functional_keras.py

'''Functional Keras is a more functional replacement for the Graph API.
'''

###################
# 2 LSTM branches #
###################
a = Input(input_shape=(10, 32))  # output is a TF/TH placeholder, augmented with Keras attributes
b = Input(input_shape=(10, 32))
encoded_a = LSTM(32)(a)  # output is a TF/TH tensor
encoded_b = LSTM(32)(b)
merged = merge([encoded_a, encoded_b], mode='concat')
decoded = RepeatVector(10)(merged)
decoded = LSTM(32, return_sequences=True)(decoded)

# this is a fully-featured Keras model, will all the goodies that come with those.
# this is made possible by Keras topology information stored in the tensors.
model = Model(input=[a, b], output=[decoded])
model.compile(optimizer=Adam(), loss='mse')
model.fit([x1, x2], y)

################
# Shared layer #
################
shared_lstm = LSTM(32)
a = Input(input_shape=(10, 32))
b = Input(input_shape=(10, 32))
encoded_a = shared_lstm(a)
encoded_b = shared_lstm(b)
merged = merge([encoded_a, encoded_b], mode='concat')
decoded = RepeatVector(10)(merged)
decoded = LSTM(32, return_sequences=True)(decoded)

##############################
# Insertion of arbitrary ops #
##############################
# NOTE: cannot do a = tf.sigmoid(a), because although 'a' is a valid tf tensor,
# it is 'augmented' with data that allows Keras to keep track of previous operations
# (thus making it possible to train a model)...
a = Input(input_shape=(10, 32))
a = Lambda(tf.sigmoid)(a)

model = Model(input=[a, b], output=[decoder])
model.compile(optimizer=Adam(), loss='mse')
model.fit([x1, x2], y)

Ok let us mock up how Layer would look like them. Do you already have an idea?

class Layer(object):
    def __init__(self, *args, *kwargs):
        ...
    def __call__(self, inputs):
        # make sure inputs is a list
        # call merge, if "not merge_mode or singleton" just pass inputs[0], else return new layer with merged values as output
        # set that merge as previous with set_previous?
        # Return "get_output" or another Layer?

Before, __call__ would return a tensor (either based theano or tensorflow) and Model would actually care to get_input and get_outputs

Let us investigate each case:

• Return get_output
  This would make graph design easier. Also the API becomes a graph annotator similar to Blocks or nngraph. Note that we would have no access to the layer properties in a case such as decoded = RepeatVector(10)([encoded_a, encoded_b], merge_mode='concat') the RepeatVector is gone and decoder is a tensor. This is probably the most minimalist and clean option. We could for example just compile K.function([a, b], [decoder])
• Return Layer
  To avoid losing much control, __call__ could return a Layer or a Container with the elements added so far. If it returns a layer or a container, we would have to keep get_output to actually get tensor elements.
  Also, returning an object Layer or Container gives us more friendliness if trying to support non symbolic backends in the future.

Unless I'm missing something, I'm more inclined to the second case where we return a Layer.

What do you guys think?

Before, call would return a tensor (either based theano or tensorflow) and Model would actually care to get_input and get_outputs

__call__ will still return a tensor. It will be, however, a tensor possessing some extra attributes making it possible to reconstruct the Keras model it just went through (independently of the backend the tensor type belongs to).

get_output will simply retrieve the input tensor, pass it to __call__, and return the result.

Thank you for this @fchollet! As another feather in the "there is only one API that you can converge to" cap, a colleague of mine who uses https://github.com/Lasagne tells me that they have a similar interface.

I think this is great @fchollet! My one concern (not sure if you addressed elsewhere, I only looked at this quickly) is that I do like the verbose named argument training/prediction with the Graph API, but I think it should be optional. Could something like

a = Input(input_shape=(10, 32), name='input_1')
b = Input(input_shape=(10, 32), name='input_2')
...
model = Model(input=[a, b], output={'y': decoded})
model.compile(optimizer=Adam(), loss='mse')
model.fit({'input_1': x1, 'input_2': x2}, {'y': y})

be possible? Not sure how feasible/desirable this is...

@lukedeo: I was thinking of unifying the Graph, Sequential and Model APIs, so that you could do:

model.compile(optimizer=Adam(), loss='mse')
model.compile(optimizer=Adam(), loss={'y': 'mse'})

model.fit({'input_1': x1, 'input_2': x2}, y)
model.fit({'input_1': x1, 'input_2': x2}, [y])
model.fit({'input_1': x1, 'input_2': x2}, {'y': y})
model.fit([x1, x2], y)
model.fit([x1, x2], [y])

It implies that inputs and outputs are ordered (by order of their definition in the code). Thoughts?

To keep backward compatibility, it should also be possible to do:

model.fit({'input_1': x1, 'input_2': x2, 'y': y}, None)

@fchollet: Very nice! Although the current Graph API is verbose, I liked that I could easily combine multiple Sequential models as inputs, e.g.:

conv1 = Sequential()  # first convnet
conv1.add(...)

conv2 = Sequential()  # second convnet, different architecture
conv2.add(...)

model = Graph()
model.add_input(name='conv1_input', input_shape=(3, 32, 32))
model.add_input(name='conv2_input', input_shape=(1, 64, 64))
model.add_node(conv1, name='conv1', input='conv1_input')
model.add_node(conv2, name='conv2', input='conv2_input')
# merge concat here

How would something like this translate to the new functional Graph API?

cannot do a = tf.sigmoid(a)

Though tf.sigmoid wont work, we could make K.sigmoid work right? K.sigmoid could just copy the said attributes of the input to the output tensor. You can also do a = Activation('sigmoid')(a). Lambda layer for every activation doesn't sound right.

