Keras predict_classes维度异常求助：机器学习新手的模型构建问题

Hey there! As someone starting out with Keras/TensorFlow in R, it's easy to mix up functions designed for different task types—let's figure out why predict_classes is throwing dimension errors and get your 4→11 vector model working right.

First, let's get to the core of the problem: predict_classes is built for single-label classification tasks (where each sample maps to one category index, like 0-10 for 11 classes). But your model outputs an 11-dimensional vector of values between 0-1 (with sum <1)—that's either a multi-label classification task or a regression task predicting a constrained vector. predict_classes has no idea how to handle that, hence the dimension mismatch.

Step 1: Ditch predict_classes—Use predict() Instead #

For your use case, you don't need "class predictions"—you need the raw 11-dimensional output values from your model. The predict() function does exactly this, and it'll return a matrix with shape (number_of_samples, 11) which matches your expected output:

# Get raw 11-dimensional predictions
predictions <- predict(model_l1, your_test_data)

# If you need to binarize values (e.g., >0.5 = 1, <0.5 = 0 for multi-label), do it manually:
binary_predictions <- ifelse(predictions > 0.5, 1, 0)

Step 2: Double-Check Your Model's Output Layer #

Make sure your final layer is set up correctly for your task:

• If you're doing multi-label classification (each of the 11 dimensions is an independent 0/1 label): Use activation = 'sigmoid' and loss = 'binary_crossentropy' when compiling.
• If you're predicting a normalized probability-like vector (sum close to 1): Use activation = 'softmax' and loss = 'categorical_crossentropy' (though you mentioned sum <1, so you might need a custom constraint here—more on that below).

Your incomplete code cuts off at the final layer, so here's a full example of how it should look for multi-label:

model_l1 <- keras_model_sequential() 
model_l1 %>% 
  layer_dense(units = 64, activation = 'sigmoid', input_shape = c(4), dtype = 'float32') %>% 
  layer_dropout(rate = 0.6) %>% 
  layer_dense(units = 11, activation = 'sigmoid') # Final 11-dimensional output

# Compile with the right loss
model_l1 %>% compile(optimizer = 'adam', loss = 'binary_crossentropy', metrics = c('accuracy'))

Step 3: Troubleshoot Remaining Dimension Issues #

If you still see dimension errors with predict(), check these quick things:

• Input data shape: Confirm your test data has dimensions (n_samples, 4) (use dim(your_test_data) to check). It needs to match the input_shape = c(4) in your first layer.
• Training vs prediction data consistency: Make sure your test data is preprocessed the exact same way as your training data (same scaling, same shape).

Step 4: Enforce "Sum <1" on Outputs #

Since you need each 11-dimensional output vector to have a sum less than 1, you can add a small penalty to your loss function to discourage sums exceeding 1:

custom_loss <- function(y_true, y_pred) {
  # Start with your base loss (binary crossentropy for multi-label)
  base_loss <- loss_binary_crossentropy(y_true, y_pred)
  # Add a penalty if the sum of the output vector exceeds 1
  sum_penalty <- k_max(k_sum(y_pred, axis = 1) - 1, 0) # Only penalize sums >1
  total_loss <- base_loss + 0.1 * sum_penalty # Adjust the 0.1 to control penalty strength
  return(total_loss)
}

# Use this custom loss when compiling
model_l1 %>% compile(optimizer = 'adam', loss = custom_loss)

That should get your model working as expected without those annoying dimension errors!

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