TensorFlow中DNNLinearCombinedRegressor回归模型损失过高求助

That’s definitely odd—with your target labels ranging from 0-40, an MSE over 1000 means your model isn’t learning anything meaningful from the data. Let’s break down the most likely issues and fixes step by step:

1. Critical Data Mix-Up (The Most Obvious Red Flag) #

Looking at your code, you’re training the model on test_data instead of a dedicated training dataset! This is a huge mistake—if you’re using held-out test data for training, the model can’t generalize, and it’ll either overfit immediately or fail to learn patterns entirely.

Fix:
Replace test_data in the model.train call with your actual training dataset:

def main():
    shutil.rmtree(modeldir, ignore_errors=True)  # Add ignore_errors to avoid missing directory errors
    model = build_estimator(modeldir)
    # Use train_data instead of test_data here!
    model.train(input_function = lambda: input_function(200, train_data))

2. Input Function Issues #

Your input function is the backbone of data feeding—if it’s mishandling features or labels, the model will never converge. Check these points:

• Feature-Label Separation: Ensure your input_function correctly splits features and the target label (returning (features, labels) instead of mixing them together).
• Normalize Continuous Features: DNNs are extremely sensitive to unnormalized numeric data. If you have continuous features (e.g., age, income), scale them to a small range (0-1 with MinMaxScaler or mean=0/variance=1 with StandardScaler).
• Batch Size & Shuffling: Add shuffling for training data and set a reasonable batch size (64-128 works for most cases) to stabilize training.

Example Corrected Input Function:

from sklearn.preprocessing import MinMaxScaler

def input_function(num_epochs, data_set, is_training=True):
    # Split features and target label
    features = data_set.drop("your_target_column", axis=1)
    labels = data_set["your_target_column"]
    
    # Normalize continuous features
    continuous_cols = ["col1", "col2", "col3"]  # Replace with your actual continuous columns
    scaler = MinMaxScaler()
    features[continuous_cols] = scaler.fit_transform(features[continuous_cols])
    
    # Build TensorFlow Dataset
    dataset = tf.data.Dataset.from_tensor_slices((dict(features), labels))
    if is_training:
        dataset = dataset.shuffle(buffer_size=len(data_set)).repeat(num_epochs).batch(64)
    else:
        dataset = dataset.batch(64)
    return dataset

3. Feature Column Misconfiguration #

Incorrect feature column setup is another common culprit for poor regression performance:

• Wide Columns: For categorical features, use categorical_column_with_vocabulary_list (if you know all possible values) or categorical_column_with_hash_bucket (for high-cardinality features). Avoid adding continuous features to wide columns unless you’re binning them.
• Deep Columns: Wrap continuous features in numeric_column, and convert categorical features to embedding_column (use a reasonable embedding dimension—min(600, round(category_cardinality**0.25)) is a safe rule of thumb).
• No Label Leakage: Double-check that your target column isn’t accidentally included in wide_columns or deep_columns.

4. Model Architecture & Hyperparameters #

Your current hidden units [1024, 512, 256] might be overkill for a target range of 0-40—too much model capacity can lead to overfitting or gradient instability. Try these adjustments:

• Shrink DNN Size: Start with smaller hidden units like [256, 128, 64] and scale up only if needed.
• Explicit Optimizers: Set custom optimizers with lower learning rates to stabilize training:

  return tf.estimator.DNNLinearCombinedRegressor(
      model_dir = model_dir,
      linear_feature_columns = wide_columns,
      dnn_feature_columns = deep_columns,
      dnn_hidden_units = [256, 128, 64],
      linear_optimizer=tf.keras.optimizers.Ftrl(learning_rate=0.01),
      dnn_optimizer=tf.keras.optimizers.Adam(learning_rate=0.001)
  )
  

• Check Logs: Enable TensorFlow logging to track loss trends:

  tf.logging.set_verbosity(tf.logging.INFO)
  

  If loss stays flat from the start, the model isn’t learning—this usually points to input data or feature column issues.

5. Validate Target Data #

Finally, confirm your target labels are clean:

• No missing values or outliers (e.g., values outside 0-40)
• Labels are numeric (not string-encoded)
• No class imbalance (though less critical for regression)

内容的提问来源于stack exchange，提问作者guru