What is the role of "Flatten" in Keras?

Question

I am trying to understand the role of the Flatten function in Keras. Below is my code, which is a simple two-layer network. It takes in 2-dimensional data of shape (3, 2), and outputs 1-dimensional data of shape (1, 4):

model = Sequential()
model.add(Dense(16, input_shape=(3, 2)))
model.add(Activation('relu'))
model.add(Flatten())
model.add(Dense(4))
model.compile(loss='mean_squared_error', optimizer='SGD')

x = np.array([[[1, 2], [3, 4], [5, 6]]])

y = model.predict(x)

print y.shape

This prints out that y has shape (1, 4). However, if I remove the Flatten line, then it prints out that y has shape (1, 3, 4).

I don't understand this. From my understanding of neural networks, the model.add(Dense(16, input_shape=(3, 2))) function is creating a hidden fully-connected layer, with 16 nodes. Each of these nodes is connected to each of the 3x2 input elements. Therefore, the 16 nodes at the output of this first layer are already "flat". So, the output shape of the first layer should be (1, 16). Then, the second layer takes this as an input, and outputs data of shape (1, 4).

So if the output of the first layer is already "flat" and of shape (1, 16), why do I need to further flatten it?

Answer 1

If you read the Keras documentation entry for Dense, you will see that this call:

Dense(16, input_shape=(5,3))

would result in a Dense network with 3 inputs and 16 outputs which would be applied independently for each of 5 steps. So, if D(x) transforms 3 dimensional vector to 16-d vector, what you'll get as output from your layer would be a sequence of vectors: [D(x[0,:]), D(x[1,:]),..., D(x[4,:])] with shape (5, 16). In order to have the behavior you specify you may first Flatten your input to a 15-d vector and then apply Dense:

model = Sequential()
model.add(Flatten(input_shape=(3, 2)))
model.add(Dense(16))
model.add(Activation('relu'))
model.add(Dense(4))
model.compile(loss='mean_squared_error', optimizer='SGD')

EDIT: As some people struggled to understand - here you have an explaining image:

Answer 2

This is how Flatten works converting Matrix to single array.

Answer 3

_{short read:}

Flattening a tensor means to remove all of the dimensions except for one. This is exactly what the Flatten layer does.

_{long read:}

If we take the original model (with the Flatten layer) created in consideration we can get the following model summary:

Layer (type)                 Output Shape              Param #   
=================================================================
D16 (Dense)                  (None, 3, 16)             48        
_________________________________________________________________
A (Activation)               (None, 3, 16)             0         
_________________________________________________________________
F (Flatten)                  (None, 48)                0         
_________________________________________________________________
D4 (Dense)                   (None, 4)                 196       
=================================================================
Total params: 244
Trainable params: 244
Non-trainable params: 0

For this summary the next image will hopefully provide little more sense on the input and output sizes for each layer.

The output shape for the Flatten layer as you can read is (None, 48). Here is the tip. You should read it (1, 48) or (2, 48) or ... or (16, 48) ... or (32, 48), ...

In fact, None on that position means any batch size. For the inputs to recall, the first dimension means the batch size and the second means the number of input features.

The role of the Flatten layer in Keras is super simple:

A flatten operation on a tensor reshapes the tensor to have the shape that is equal to the number of elements contained in tensor non including the batch dimension.

---

Note: I used the model.summary() method to provide the output shape and parameter details.

Answer 4

I came across this recently, it certainly helped me understand: https://www.cs.ryerson.ca/~aharley/vis/conv/

So there's an input, a Conv2D, MaxPooling2D etc, the Flatten layers are at the end and show exactly how they are formed and how they go on to define the final classifications (0-9).

Answer 5

It is rule of thumb that the first layer in your network should be the same shape as your data. For example our data is 28x28 images, and 28 layers of 28 neurons would be infeasible, so it makes more sense to 'flatten' that 28,28 into a 784x1. Instead of wriitng all the code to handle that ourselves, we add the Flatten() layer at the begining, and when the arrays are loaded into the model later, they'll automatically be flattened for us.

Answer 6

Flattening is converting the data into a 1-dimensional array for inputting it to the next layer. We flatten the output of the convolutional layers to create a single long feature vector. In some architectures, e.g. CNN an image is better processed by a neural network if it is in 1D form rather than 2D.

Answer 7

Flatten make explicit how you serialize a multidimensional tensor (tipically the input one). This allows the mapping between the (flattened) input tensor and the first hidden layer. If the first hidden layer is "dense" each element of the (serialized) input tensor will be connected with each element of the hidden array. If you do not use Flatten, the way the input tensor is mapped onto the first hidden layer would be ambiguous.

Answer 8

Here I would like to present another alternative to Flatten function. This may help to understand what is going on internally. The alternative method adds three more code lines. Instead of using

#==========================================Build a Model
model = tf.keras.models.Sequential()

model.add(keras.layers.Flatten(input_shape=(28, 28, 3)))#reshapes to (2352)=28x28x3
model.add(layers.experimental.preprocessing.Rescaling(1./255))#normalize
model.add(keras.layers.Dense(128,activation=tf.nn.relu))
model.add(keras.layers.Dense(2,activation=tf.nn.softmax))

model.build()
model.summary()# summary of the model

we can use

    #==========================================Build a Model
    tensor = tf.keras.backend.placeholder(dtype=tf.float32, shape=(None, 28, 28, 3))
    
    model = tf.keras.models.Sequential()
    
    model.add(keras.layers.InputLayer(input_tensor=tensor))
    model.add(keras.layers.Reshape([2352]))
model.add(layers.experimental.preprocessing.Rescaling(1./255))#normalize
    model.add(keras.layers.Dense(128,activation=tf.nn.relu))
    model.add(keras.layers.Dense(2,activation=tf.nn.softmax))
    
    model.build()
    model.summary()# summary of the model

In the second case, we first create a tensor (using a placeholder) and then create an Input layer. After, we reshape the tensor to flat form. So basically,

Create tensor->Create InputLayer->Reshape == Flatten

Flatten is a convenient function, doing all this automatically. Of course both ways has its specific use cases. Keras provides enough flexibility to manipulate the way you want to create a model.

Answer 9

Keras flatten class is very important when you have to deal with multi-dimensional inputs such as image datasets. Keras.layers.flatten function flattens the multi-dimensional input tensors into a single dimension, so you can model your input layer and build your neural network model, then pass those data into every single neuron of the model effectively.

You can understand this easily with the fashion MNIST dataset. The images in this dataset are 28 * 28 pixels. Hence if you print the first image in python you can see a multi-dimensional array, which we really can't feed into the input layer of our Deep Neural Network.

print(train_images[0])

first image of fashion MNIST

To tackle this problem we can flatten the image data when feeding it into a neural network. We can do this by turning this multidimensional tensor into a one-dimensional array. In this flattened array now we have 784 elements (28 * 28). Then we can create out input layer with 784 neurons to handle each element of the incoming data.

We can do this all by using a single line of code, sort of...

keras.layers.flatten(input_shape=(28,28))

Answer 10

As the name suggests it just flattens out the input Tensor. A very good visual to understand this is given below. Please let me know if there is any confusion. Flatten Input Tensor