Weka Multilayer Perceptron训练耗时过长，求耗时评估及提速方案

Is this normal? #

Short answer: It could be normal, but it hinges on what the a hidden layer setting is doing. In Weka, setting hidden layers to a triggers automatic optimization of hidden node counts and layer structure—this isn’t just training a fixed MLP, it’s searching through dozens of potential architectures to find the best fit. That search adds massive computational overhead, especially with 3310 instances. On a standard consumer CPU, waiting a few hours isn’t out of the question, but if it’s stuck for 6+ hours with zero progress, you might want to check for memory bottlenecks or data anomalies.

Estimated completion time? #

There’s no one-size-fits-all answer—it varies drastically based on:

• Your hardware (CPU core count, clock speed, available RAM)
• The range of hidden layer architectures the algorithm is testing
• How quickly each candidate model converges
• Data traits (scaling, noise, class balance)

As a rough rule of thumb: If a fixed-structure MLP (e.g., 5 hidden nodes) takes 5-10 minutes to train on your setup, the auto-optimization (a) could take 10-100x longer—anywhere from an hour to a full day.

Speed-up tips #

Here are actionable steps to get your model training faster:

• Drop the auto-hidden-layer setting: Replace a with a fixed number of hidden nodes using a rule-of-thumb, like (number of input attributes + number of output classes)/2 (for 5 inputs, that’s 3-8 nodes). This eliminates the architecture search entirely, which is the biggest time drain. Start with 5—you can tune it later if needed.
• Preprocess your data: MLPs are hyper-sensitive to feature scaling. Normalize all numeric attributes to the [0,1] range or standardize them to mean 0, std dev 1 using Weka’s Normalize or Standardize filter. This helps backpropagation converge much faster.
• Tweak training parameters:
  • Lower the maximum iterations (maxIterations) from the default (often 1000) to a reasonable value like 500—you can check if the error plateaus early and stop training manually if needed.
  • Switch to QuickPropagation instead of standard backpropagation—it’s a faster-converging algorithm built into Weka’s MLP.
  • Adjust the learning rate (start with 0.1) and momentum (0.9 is a safe default) to speed up convergence without destabilizing the model.
• Optimize hardware and JVM settings:
  • Launch Weka with increased memory allocation using JVM flags like -Xmx8g (replace 8g with half your system’s RAM) to avoid memory swapping, which cripples performance.
  • Enable parallel garbage collection with -XX:+UseParallelGC to reduce training pauses caused by memory cleanup.
• Test with a smaller dataset first: Take a 10% sample of your 3310 instances and run training to gauge baseline time. This helps you estimate full training duration and catch issues early.
• Check for data problems: Ensure there are no missing values, extreme outliers, or heavily imbalanced classes—all of these can make the model struggle to converge, dragging out training time.

内容的提问来源于stack exchange，提问作者J.Extra2