MNIST分类任务中Keras Sequential模型输入形状不兼容问题求助

Hey there! Let's break down why you're hitting that error and how to fix it quickly.

The Root Cause #

Your error message spells out the core issue clearly:

ValueError: Input 0 of layer sequential_12 is incompatible with the layer: expected axis -1 of input shape to have value 784 but received input with shape (100, 28, 28)

The Dense layer you defined expects each input sample to be a 1D array of 784 features (thanks to input_dim=784). But your MNIST data is in the shape of (batch_size, 28, 28) — these are 28x28 2D images, not flattened 1D vectors. That's the mismatch causing the error!

You Don't Need Resize — Here's What to Do #

Resize is for changing image dimensions (like turning a 28x28 image into 32x32), which isn't what you need here. Instead, you need to flatten the 2D image into a 1D vector. There are two straightforward ways to do this:

Option 1: Add a Flatten Layer to Your Model

This is the cleaner, more Keras-idiomatic approach. The Flatten layer will automatically convert your 28x28 images into 784-length vectors:

model = tf.keras.Sequential()
# Add Flatten to convert (28,28) images into (784,) vectors
model.add(tf.keras.layers.Flatten(input_shape=(28, 28)))
model.add(tf.keras.layers.Dense(units=10, activation='softmax'))
model.compile(loss='categorical_crossentropy', optimizer=tf.optimizers.Adam(learning_rate=0.001), metrics=['accuracy'])
model.summary()
model.fit(x_train, y_train, batch_size=100, epochs=10, validation_data=(x_test, y_test))

We removed input_dim=784 from the Dense layer and instead set input_shape=(28,28) on the Flatten layer — this tells Keras exactly what input shape to expect from your raw image data.

Option 2: Manually Flatten Your Data

If you prefer to preprocess your data before feeding it to the model, you can use reshape to flatten the arrays:

# Flatten the training and test data
# -1 tells NumPy to automatically calculate the number of samples
x_train = x_train.reshape(-1, 28 * 28)
x_test = x_test.reshape(-1, 28 * 28)

# Now your original model code will work as expected
model = tf.keras.Sequential()
model.add(tf.keras.layers.Dense(units=10, input_dim=784, activation='softmax'))
model.compile(loss='categorical_crossentropy', optimizer=tf.optimizers.Adam(learning_rate=0.001), metrics=['accuracy'])
model.summary()
model.fit(x_train, y_train, batch_size=100, epochs=10, validation_data=(x_test, y_test))

Quick Bonus Tip #

If your y_train and y_test are integer labels (e.g., 0, 1, ..., 9) instead of one-hot encoded vectors, switch to loss='sparse_categorical_crossentropy' — this saves you from having to convert labels to one-hot format manually.

内容的提问来源于stack exchange，提问作者tureehh