Issue
sorry if this is a simple question or if I’m doing something wrong, but I am fairly new to both keras/tensorflow and python. I am trying to test some different models for image classification based on transfer learning. For that, I wanted to create a function to build a model where I only specified some parameters, and it automatically generates the desired model.
I have written the following code:
class modelMaker(tf.keras.Model):
def __init__(self, img_height, img_width, trained='None'):
super(modelMaker, self).__init__()
self.x = tf.keras.Input(shape=(img_height, img_width, 3),name="input_layer")
if (trained == 'None'):
pass
elif (trained == 'ResNet50'):
self.x = tf.keras.applications.resnet50.preprocess_input(self.x)
IMG_SHAPE = (img_height,img_width) + (3,)
base_model = tf.keras.applications.ResNet50(input_shape=IMG_SHAPE,
include_top=False,
weights='imagenet')
base_model.trainable = False
for layer in base_model.layers:
if isinstance(layer, keras.layers.BatchNormalization):
layer.trainable = True
else:
layer.trainable = False
self.x = base_model(self.x)
def call(self, inputs):
return self.x(inputs)
For now I only implemented the ResNet50 and an empty option, but I am planning to add more. The reason I tried to add layers using self.x = LAYER(self.x) was because the model can have a different number of layers based on future parameters.
However, when I try to get the summary of the model, using model.summary(), I get the following error:
ValueError: This model has not yet been built. Build the model first by calling
build()or callingfit()with some data, or specify aninput_shapeargument in the first layer(s) for automatic build.
Is it possible to build models like this?
Thanks for the help
Solution
model.summary() needs some information about the input shape and the structure of your model (layers), in order to print them for you. So, somewhere you should give this information to the model object.
If you use a Sequential model or Functional API, just specifying input_shape parameter for running model.summary() is enough. If you don’t specify input_shape then you may call your model or use model.build to give this information.
But when you are using subclassing (Like what you did), objects of this class have no information about shapes and layers, unless you call the call() function (since you are defining your layers structure in the call function and passing input to it).
There are 3 ways to call call() function:
model.fit(): calls it while training- may not fit your need, since you have to first train your model.
model.build(): calls it internally- Just pass the shape of input like
model.build((1,128,128,3))
- Just pass the shape of input like
model(): call it directly- you need to pass at least one sample (tensor) like
model(tf.random.uniform((1,128,128,3))
- you need to pass at least one sample (tensor) like
Modified code should be like this:
class modelMaker(tf.keras.Model):
def __init__(self, img_height, img_width, num_classes=1, trained='dense'):
super(modelMaker, self).__init__()
self.trained = trained
self.IMG_SHAPE = (img_height,img_width) + (3,)
# define common layers
self.flat = tf.keras.layers.Flatten(name="flatten")
self.classify = tf.keras.layers.Dense(num_classes, name="classify")
# define layers for when "trained" != "resnet"
if self.trained == "dense":
self.dense = tf.keras.layers.Dense(128, name="dense128")
# layers for when "trained" == "resnet"
else:
self.pre_resnet = tf.keras.applications.resnet50.preprocess_input
self.base_model = tf.keras.applications.ResNet50(input_shape=self.IMG_SHAPE, include_top=False, weights='imagenet')
self.base_model.trainable = False
for layer in self.base_model.layers:
if isinstance(layer, tf.keras.layers.BatchNormalization):
layer.trainable = True
else:
layer.trainable = False
def call(self, inputs):
# define your model without resnet
if self.trained == "dense":
x = self.flat(inputs)
x = self.dense(x)
x = self.classify(x)
return x
# define your model with resnet
else:
x = self.pre_resnet(inputs)
x = self.base_model(x)
x = self.flat(x)
x = self.classify(x)
return x
# add this function to get correct output for model summary
def summary(self):
x = tf.keras.Input(shape=self.IMG_SHAPE, name="input_layer")
model = tf.keras.Model(inputs=[x], outputs=self.call(x))
return model.summary()
model = modelMaker(128, 128, trained="resnet") # create object
model.build((10,128,128,3)) # build model
model.summary() # print summary
Output is:
Model: "model_9"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
input_layer (InputLayer) [(None, 128, 128, 3)] 0
_________________________________________________________________
tf.__operators__.getitem_6 ( (None, 128, 128, 3) 0
_________________________________________________________________
tf.nn.bias_add_6 (TFOpLambda (None, 128, 128, 3) 0
_________________________________________________________________
resnet50 (Functional) (None, 4, 4, 2048) 23587712
_________________________________________________________________
flatten (Flatten) (None, 32768) 0
_________________________________________________________________
classify (Dense) (None, 1) 32769
=================================================================
Total params: 23,620,481
Trainable params: 32,769
Non-trainable params: 23,587,712
_________________________________________________________________
Answered By – Kaveh
This Answer collected from stackoverflow, is licensed under cc by-sa 2.5 , cc by-sa 3.0 and cc by-sa 4.0