Also, what about operators ? c = Dense(10)(a + b) and a = 0.2 * a ?
One option would be to write an abstract class K.Tensor that wraps both theano and tensorflow tensors, which would take care of all the attribute business.

One option would be to write an abstract class K.Tensor that wraps both theano and tensorflow tensors, which would take care of all the attribute business.

That seems like a good idea, I'll think about it a bit more.

How would something like this translate to the new functional Graph API?

Models would be callable, so you could do:

conv1 = Sequential()  # first convnet
conv1.add(...)

conv2 = Sequential()  # second convnet, different architecture
conv2.add(...)

a = Input(input_shape=(3, 32, 32))
b = Input(input_shape=(1, 64, 64))
conv1a = conv1(a)
conv2b = conv2(b)
merged = merge([conv1a, conv2b], mode='concat', axis=1)

@farizrahman4u: upon reflexion, that is not feasible, because we need to be able to construct a layer-level computation graph in Keras land. When using Lambda, then the TF/TH ops are associated to the Lambda layer. When using arbitrary ops outside of Lambda, we would have no layer for these ops to be attached to, and the entire concept of "model" (as a network of layers) would break down.

The way to solve this would be to associated "layers" to every op in the Keras backend, but I don't want to go that way. Let's keep it simple.

The inputs a and b in the first example get used twice: once to go into the two LSTM and then again when specifying the input in Model.

a = Input(input_shape=(10, 32))  
b = Input(input_shape=(10, 32))
encoded_a = LSTM(32)(a) 
encoded_b = LSTM(32)(b)
# skip some code here
model = Model(input=[a, b], output=[decoded])

This looks odd to me.

Would it be sensible to have the Model optionally figure out what the inputs are based on where they get used in the graph? In this case, it seems like it would be straightforward, but maybe that is not generally the case.

I was caught a little off-guard by

encoded_a = shared_lstm(a)
encoded_b = shared_lstm(b)

I expected the definition of a layer to happen in one place (i.e. the meaning of the first line doesn't change based on the second line in this example).

In this API, could I replace those two lines with something like

encoded_a, encoded_b = shared_lstm([a,b])

I expected the definition of a layer to happen in one place (i.e. the meaning of the first line doesn't change based on the second line in this example).

Which is definitely the case. The definition of the layer happens at shared_lstm = LSTM(32). From then on it acts as a function (of course it should only be called on tensors with compatible shapes).

Would it be sensible to have the Model optionally figure out what the inputs are based on where they get used in the graph?

It is possible, but I believe it is a better practice to have the user explicitly states the inputs in the model definition, since it allows us to raise an error message if there is a discrepancy between the user's view of the world and the actual graph. Because the user will need to know what inputs are required anyway, when passing data to the model. Might as well prevent issues as early as possible.

@fchollet

I have a few doubts regarding the new api:

• Masking : How are the masks of sequences passed around among different layers?
• The Graph model can be used as a query-able data structure when copying weights from a model to another model of different config. Will this be possible in the new api?

I still recommend the K.Tensor approach because:

• Can write arbitrary TH/TF expressions without loss of topology information. No confusion between Lambda and LambdaMerge and all those stuff.
• Mask can be made an attribute of the Tensor. This answers my first question. This way all mask stuff would be completely hidden from the user. Otherwise __call__ would have to return a tuple, (sequence, mask), then the user would have to reroute it to the mask arg of next layer:

(y, mask) = LSTM(10, return_sequences=True)(x, mask)
(z, mask) = LSTM(5)(y, mask)

which is not very interesting.

• It is not that complicated. A class with lot of operator overloads. All this is a lot of work anyway:)

One complication is that there would be a lot of dummy "op layers".

Masking : How are the masks of sequences passed around among different layers?

Like they were before. Some layers can generate a mask based on their input tensor and the previous mask. The mask is then propagated forward. If a layer that does not supports masking receives a non-None mask, it raises an error.

Importantly the new approach is more general than the previous one, so it will be possible for a multi-input layer to handle masking.

How it works in practice:

a = Input(shape)

# This creates a node in a graph linking a to b.
# the mask generated by Masking is stored inside the node.
b = Masking()(a)

# the lstm retrieves the node that b came from, and reads the mask from there
c = LSTM(32)(b)

The Graph model can be used as a query-able data structure when copying weights from a model to another model of different config. Will this be possible in the new api?

Yes. This is an important feature. You will still be able to iterate over the layers in a graph and query a layer by name.

a = Input(shape)
b = Dense(32, name='my_dense')(a)
c = Dense(32, name='output')(b)

model = Model(a, c)

# list of all layers in order of horizontal graph traversal.
# So for a sequential model it's just the ordered list of layers, starting with the input layer
model.layers

first_dense_instance = model.get_layer(name='my_dense')
first_dense_instance = model.get_layer(index=0)

Very interesting new API and less verbose, more readable and avoid a lot of input=X name=X, I like it!
What about the access of an intermediary layer like in a Siamese Network?