tf.keras.layers.Flatten() reshapes a multi-dimensional input into a single long 1D vector per example, without changing any values or the batch dimension.

Syntax

tf.keras.layers.Flatten()

Deep Dive Course

Flatten collapses every dimension of an input except the batch dimension into one single dimension, preserving the values and their order but discarding any multi-dimensional spatial structure. It's most commonly used as the bridge between convolutional/pooling layers, which output multi-dimensional feature maps, and Dense layers, which require a flat 1D vector of features per example — Flatten has no learned parameters and performs no computation, it purely rearranges an existing tensor's shape.

1Understanding tf.keras.layers.Flatten()

💡

Flatten produces no computation and no learned parameters — its output length is simply the product of all the non-batch dimensions of its input, which is exactly the number that determines how many parameters the very next Dense layer will need.

editor.html

import tensorflow as tf
from tensorflow.keras import layers

layer = layers.Flatten()
output = layer(tf.zeros([1, 13, 13, 8]))
print(output.shape)

localhost:3000

2Practical Example

Here is a real-world application of tf.keras.layers.Flatten() showing how it is used in production TensorFlow code.

editor.html

import tensorflow as tf
from tensorflow.keras import layers

layer = layers.Flatten()
x = tf.constant([[[1.0, 2.0], [3.0, 4.0]]])
print(layer(x).numpy())

localhost:3000

3Best Practices

Follow these guidelines when working with tf.keras.layers.Flatten():

1. Place Flatten immediately before the first Dense layer when transitioning from convolutional/pooling layers to a fully-connected classifier head

2. Check the flattened output size, the product of the preceding layer's non-batch dimensions, before adding a large Dense layer after it, since that size directly controls the next layer's parameter count

3. Consider GlobalAveragePooling2D as an alternative to Flatten when you want to avoid a huge Dense layer following a large feature map

⚠️

Tip: Flatten produces no computation and no learned parameters — its output length is simply the product of all the non-batch dimensions of its input, which is exactly the number that determines how many parameters the very next Dense layer will need.

editor.html

import tensorflow as tf
from tensorflow.keras import layers

layer = layers.Flatten()
output = layer(tf.zeros([1, 13, 13, 8]))
print(output.shape)

localhost:3000

Examples

Example 01Basic Usage

import tensorflow as tf
from tensorflow.keras import layers

layer = layers.Flatten()
output = layer(tf.zeros([1, 13, 13, 8]))
print(output.shape)

Example 02Advanced Example

import tensorflow as tf
from tensorflow.keras import layers

layer = layers.Flatten()
x = tf.constant([[[1.0, 2.0], [3.0, 4.0]]])
print(layer(x).numpy())

Best Practices

Place Flatten immediately before the first Dense layer when transitioning from convolutional/pooling layers to a fully-connected classifier head
Check the flattened output size, the product of the preceding layer's non-batch dimensions, before adding a large Dense layer after it, since that size directly controls the next layer's parameter count
Consider GlobalAveragePooling2D as an alternative to Flatten when you want to avoid a huge Dense layer following a large feature map

Interview Question

Why is Flatten often followed directly by a Dense layer, rather than another Conv2D layer, in a typical image classification model?▼

Hint: Think about what kind of structure Conv2D relies on in its input, and what Flatten does to that structure.

Conv2D relies fundamentally on the spatial structure of its input, sliding a small kernel across a preserved height and width to detect local patterns — that only makes sense on a genuinely multi-dimensional, spatially-organized tensor. Flatten deliberately destroys that spatial structure, collapsing height, width, and channels into one flat, unstructured vector, which means a Conv2D layer placed after it would have no meaningful spatial neighborhood left to slide a kernel across. A Dense layer, in contrast, doesn't rely on any spatial structure at all — it simply connects every input value to every output unit — making it the natural next step once Flatten has produced a plain feature vector, typically used as the classifier head that turns extracted features into final class scores.

Exercises

MediumPractice using tf.keras.layers.Flatten() in a real scenario.

View Solution

import tensorflow as tf
from tensorflow.keras import layers

layer = layers.Flatten()
output = layer(tf.zeros([1, 13, 13, 8]))
print(output.shape)

Frequently Asked Questions

Why is Flatten often followed directly by a Dense layer, rather than another Conv2D layer, in a typical image classification model?▼

Related Functions

tf-keras-layers-maxpooling2dtf-reshapemodel-summary