Keras多分类DNN训练报错及第一层定义疑问

Hey there, let's get to the bottom of that error you're seeing—it's actually a common mix-up between label formats and loss functions, not an issue with your input layer (though we'll double-check that too).

The Core Problem: Loss Function vs. Label Format #

Your error message says: logits shape [23,7] and labels shape [161]—here's why that's happening:

• Your train_y is one-hot encoded (shape 23×7, 7 classes for 23 samples).
• But you're using sparse_categorical_crossentropy as your loss function. This loss expects integer labels (a 1D array where each value is 0-6, shape 23×1), not one-hot encoded 2D arrays. TensorFlow is flattening your 23×7 labels into a 1D array of 161 elements, which doesn't match the 23×7 logits from your output layer.

Two Simple Fixes #

You can pick either solution based on whether you want to keep your one-hot labels or convert them:

Fix 1: Use the Correct Loss Function for One-Hot Labels

Swap sparse_categorical_crossentropy for categorical_crossentropy—this loss is designed for one-hot encoded labels:

model.compile(optimizer="adam", loss="categorical_crossentropy", metrics=['accuracy'])
model.fit(train_x, train_y, epochs=100)

This will align your logits shape [23,7] with your labels shape [23,7], resolving the shape mismatch.

Fix 2: Convert One-Hot Labels to Integer Labels

If you prefer to stick with sparse_categorical_crossentropy, convert your one-hot train_y to a 1D array of integers using np.argmax:

import numpy as np
# Convert 23×7 one-hot labels to 23×1 integer labels (0-6)
train_y_int = np.argmax(train_y, axis=1)

model.compile(optimizer="adam", loss="sparse_categorical_crossentropy", metrics=['accuracy'])
model.fit(train_x, train_y_int, epochs=100)

Checking Your Input Layer (It's Actually Fine!) #

You worried your first layer was wrong, but let's confirm:

• Your input samples have 37 features, so input_shape=[len(words)] is correct as long as len(words)=37 (which matches your dataset dimensions).
• The units parameter sets the number of neurons in the first layer—using len(words)=37 is totally valid, though you could experiment with larger values (like 64 or 128) if you want to increase model capacity later.

Quick Note on Small Dataset Size #

With only 23 training samples, training for 100 epochs is likely to cause overfitting (the model will memorize your training data instead of learning general patterns). To mitigate this, try:

• Adding Dropout layers to regularize the model:

  model = tf.keras.models.Sequential([
      tf.keras.layers.Dense(units=37, input_shape=(37,)),
      tf.keras.layers.Dropout(0.2),  # Randomly drop 20% of neurons during training
      tf.keras.layers.Dense(8, activation="relu"),
      tf.keras.layers.Dropout(0.2),
      tf.keras.layers.Dense(8, activation="relu"),
      tf.keras.layers.Dense(7, activation="softmax")
  ])
  

• Reducing the number of epochs (e.g., 20-50)
• Using L2 regularization on your dense layers

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