TensorFlow Keras中model.predict()预测结果异常问题求助

Hey there, let's break down why your model's predictions are all identical and work through the fixes step by step.

1. First: Make Sure Your Input Shape Matches the Model's Expectations #

Your model's input layer is defined with input_dim=w (and based on your test samples, w is 6), which means the model expects inputs in the shape (None, 6) — a 2D array where each row is a sample with 6 features.

Common Input Mistakes & Fixes: #

• When you grab a single sample like b = X[11], first check its shape:

  print("Single sample shape:", b.shape)
  

  • If the output is (6,) (a 1D array): Convert it to a 2D array with shape (1,6) using reshape (this works reliably no matter how many features you have):

    input_data = b.reshape(1, -1)  # The -1 lets Python auto-calculate the dimension size
    n = model.predict(input_data, verbose=0)
    

  • If the output is (1,6) (already 2D): Pass it directly to model.predict(b) — don’t wrap it in np.array([b]), since that would create a 3D array (1,1,6) and trigger the shape mismatch warning you saw earlier.

2. The Core Issue: Mismatched Feature Scales #

Looking at your test samples, your features have wildly different value ranges:

• The first 5 features are between 0.004 and 0.45
• The last feature ranges from 8.055 all the way up to 526.18

Neural networks are extremely sensitive to input scale. Without normalization/standardization, the model can’t learn effectively — it either suffers from vanishing/exploding gradients, or gets dominated by the large-scale feature, leading to identical predictions for all inputs.

Fix: Normalize Your Data Before Training #

Use a scaler from Scikit-learn to standardize your features (always fit the scaler only on training data to avoid data leakage):

from sklearn.preprocessing import StandardScaler

# Initialize scaler and fit it to training data
scaler = StandardScaler()
X_train_scaled = scaler.fit_transform(X[train])
X_test_scaled = scaler.transform(X[test])

# Train with scaled data
history = model.fit(X_train_scaled, Y[train], epochs=epoc, batch_size=bs)

# When predicting, scale the input with the same scaler
input_data_scaled = scaler.transform(np.array([b]).reshape(1, -1))
n = model.predict(input_data_scaled)

3. Verify Your Model Is Actually Learning #

Check the training loss history to confirm the model is improving:

print("Training loss over epochs:", history.history['loss'])

If the loss drops steadily and stabilizes, the model is training correctly. If it stays flat or barely changes, you may need to:

• Adjust the learning rate (try 1e-3 or 1e-4 instead of your current lr)
• Tweak the network structure (enable the commented-out hidden layer, or adjust the size of lw1)
• Increase the number of training epochs (epoc)

Wrap-Up #

Start with normalizing your data — this is the most likely cause of your constant predictions. Then double-check your input shape matches what the model expects. After these fixes, your model.predict() should return varied, reasonable results.

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