关于Keras中LSTM工作机制的技术问询及输出形状验证

Let's break down exactly why your model outputs a shape of (3,1) and walk through how the LSTM processes your input step-by-step.

First, Clarify the Key Parameters #

Your input is a sequence of 3 words, each represented as a 4-dimensional vector — so the input shape (ignoring batch size) is (3,4). Here's what each part of your LSTM layer does:

• LSTM(1): The 1 defines the dimension of the LSTM's hidden state output. Every time the LSTM processes a time step, it outputs a 1-dimensional vector representing its current state.
• return_sequences=True: This is the critical setting here. When enabled, the LSTM doesn't just return the hidden state from the final time step — it returns the hidden state generated at every single time step in the input sequence.

Why the Output Shape is (3,1) #

Let's map this directly to your example sequence (how are you, 3 time steps):

1. Time Step 1: The LSTM takes the first word's 4D vector, processes it, and outputs a 1D hidden state (h1).
2. Time Step 2: The LSTM takes the second word's 4D vector plus the hidden state h1 from the previous step, processes them together, and outputs a new 1D hidden state (h2).
3. Time Step 3: The LSTM takes the third word's 4D vector plus h2, processes them, and outputs h3.

Since return_sequences=True, the model stacks these three 1D states into a sequence: [h1, h2, h3], which gives you a shape of (3,1) (note: Keras actually shows (None, 3,1) in model.summary() — the None represents a variable batch size, but when feeding a single sample, it simplifies to (3,1)).

How This Fits Into a Translation Task #

In translation scenarios, this setup makes sense if you're doing word-level sequence-to-sequence mapping (e.g., generating a translated word for every input word). If you instead wanted a single "summary" vector of the entire input sequence to feed into a decoder LSTM (the standard encoder-decoder translation setup), you'd set return_sequences=False (the default). That would give you only the final hidden state h3, with a shape of (1) (or (batch_size,1) for batches).

Quick Test: Adjusting return_sequences #

If you modify your code to:

model.add(LSTM(1, input_shape=(3,4)))  # return_sequences=False by default

Your output shape would change to (1) (or (batch_size,1)), since only the final time step's hidden state is returned.

内容的提问来源于stack exchange，提问作者Pusheen_the_dev