基于LSTM/GRU的RNN回归：三维目标值的格式适配咨询

Great question! Let's break this down clearly based on your specific task—since you're mapping one entire IMU time sequence to a single 3-dimensional PAD value (Pleasure/Arousal/Dominance), you do NOT need to reshape your target data into a 3D format matching your input. Here's why, plus a concrete implementation approach:

Key Background: Task Type Matters #

Your setup is a sequence-to-vector regression task, not sequence-to-sequence. That means:

• Input: 3D tensor (batch_size, timesteps, imu_feature_count) (you already have this handled)
• Target: 2D tensor (batch_size, 3) — one 3-value PAD vector per time sequence in your batch.

This is the right shape because you want your model to learn to summarize the entire time series into a single output vector, not predict a PAD value at every time step.

Model Implementation Example (Keras/TensorFlow) #

Here's how to structure your model to match this target shape:

from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import LSTM, Dense

# Define your input parameters (adjust timesteps/imu_features to your data)
timesteps = 100  # Example: each IMU sequence has 100 time steps
imu_features = 6  # Example: IMU has 6 features (accel x/y/z, gyro x/y/z)

# Build the sequence-to-vector model
model = Sequential([
    # LSTM layer with return_sequences=False: outputs a single vector per batch sample
    LSTM(64, return_sequences=False, input_shape=(timesteps, imu_features)),
    # Optional dense layer for intermediate processing
    Dense(32, activation='relu'),
    # Final dense layer outputs 3 values for PAD
    Dense(3)
])

# Compile with regression-friendly loss/optimizer
model.compile(optimizer='adam', loss='mean_squared_error')

Why 3D Targets Are Unnecessary (And Harmful) For Your Task #

If you reshaped your targets to (batch_size, timesteps, 3), you'd be training the model for sequence-to-sequence regression—meaning it would predict a PAD value at every time step in the input sequence. That's not what you want here, since your problem states each entire sequence maps to one PAD vector.

Edge Case: If You Ever Need Time-Step PAD Predictions #

If you later decide to predict PAD values at individual time steps (instead of per-sequence), then you would:

• Reshape targets to 3D (batch_size, timesteps, 3)
• Set return_sequences=True in your LSTM/GRU layer
• Use a TimeDistributed(Dense(3)) layer to output predictions for every time step

Final Takeaway #

Stick with a 2D target shape (batch_size, 3) for your current task. This aligns perfectly with a sequence-to-vector LSTM/GRU architecture, which will learn to distill the entire IMU time series into your three PAD values.

内容的提问来源于stack exchange，提问作者Fawzi Rida