Keras predict工作原理咨询及商品量价训练模型predict异常问题

Hey there! Let's break this down clearly—first explaining how Keras' prediction methods work, then diving into your stock price/volume model scenario.

At their core, these methods run your trained model in inference mode—meaning all training-specific behaviors (like Dropout, BatchNormalization's training state, or random data augmentation) are turned off. Here's the nitty-gritty:

• Forward propagation only: The model uses the weights it learned during training to pass your input data through each layer, calculating the output predictions for every sample.
• Key differences between the two:
  • predict(): Handles any size of input by automatically splitting it into batches (using the batch_size you set during training, defaulting to 32). It combines results from all batches and returns a single array of predictions.
  • predict_on_batch(): Only processes one single batch of input. If your input doesn't match the expected batch shape (e.g., wrong number of features or samples), it'll throw an error right away.
• Output shape: This depends entirely on your model's output layer. For regression tasks like predicting stock prices (a continuous value), you probably have a single-neuron Dense layer, so predictions will be a 2D array shaped like (number_of_samples, 1).

You mentioned that predictions on training data (orange curve) and validation data (green curve) look good, but cut off with a "but..."—my guess is you're hitting issues when predicting unseen future data? That's super common in time series forecasting, even with simple models:

• Why training/validation fit well: Your model might just be memorizing local patterns (or even noise) in the historical data instead of learning generalizable relationships. If your input features (close price, volume) have simple lagged correlations with the target, the model can easily nail known data but fail when future trends shift.
• Potential pitfalls to check:
  • Data preprocessing mistakes: If you didn't normalize/standardize your data, or used validation set stats to normalize training data (you should only use training set mean/std for all splits), your predictions might be skewed.
  • Bad validation split: If you randomly split your data instead of using a time-based split (validation data must come after training data in time), your model is essentially cheating by seeing future data, making validation results look better than they are.
  • Limited feature set: Even with a simple model, adding time-based features (like moving averages, daily returns, or technical indicators) can help it learn more robust patterns instead of just fitting raw price/volume.

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