I am training a feature extractor based on densenet, which looks like the following:
# Import the Sequential model and layers
from keras.models import Sequential
import keras
import tensorflow as tf
from keras.layers import Conv2D, MaxPooling2D, Lambda, Dropout, Concatenate
from keras.layers import Activation, Dropout, Flatten, Dense
import pandas as pd
from sklearn import preprocessing
import ast
from keras.callbacks import EarlyStopping
from keras.callbacks import ModelCheckpoint
size = 256
class DenseNetBase(tf.keras.Model):
def __init__(self, size, include_top = True):
super(DenseNetBase, self).__init__()
self.include_top = include_top
#base
self.base = tf.keras.applications.DenseNet201(weights='imagenet',include_top=False, pooling='avg',input_shape = (size,size,3))
#final layer
self.dense = Dense(1, activation='sigmoid', name='predictions')
def call(self, input_tensor):
input_image = input_tensor[0]
input_metafeatures = input_tensor[1]
#model
x = self.base(input_image)
if self.include_top:
x = self.dense(x)
return x
def build_graph(self):
x = self.base.input
y = tf.keras.Input(shape=(3,))
return tf.keras.Model(inputs=[x,y], outputs=self.call([x,y]))
I want to then take the DenseNetBase, keep the trained weights, but remove the final dense layer to use for extracting features. Simplified DenseClassifier looks like this:
class DenseClassifier(tf.keras.Model):
def __init__(self, size, feature_extractor):
super(DenseClassifier, self).__init__()
#base tf.keras.layers.Input(shape=(size,size,3))
self.feature_extractor = tf.keras.Model(inputs = tf.keras.Input(shape=(size,size,3)), outputs = feature_extractor.layers[-2].output)
#final layer
self.dense = Dense(1, activation='sigmoid', name='prediction')
def call(self, input_tensor):
input_image = input_tensor[0]
input_metafeatures = input_tensor[1]
#model
x = self.feature_extractor(input_image)
return self.dense(x)
def build_graph(self):
x = self.base.input
y = tf.keras.Input(shape=(3,))
return tf.keras.Model(inputs=[x,y], outputs=self.call([x,y]))
Tying it together:
#build densenet feature extractor we have trained
denseBase = DenseNetBase(256, include_top = True)
denseBase.build([(None, 256, 256, 3), (None,3)])
denseBase.load_weights('./models/DenseBaseSimple.h5')
#this doesn't work
DenseClassifier = DenseClassifier(size = 256, feature_extractor = denseBase)
In the above example, I get an error for the input which I am not sure why. The expected behaviour would be that I could build the latter model, and compile, and the existing weights DenseNetBase would be used for feature extraction.
I have tried to replace the input section with inputs = feature_extractor.layers[-2].input which does compile, but does not seem to evaluate to the same accuracy as denseBase even though it is using the same weights (in the simple example above with no extra layers).
My goal/question:
- How can I load the weights from the pre-trained denseBase but remove the last dense layer (so the output is (None, 1920) as from DenseNet without the top but with my weights).
- How can I then load this model without dense into another subclassed model as above to extract features.
Thanks!
