LSTM代码中num_units参数含义疑问：是隐藏状态数还是堆叠单元数

Great question—this is such a common point of confusion when getting started with recurrent networks, so let’s break it down clearly:

• The 15 in LSTM(15, return_sequences=True) refers to the dimension size of the hidden state (h_t) for each LSTM unit in the layer—it is NOT 15 separate LSTM units stacked at the same time step.

Here’s a more concrete breakdown to solidify the idea:

1. Hidden State Basics: Every LSTM unit in the layer calculates a hidden state vector at each time step. The num_units parameter defines how many elements are in this vector. In your example, every time step’s hidden state is a 15-dimensional vector.
2. Input/Output Dimension Mapping: When your input is a T×D matrix (T time steps, D features per time step), setting return_sequences=True tells the layer to output the hidden state from every time step. Since each hidden state is 15-dimensional, you end up with a T×15 output matrix—one 15D vector per time step.
3. Stacked LSTMs vs. Hidden State Dimension: If you wanted 15 distinct LSTM layers stacked on top of each other (not the same as hidden state size), you’d chain multiple LSTM layers like this:

   x = LSTM(15, return_sequences=True)(x)
   x = LSTM(15, return_sequences=True)(x)
   # ... repeat 13 more times to get 15 total stacked layers
   

   This is entirely different from the single LSTM(15) layer you’re asking about.

To make it even more tangible: if your input is a 10×8 matrix (10 time steps, 8 features each), passing it through LSTM(15, return_sequences=True) will output a 10×15 matrix—10 rows (one per time step), each containing 15 values that represent that time step’s hidden state.

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