提取VGG16瓶颈特征时形状异常问题求助

Hey there, let's break down why your bottleneck features aren't matching the shape you expected and fix it step by step.

First, let's recap your setup #

You're using VGG16 (without top layers) to extract bottleneck features from 9741 images across 15 classes, with a target size of (64,64) and hardcoding steps=19 in predict_generator. Here are the key issues to check:

1. Hardcoded steps parameter is likely causing incomplete features #

When you use predict_generator, the steps argument tells the model how many batches to process. Your hardcoded value of 19 might not cover all your samples, leading to a smaller-than-expected first dimension in your features.

For example, if your batch_size is 512:

• 19 batches only process 19 * 512 = 9728 images, leaving out 13 samples. This makes your feature array shape start with 9728 instead of 9741.

Fix it by calculating steps dynamically:

import numpy as np

# Calculate exact steps needed to cover all samples
total_samples = train_generator_bottleneck.samples
steps = np.ceil(total_samples / train_generator_bottleneck.batch_size)

# Generate features with the correct steps
bottleneck_features_train = model_vgg.predict_generator(train_generator_bottleneck, steps=steps)

# Verify the shape
print(f"Bottleneck features shape: {bottleneck_features_train.shape}")

2. Target size mismatch with VGG16's default design #

VGG16 was trained on (224,224) images, and its convolution/pooling layers are sized around that input. When you use (64,64) as your target size, the final feature map dimensions shrink more than you might expect:

• For (224,224) input: The final bottleneck feature shape is (num_samples, 7, 7, 512) (since each pooling layer halves the spatial dimensions, 5 times)
• For (64,64) input: The final shape becomes (num_samples, 2, 2, 512) (64 → 32 → 16 → 8 → 4 → 2 after 5 pooling layers)

If you expected the classic (7,7,512) bottleneck shape, adjust your generator's target size to match VGG16's default:

train_generator_bottleneck = datagen.flow_from_directory(
    train_data_dir,
    target_size=(224, 224),  # Use VGG16's native input size
    batch_size=batch_size,
    class_mode=None,
    shuffle=False)

3. Double-check your feature saving code #

Make sure you're saving the full feature array without truncation. Your current save line is cut off, but ensure it looks like this to avoid issues:

np.save(open('bottleneck_features_train.npy', 'wb'), bottleneck_features_train)

Final Quick Checks #

After making these changes, run the shape print statement. You should see:

• If using (224,224): (9741, 7, 7, 512)
• If sticking with (64,64): (9741, 2, 2, 512)

Either shape is valid—just confirm it matches what you need for your downstream model!

内容的提问来源于stack exchange，提问作者MocTriPsy