训练整数温度深度学习模型后，如何预测非整数温度？

Great question—this is a super common scenario in regression tasks where you can’t possibly train on every possible input value. Let’s walk through practical, actionable approaches to make this work:

1. Reframe Your Task as Regression (Not Classification) #

Chances are, if you’re only predicting integer temperatures right now, you might have set this up as a classification task. Instead, shift to regression:

• Train your model to output a continuous numerical value (not a discrete class label) for temperature.
• Use a regression-focused loss function like Mean Squared Error (MSE) or Mean Absolute Error (MAE) instead of cross-entropy.
• Swap out any classification-specific output layers (like softmax) for a linear activation layer, which lets the model predict any real number within a reasonable range.

This way, your model learns the underlying continuous relationship between your input features (e.g., sensor readings, environmental data) and temperature, so it naturally interpolates between the integer points you trained on.

2. Generate Synthetic Training Data via Interpolation #

If you have limited training data, you can safely create synthetic samples to fill in the gaps between your integer temperatures:

• Take pairs of existing samples (e.g., 2℃ and 3℃) and generate weighted averages of their input features to simulate what those features might look like at 2.5℃.
• Use the interpolated temperature (2.5℃) as the label for this synthetic sample.
• Just make sure this interpolation aligns with real-world physics—don’t generate data that’s physically impossible for your use case.

This gives your model more data to learn the smooth transition between temperature values.

3. Choose a Regression-Focused Model Architecture #

Stick with architectures that excel at continuous value prediction:

• Simple fully connected neural networks work great for most straightforward regression tasks.
• If your input has spatial patterns (like thermal images) use a CNN; if it’s time-series data (like sensor logs over time) use an RNN or Transformer.
• The key rule: no softmax at the output—use a linear layer to output your continuous temperature value.

4. Post-Processing Interpolation (Last Resort) #

If you absolutely can’t rework your model to do regression, you can add a post-processing step:

• Train your existing model to predict integer temperatures as before.
• Fit a simple regression model (linear or polynomial) on your model’s predictions and the corresponding integer temperature labels.
• When you need to predict 2.5℃, feed the input to your original model to get nearby integer predictions, then use the fitted regression model to interpolate between those values.

This is a workaround though—training a regression model directly will always be more accurate and cleaner.

Quick Pro Tip #

When evaluating your model, use regression metrics like MSE, MAE, or R² score instead of classification metrics like accuracy. These metrics will properly measure how well your model predicts continuous temperature values.

内容的提问来源于stack exchange，提问作者Wesa