GRU: ทางเลือกที่กระชับกว่า

The GRU is a leaner cousin of the LSTM. It chases the same goal, long memory, but with fewer gates and fewer parameters. ⚡

Unlike the LSTM, a GRU has no separate cell state. It keeps everything in a single hidden state that it updates each step.

A GRU uses only two gates: an update gate and a reset gate. That is one fewer than the LSTM, which trims compute and memory.

The update gate decides how much of the old hidden state to carry forward versus how much new information to let in.

The reset gate controls how much past memory feeds into the new candidate, letting the cell ignore stale context when needed.

Using the reset gate, the GRU builds a candidate hidden state, a fresh proposal for what this step's memory could be.

The new hidden state is a smooth blend: the update gate mixes the previous state with the candidate in one clean equation.

With one less gate and no cell state, a GRU has fewer parameters. It often trains faster and needs less data to fit well.

On many tasks a GRU matches LSTM accuracy. The smaller cell rarely hurts, so it is a strong default for sequence models.

Try a GRU first for speed and small datasets. Reach for an LSTM when very long dependencies demand its extra memory control.

In Keras swapping is trivial: replace the LSTM layer with a GRU layer and keep the rest of your model unchanged.

Check what sets a GRU apart from an LSTM.

A GRU trims the LSTM to two gates and one state. It is faster and leaner while delivering similar accuracy on most tasks. ✅