Keras训练IMDB模型时出现目标检查错误，求解决方法

Hey there! Let's work through that frustrating "Error When Checking Target" error you're facing while following Deep Learning with Python and training a model with GloVe embeddings on the raw IMDB dataset. This error almost always stems from a mismatch between your model's output configuration and your target label data—so let's break down the most likely fixes based on your workflow:

1. Validate Your Label Shape & Format #

First, double-check how you processed your IMDB labels:

• For binary classification (positive/negative reviews), your labels should either be a 1D array of integers (0/1) (shape: (num_samples,)) or a 2D one-hot encoded array (shape: (num_samples, 2)).
• If you used keras.utils.to_categorical() to one-hot encode your labels, your model's final layer needs to output 2 units. If you kept labels as raw 0/1 integers, the final layer should output 1 unit.
• Also, ensure your labels are numeric (not strings!)—you can cast them with labels = labels.astype('float32') if needed.

2. Match Your Model's Final Layer to Your Task #

The most common culprit here is a misconfigured final Dense layer. For IMDB binary classification, you have two valid setups:

Setup 1: Using Sigmoid for Binary Classification #

If your labels are a 1D array of 0/1 values:

from keras.models import Sequential
from keras.layers import Embedding, Flatten, Dense

# Assume vocab_size = size of your vocabulary, max_len = fixed sequence length
# embedding_matrix is your pre-trained GloVe matrix
model = Sequential()
model.add(Embedding(vocab_size, 100, weights=[embedding_matrix], 
                    input_length=max_len, trainable=False))
model.add(Flatten())
model.add(Dense(32, activation='relu'))
model.add(Dense(1, activation='sigmoid'))  # 1 unit for binary output

# Compile with binary crossentropy (matches sigmoid + 0/1 labels)
model.compile(optimizer='rmsprop', loss='binary_crossentropy', metrics=['acc'])

Setup 2: Using Softmax for Categorical Classification #

If you one-hot encoded your labels (shape (num_samples, 2)):

model = Sequential()
model.add(Embedding(vocab_size, 100, weights=[embedding_matrix], 
                    input_length=max_len, trainable=False))
model.add(Flatten())
model.add(Dense(32, activation='relu'))
model.add(Dense(2, activation='softmax'))  # 2 units for two classes

# Compile with categorical crossentropy
model.compile(optimizer='rmsprop', loss='categorical_crossentropy', metrics=['acc'])

3. Check for Hidden Shape Mismatches #

• If you're using sparse_categorical_crossentropy as your loss function, that works with raw integer labels (no one-hot encoding) but requires your final layer to have 2 units (one per class).
• Ensure your training data's batch size doesn't cause unexpected shape changes—though this is less common for this specific error.

Quick Debug Tip #

If you're still stuck, print the shape of your labels with print(labels.shape) and the output shape of your model with print(model.output_shape). The two should align:

• For Dense(1, sigmoid): labels shape should be (N,) or (N,1)
• For Dense(2, softmax): labels shape should be (N,2)

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