TensorFlow · Class

layers.GRUCell

tf.keras.layers.GRUCell implements one timestep of a gated recurrent unit. Wrap it with tf.keras.layers.RNN when a sequence-level recurrent layer is needed.

How it works

GRUCell performs one timestep of gated recurrent computation. It receives a batch of feature vectors and the previous state, applies input and recurrent transformations with configurable activations, dropout, and reset-gate ordering, then produces the timestep output and updated state.

Syntax

tf.keras.layers.GRUCell(units, activation='tanh', recurrent_activation='sigmoid', use_bias=True, kernel_initializer='glorot_uniform', recurrent_initializer='orthogonal', bias_initializer='zeros', kernel_regularizer=None, recurrent_regularizer=None, bias_regularizer=None, kernel_constraint=None, recurrent_constraint=None, bias_constraint=None, dropout=0.0, recurrent_dropout=0.0, reset_after=True, seed=None, **kwargs)

Parameters

Name	Type	Description
`units`	`int`	Number of units in the GRU state.
`activation`	`callable or string`	Activation applied to the candidate state; defaults to tanh, or no activation when None.
`recurrent_activation`	`callable or string`	Activation used for the recurrent gates; defaults to sigmoid.
`use_bias`	`bool`	Whether to include a bias vector.
`kernel_initializer`	`Initializer`	Initializer for the input kernel.
`recurrent_initializer`	`Initializer`	Initializer for the recurrent kernel.
`bias_initializer`	`Initializer`	Initializer for the bias.
`kernel_regularizer`	`Regularizer`	Regularizer for the input kernel.
`recurrent_regularizer`	`Regularizer`	Regularizer for the recurrent kernel.
`bias_regularizer`	`Regularizer`	Regularizer for the bias.
`kernel_constraint`	`Constraint`	Constraint for the input kernel.
`recurrent_constraint`	`Constraint`	Constraint for the recurrent kernel.

Returns

The cell returns the output for the current timestep and the updated recurrent state.

Key points

GRUCell is the recurrent cell used to implement a GRU layer.
Each invocation handles one timestep rather than an entire temporal sequence.
At call time, inputs contain one batch of feature vectors.
The state argument carries the preceding timestep's hidden representation.
dropout and recurrent_dropout specify fractions of input and recurrent transformations to drop.
reset_after chooses the reset-gate ordering used by the cell.
seed controls the randomness used by dropout.

Common mistakes

GRUCell is a single-timestep cell, not a complete sequence processor; use tf.keras.layers.RNN to iterate it over a sequence.
The training argument affects behavior only when input or recurrent dropout is enabled.
The previous state must have batch-by-units shape, matching the cell's units.

Example: Run a GRU cell through an RNN wrapper

import tensorflow as tf

cell = tf.keras.layers.GRUCell(4)
rnn = tf.keras.layers.RNN(cell, return_state=True)
inputs = tf.random.normal((2, 3, 5))
output, state = rnn(inputs)
print(output.shape, state.shape)

Wrap the cell in RNN to process a batch of three-step sequences and return the final output plus state.

When to use it

Build a customized GRU recurrence by configuring activations, initializers, regularization, constraints, or gate orderin

Custom recurrent computation

Use the cell inside tf.keras.layers.RNN to process sequences timestep by timestep.

Sequence processing

Feed one timestep at a time while carrying the returned state to the next call.

Stateful step-wise inference

Interview questions

QHow does GRUCell differ from tf.keras.layers.GRU?

GRUCell computes one timestep, whereas the GRU layer handles a complete sequence. To process multiple timesteps with the cell, place it inside tf.keras.layers.RNN.

Follow-up: What shape does the cell receive? It takes inputs shaped (batch, features) and a previous state shaped (batch, units).

QWhat does reset_after control?

The reset_after option selects whether the reset gate is applied after or before matrix multiplication. Its default is True, which is the cuDNN-compatible convention.

Follow-up: Which settings support the cuDNN implementation? The GRU documentation lists tanh and sigmoid activations, zero dropout, no unrolling, bias enabled, reset_after=True, suitable padding, and eager execution.

QWhat are the call inputs and state shapes?

The inputs argument is a 2D tensor with shape (batch, features), while states is a 2D tensor with shape (batch, units). The training flag selects training or inference behavior when dropout is configured.

Follow-up: Why wrap the cell in RNN? The wrapper supplies sequence iteration and manages recurrent state across timesteps.

tf.keras.layers.GRU
tf.keras.layers.RNN
tf.keras.layers.SimpleRNNCell
tf.keras.layers.LSTMCell

Source: tf.keras.layers.GRUCell | TensorFlow v2.16.1. Checked 2026-10-15.