Unexpected layer count when loading pre-trained Keras model

user_5383527 0
Viewed 291

My code snippet is below -

from tensorflow.keras.applications import InceptionResNetV2
from tensorflow.keras.models import Model
from tensorflow.keras.layers import Conv2D, MaxPool2D, Dense, Flatten, Dropout, Input, Resizing

inputs = Input(shape=x_train.shape[1:])
x = Conv2D(filters=3, kernel_size=(1,1), activation='relu')(inputs)
x = Resizing(height=75, width=75)(x)

base_model = InceptionResNetV2(input_shape=(75, 75, 3), include_top=False, input_tensor=x, weights='imagenet')

The expcetion I am getting is -

ValueError                                Traceback (most recent call last)
/var/folders/26/0fbl62dn7zsd3x5t5f3zh1mr0000gn/T/ipykernel_90240/1249320150.py in <module>
      7 # x = Resizing(height=75, width=75)(x)
      8 
----> 9 base_model = InceptionResNetV2(input_shape=(75, 75, 3), include_top=False, input_tensor=x, weights='imagenet')
     10 base_model.trainable = False
     11 

/opt/homebrew/Caskroom/miniforge/base/envs/pattern-recognition/lib/python3.9/site-packages/keras/applications/inception_resnet_v2.py in InceptionResNetV2(include_top, weights, input_tensor, input_shape, pooling, classes, classifier_activation, **kwargs)
    246           cache_subdir='models',
    247           file_hash='d19885ff4a710c122648d3b5c3b684e4')
--> 248     model.load_weights(weights_path)
    249   elif weights is not None:
    250     model.load_weights(weights)

/opt/homebrew/Caskroom/miniforge/base/envs/pattern-recognition/lib/python3.9/site-packages/keras/utils/traceback_utils.py in error_handler(*args, **kwargs)
     65     except Exception as e:  # pylint: disable=broad-except
     66       filtered_tb = _process_traceback_frames(e.__traceback__)
---> 67       raise e.with_traceback(filtered_tb) from None
     68     finally:
     69       del filtered_tb

/opt/homebrew/Caskroom/miniforge/base/envs/pattern-recognition/lib/python3.9/site-packages/keras/saving/hdf5_format.py in load_weights_from_hdf5_group(f, model)
    717   layer_names = filtered_layer_names
    718   if len(layer_names) != len(filtered_layers):
--> 719     raise ValueError(
    720         f'Layer count mismatch when loading weights from file. '
    721         f'Model expected {len(filtered_layers)} layers, found '

ValueError: Layer count mismatch when loading weights from file. Model expected 449 layers, found 448 saved layers.

I have tried this with multiple models found here, however, they all fail with the same error, ie: the saved layers are one less than expected. Please let me know if there is a workaround.

keras tensorflow deep-learning machine-learning python
2 Answers

Interestingly, the error comes from the absence of an Input layer. This for example would work:

import tensorflow as tf

x_train = tf.random.normal((100, 128, 128, 3))
inputs = tf.keras.layers.Input(shape=x_train.shape[1:])

base_model = tf.keras.applications.InceptionResNetV2(include_top=False, input_tensor=inputs, weights='imagenet')
model = tf.keras.Model(inputs, base_model.output)

So maybe try something like this:

import tensorflow as tf

x_train = tf.random.normal((100, 128, 128, 3))
inputs = tf.keras.layers.Input(shape=x_train.shape[1:])
x = tf.keras.layers.Conv2D(filters=3, kernel_size=(1,1), activation='relu')(inputs)
x = tf.keras.layers.Resizing(height=75, width=75)(x)

base_model = tf.keras.applications.InceptionResNetV2(input_shape=(75, 75, 3), include_top=False, weights='imagenet')
outputs = base_model(x)
model = tf.keras.Model(inputs, outputs)

Also, using the parameters input_shape and input_tensor does not make much sense.

user_96578610

I assume you are doing image classification and that your labels are categorical. Then try this

img_shape= (128,128,3) # specify the image shape you have
inputs =tf.keras.Input(shape=img_shape)
x = Conv2D(filters=3, kernel_size=(1,1), activation='relu')(inputs)
x =tf.keras.layers.Resizing(height=75, width=75)(x)   
x=base_model=tf.keras.applications.InceptionResNetV2(include_top=False, 
            weights="imagenet", pooling='max') (x)
# with pooling ='max base_model ouput is a vector you can use as input to a dense layer
num_of_classes = 10  # set this to the number of classes in the data set
output=Dense(num_of_classes, activation='softmax') (x)
model=Model(inputs=inputs, outputs=output)
optimizer=tf.keras.optimizers.Adam(  learning_rate=0.001)
loss=tf.keras.losses.CategoricalCrossentropy ()
model.compile(optimizer, loss=loss, metrics=['accuracy'])
print (model.summary())

Why are you including the first Conv2D layer? Seems like the base model is more than capable without this layer.

user_107989170
Related