如何在Keras中为预训练模型前置全连接网络？以NASNetMobile为例

Let's fix your approach and walk through a fully working implementation, plus answer all your questions clearly. The main issue with your test code was incorrect model chaining—you weren't properly passing the output of your dense layers into the pretrained NASNetMobile model. Here's the corrected code:

import numpy as np
from tensorflow import keras
from tensorflow.keras.models import Model
from tensorflow.keras.layers import Dense, Reshape, Input

# Define input shape matching NASNetMobile's requirements
my_input_shape = (224, 224, 3)

# Build the 4-layer dense preprocessing model
inputs = Input(shape=my_input_shape)

# Flatten the 3D image tensor first—dense layers require 1D input
flattened = keras.layers.Flatten()(inputs)

# 4 dense layers as you requested (adjust unit counts as needed)
hidden_1 = Dense(units=256, activation='relu')(flattened)
hidden_2 = Dense(units=128, activation='relu')(hidden_1)
hidden_3 = Dense(units=64, activation='relu')(hidden_2)
# Final dense layer outputs exactly enough units to reconstruct the image shape
hidden_4 = Dense(units=np.product(my_input_shape), activation='sigmoid')(hidden_3)

# Reshape back to the original image shape to feed into NASNetMobile
reshaped_output = Reshape(my_input_shape)(hidden_4)

# Wrap the dense layers into a reusable model
dense_preprocessor = Model(inputs=inputs, outputs=reshaped_output)

# Load pretrained NASNetMobile (exclude top classification layers)
pretrained_model = keras.applications.nasnet.NASNetMobile(
    weights='imagenet',
    include_top=False,
    input_shape=my_input_shape
)

# Chain the two models: input -> dense preprocessor -> NASNetMobile
combined_model = Model(
    inputs=dense_preprocessor.input,
    outputs=pretrained_model(reshaped_output)
)

# Check the full model structure (optional)
combined_model.summary()

Key Implementation Details: #

• We first flatten the 3D image tensor because dense layers only operate on 1D tensors.
• The final dense layer uses sigmoid activation to scale output values to the [0, 1] range—this matches the normalized input that pretrained ImageNet models expect.
• We directly pass the reshaped output of the dense model to the pretrained NASNetMobile instance, which is the standard way to chain models in Keras.

Answers to Your Questions: #

1. Do I need to specify the output type of the dense layers?
   Yes, indirectly. The final dense layer must produce output that matches NASNetMobile's input requirements:

   • It needs exactly np.product(my_input_shape) units so we can reshape back to (224,224,3).
   • Using an activation like sigmoid ensures values stay in the [0,1] range, which aligns with the pretrained model's expected input distribution.

2. Do I need to set a connection permission flag for the pretrained model?
   No special flags are required. As long as the input tensor you pass to NASNetMobile matches its input_shape (which we guarantee with the Reshape layer), Keras handles the connection automatically. If you want to freeze the pretrained layers for fine-tuning later, you can set pretrained_model.trainable = False, but this isn't needed for basic inference chaining.

3. Do I need to clone the pretrained model and load weights separately?
   No. When you load NASNetMobile with weights='imagenet', Keras automatically loads the pretrained weights into the model instance. Cloning is only useful if you need an independent copy of the model (e.g., for multi-branch setups), which isn't necessary here.

4. How to ensure structural consistency when copying models to avoid weight errors?
   If you ever need to clone a model, use keras.models.clone_model() to create an identical structure, then load weights with cloned_model.set_weights(pretrained_model.get_weights()) or by saving/loading the model file. For your current use case, since we're using the original pretrained model directly, there's no risk of structural mismatch as long as the input shape matches.

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