utility functions or loading weights from HDF5 files should be removed.

Remark: We have tested for migrating Keras code definition of VGG16, VGG19, ResNet50 and InceptionV3 into BigDL.