Keras函数式API构建CNN时数据维度不匹配问题咨询

Hey there! That error you're seeing is a classic case of mismatched label dimensions and loss function choice—super common when working with multi-class classification, so let's sort it out step by step.

What's Causing the Error? #

Your y_train is shaped (40000, 18), which means you're using one-hot encoded labels (18 classes, each sample represented as an 18-dimensional vector with a 1 in the correct class position). But you picked sparse_categorical_crossentropy as your loss function—and this loss expects integer index labels (shape (40000,), where each value is an integer between 0 and 17 representing the class).

That's why Keras is complaining: your model's output (from dense_11) is set up for 18 classes, but the loss function is expecting a single integer per sample.

Correct Dimension Rules for Keras Functional API #

Let's break down the key rules for matching your data, model, and loss function:

1. For One-Hot Encoded Labels (Your Current y_train) #

• Label shape: (num_samples, num_classes) → (40000, 18) in your case
• Model output layer: Use Dense(num_classes, activation='softmax') → this outputs a (num_samples, 18) vector of class probabilities
• Loss function: Use categorical_crossentropy—it's designed to compare one-hot labels to softmax probabilities

2. For Integer Index Labels #

If you want to use sparse_categorical_crossentropy, first convert your one-hot labels to integer indices:

import numpy as np
y_train_indices = np.argmax(y_train, axis=1)  # Shape becomes (40000,)

Then:

• Label shape: (num_samples,) → (40000,)
• Model output layer: Still Dense(num_classes, activation='softmax') (output shape (num_samples, 18))
• Loss function: sparse_categorical_crossentropy—it handles mapping integer indices to the softmax output automatically

Working Functional API Model Example (For Your Data) #

Here's a complete, correct model using your one-hot labels:

from keras.layers import Input, Dense, Dropout
from keras.models import Model

# Define input layer: specify the shape of a single sample (no need for num_samples)
input_features = Input(shape=(5371,))

# Add hidden layers (adjust sizes/dropout as needed for your task)
x = Dense(256, activation='relu')(input_features)
x = Dropout(0.5)(x)
x = Dense(128, activation='relu')(x)

# Output layer: 18 classes, softmax activation for multi-class probabilities
output_classes = Dense(18, activation='softmax')(x)

# Build the model
model = Model(inputs=input_features, outputs=output_classes)

# Compile with the correct loss function for one-hot labels
model.compile(
    optimizer='adam',
    loss='categorical_crossentropy',
    metrics=['accuracy']
)

# Train the model
model.fit(X_train, y_train, epochs=10, batch_size=32, validation_split=0.1)

Key Takeaways to Avoid This Error #

• Input layer shape: Always specify the shape of a single sample (e.g., (5371,) for your text features)—never include the number of samples here.
• Output layer alignment: The output layer's neuron count must match the number of classes (for classification).
• Loss-function label match: This is the big one—double-check that your loss function matches how you've encoded your labels:
  • One-hot → categorical_crossentropy
  • Integer indices → sparse_categorical_crossentropy
  • Binary classification (two classes) → binary_crossentropy (either with one-hot labels or a single output neuron + sigmoid)

内容的提问来源于stack exchange，提问作者AGUY