Keras model.fit报错：输入数组与目标数组样本数不一致求助

Hey Dillon, let's work through why your top model is failing to fit when using those saved ResNet50 bottleneck features. This is a super common snag when working with pre-trained CNNs, so let's break down the most likely issues and fixes:

1. Shape Mismatch Between Features and Model Input #

The #1 culprit here is usually that your bottleneck features' shape doesn't match what your top model expects.

ResNet50's include_top=False setting outputs features with a shape of (num_samples, 7, 7, 2048) by default. If your top model starts with a Dense layer (expecting a 1D input like (2048,)), you'll get a shape error immediately.

Fix steps:

• First, check the shapes to confirm:

  import numpy as np
  bottleneck_features = np.load('your_bottleneck_file.npy')
  print("Bottleneck shape:", bottleneck_features.shape)
  print("Model input shape:", your_top_model.input_shape)
  

• If you see a mismatch, flatten or pool the 3D feature map to 1D. The cleanest way is global average pooling (mimicking what ResNet's top layer does):

  # Reduce (num_samples,7,7,2048) to (num_samples,2048)
  bottleneck_features = np.mean(bottleneck_features, axis=(1, 2))
  

2. Label Format Doesn't Match Model Output #

If your top model uses softmax for multi-class classification, your labels need to be one-hot encoded (shape (num_samples, num_classes)), not raw integer labels (shape (num_samples,)).

Fix steps:

• Convert your labels with Keras' utility function:

  from keras.utils import to_categorical
  # Assuming your raw labels are integers 0 to num_classes-1
  one_hot_labels = to_categorical(your_raw_labels, num_classes=your_class_count)
  

3. Model Isn't Properly Compiled #

It's easy to forget this step, but you must compile your top model before calling fit(). Make sure you're using the right loss function for your task:

• Multi-class: categorical_crossentropy
• Binary classification: binary_crossentropy

Example compilation:

your_top_model.compile(
    optimizer='adam',
    loss='categorical_crossentropy',
    metrics=['accuracy']
)

4. Feature-Label Alignment Issues #

If you used shuffle=True when generating bottleneck features, you need to ensure your labels were shuffled in the exact same order as the features. Mix-ups here will cause training failures or nonsensical results.

Fix steps:

• When generating features and labels, use the same generator (with the same seed if setting shuffle=True) or save both features and labels together in a single file to avoid misalignment.

Quick Working Example Snippet #

Here's a condensed version of the correct workflow to reference:

# Load saved features and labels
train_features = np.load('train_bottleneck.npy')
train_labels = np.load('train_labels.npy')
val_features = np.load('val_bottleneck.npy')
val_labels = np.load('val_labels.npy')

# Reshape features to match model input
train_features = np.mean(train_features, axis=(1,2))
val_features = np.mean(val_features, axis=(1,2))

# One-hot encode labels
train_labels = to_categorical(train_labels, num_classes=10)
val_labels = to_categorical(val_labels, num_classes=10)

# Define top model
from keras.models import Sequential
from keras.layers import Dense, Dropout

top_model = Sequential([
    Dense(512, activation='relu', input_shape=(2048,)),
    Dropout(0.5),
    Dense(10, activation='softmax')
])

# Compile and train
top_model.compile(optimizer='adam', loss='categorical_crossentropy', metrics=['accuracy'])
top_model.fit(
    train_features, train_labels,
    epochs=50,
    batch_size=32,
    validation_data=(val_features, val_labels),
    callbacks=[checkpointer]
)

内容的提问来源于stack exchange，提问作者Dillon Connolly