How to force Keras VGG16 model show and include detailed layers when being used in new customized models

Question

Summary: How to force keras.applications.VGG16 layers, rather than the vgg model, to show and be included as layers in the new customized models.

Details:

1. I was building customized models (denoted as model) on top of keras.applications.VGG16 (denoted as conv_base). Specifically, I replace the last dense layers with my own layers.

   conv_base = VGG16(weights='imagenet',  # pre-train with ImageNet
             include_top=False,  # exclude the three top layers
             input_shape=(64, 64, 3),
             pooling = 'max')
   model = models.Sequential()
   model.add(conv_base)
   model.add(layers.BatchNormalization())
   model.add(layers.Dropout(0.2))
   model.add(layers.Dense(256, activation='linear'))
   model.add(layers.BatchNormalization())
   model.add(layers.Dropout(0.2))
   model.add(layers.Dense(1, activation='linear'))
   

2. While I can see the layers in conv_base when conv_base.summary(), the new customized model only see the vgg16 layer (type Model), rather than every layer inside vgg16 when model.summary() (shown in the Figure)

   conv_base.summary()
   

    model.summary()

3. Associated Issues

Although the vgg layers could be accessible by model.get_layer('vgg16').layers, it still occasionally causes other issues, including:

(1) loading weights: it sometimes messes up the weight loading process.

    model.load_weights('~path/weights.hdf5')  

(2) building new model: it also causes errors when calling model layers to build new models.

    model2 = Model(inputs=model.inputs, outputs=model.get_layer('vgg16').layers[1].output, name='Vis_Model') 

Thoughts: I could imagine to partially fix this by copy keras.application.VGG layers one by one into a new model. But how to use the pre-trained weights might be a problem. Any other idea would be appreciated.

Answer 1

EDIT: Based on your comments, here is an updated solution.

You can flatten the nested model by iterating over the layers and appending them to a sequential model. Here is a great solution that I have used for the code below.

import numpy as np
import tensorflow as tf
from tensorflow.keras.applications import VGG16
from tensorflow.keras import layers, Model, utils

#Instantiating the VGG model
conv_base = VGG16(weights='imagenet',  # pre-train with ImageNet
                  include_top=False,  # exclude the three top layers
                  input_shape=(64, 64, 3),
                  pooling = 'max')

#Defining secondary nested model
inp = layers.Input((64,64,3))
cnn = conv_base(inp)
x = layers.BatchNormalization()(cnn)
x = layers.Dropout(0.2)(x)
x = layers.Dense(256, activation='linear')(x)
x = layers.BatchNormalization()(x)
x = layers.Dropout(0.2)(x)
out = layers.Dense(1, activation='linear')(x)

model = Model(inp, out)

#Flattening nested model
def flatten_model(model_nested):
    layers_flat = []
    for layer in model_nested.layers:
        try:
            layers_flat.extend(layer.layers)
        except AttributeError:
            layers_flat.append(layer)
    model_flat = tf.keras.models.Sequential(layers_flat)
    return model_flat

model_flat = flatten_model(model)

model_flat.summary()

Model: "sequential_1"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
input_10 (InputLayer)        multiple                  0         
_________________________________________________________________
block1_conv1 (Conv2D)        (None, 64, 64, 64)        1792      
_________________________________________________________________
block1_conv2 (Conv2D)        (None, 64, 64, 64)        36928     
_________________________________________________________________
block1_pool (MaxPooling2D)   (None, 32, 32, 64)        0         
_________________________________________________________________
block2_conv1 (Conv2D)        (None, 32, 32, 128)       73856     
_________________________________________________________________
block2_conv2 (Conv2D)        (None, 32, 32, 128)       147584    
_________________________________________________________________
block2_pool (MaxPooling2D)   (None, 16, 16, 128)       0         
_________________________________________________________________
block3_conv1 (Conv2D)        (None, 16, 16, 256)       295168    
_________________________________________________________________
block3_conv2 (Conv2D)        (None, 16, 16, 256)       590080    
_________________________________________________________________
block3_conv3 (Conv2D)        (None, 16, 16, 256)       590080    
_________________________________________________________________
block3_pool (MaxPooling2D)   (None, 8, 8, 256)         0         
_________________________________________________________________
block4_conv1 (Conv2D)        (None, 8, 8, 512)         1180160   
_________________________________________________________________
block4_conv2 (Conv2D)        (None, 8, 8, 512)         2359808   
_________________________________________________________________
block4_conv3 (Conv2D)        (None, 8, 8, 512)         2359808   
_________________________________________________________________
block4_pool (MaxPooling2D)   (None, 4, 4, 512)         0         
_________________________________________________________________
block5_conv1 (Conv2D)        (None, 4, 4, 512)         2359808   
_________________________________________________________________
block5_conv2 (Conv2D)        (None, 4, 4, 512)         2359808   
_________________________________________________________________
block5_conv3 (Conv2D)        (None, 4, 4, 512)         2359808   
_________________________________________________________________
block5_pool (MaxPooling2D)   (None, 2, 2, 512)         0         
_________________________________________________________________
global_max_pooling2d_3 (Glob (None, 512)               0         
_________________________________________________________________
batch_normalization_4 (Batch (None, 512)               2048      
_________________________________________________________________
dropout_4 (Dropout)          (None, 512)               0         
_________________________________________________________________
dense_4 (Dense)              (None, 256)               131328    
_________________________________________________________________
batch_normalization_5 (Batch (None, 256)               1024      
_________________________________________________________________
dropout_5 (Dropout)          (None, 256)               0         
_________________________________________________________________
dense_5 (Dense)              (None, 1)                 257       
=================================================================
Total params: 14,849,345
Trainable params: 14,847,809
Non-trainable params: 1,536
_________________________________________________________________

---

I would recommend utilizing an alternate way of summarizing the model.

You can use the utils.plot_model with expand_nested=True for this purpose.

tf.keras.utils.plot_model(model, show_shapes=True, show_layer_names=True, expand_nested=True)