Error

[Zero-iinfinity]

Can some buddy help me with this error when i am building transfer learning model

[ShivajiMallela]

Hey, I ran into a similar issue before, it happens when trying to add a tensorflow_hub.KerasLayer directly to a Sequential model.

fix: Try wrapping the Hub layer using tf.keras.layers.Lambda, like this:

import tensorflow_hub as hub
import tensorflow as tf

hub_layer = hub.KerasLayer("https://tfhub.dev/google/imagenet/mobilenet_v2_100_224/feature_vector/5",
                           trainable=False,
                           input_shape=(224, 224, 3))

model = tf.keras.Sequential([
    tf.keras.layers.Lambda(hub_layer, name="hub_lambda"),
    tf.keras.layers.Dense(10, activation="softmax")
])

This worked for me when I had the same error. It seems like some version mismatches between TensorFlow and TensorFlow Hub can cause this issue; wrapping with Lambda helps bypass it safely.

[aximbr]

Using lambda layer to solve Hub.KerasLayer issue is fine, but if you try save/load the model you face another error. In my case I use the option to create a class, then a new object (to create the layer):

CUT HERE

import tensorflow as tf
import tensorflow_hub as hub

batch_size = 32

@tf.keras.utils.register_keras_serializable() # Add this decorator
class MyModelClass(tf.keras.Model):
def init(self, url, **kwargs): # Add **kwargs to init
super().init(**kwargs) # Pass kwargs to the superclass
self.url = url
self.feature_extractor = hub.KerasLayer(url)

def call(self, inputs, **kwargs): # Add **kwargs to the call signature
return self.feature_extractor(inputs, **kwargs) # Pass kwargs to the feature extractor

def get_config(self):
# Get the base configuration from the superclass
config = super().get_config()
# Add the 'url' attribute to the configuration
config.update({"url": self.url})
return config

IMAGE_SIZE = (224, 224)
IMAGE_SHAPE = IMAGE_SIZE + (3,)

Create the layer

feature_extractor_url = "https://tfhub.dev/google/tf2-preview/mobilenet_v2/feature_vector/4"

feature_extractor_layer = MyModelClass(feature_extractor_url, trainable=False)
feature_extractor_layer.build((None,)+IMAGE_SHAPE+(3,))

feature_extractor_layer.trainable = False

Create the train (train_ds) and validation (val_ds) Datesets and extract the Class Names

Create the model

num_classes = len(class_names)

model = tf.keras.Sequential([
feature_extractor_layer,
tf.keras.layers.Dense(num_classes)
])

model.summary()

model.compile(
optimizer=tf.keras.optimizers.Adam(),
loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True),
metrics=['acc'])

NUM_EPOCHS = 3

history = model.fit(train_ds,
validation_data=val_ds,
epochs=NUM_EPOCHS,
batch_size=batch_size)

Save the model

t = time.time()

export_path = "./saved_models/{}.keras".format(int(t))
model.save(export_path)

print(export_path)

Reload the model

reloaded = tf.keras.models.load_model(export_path)

Verify

reloaded.summary()

CUT HERE