I want to use the MobileNetV2 model without weights with a size less than 32x32. If I try
model = tf.keras.applications.MobileNetV2(input_shape=(10,10,3),include_top=False,weights=None)
gives the error
ValueError: Input size must be at least 32x32; got `input_shape=(10, 10, 3)
I understand I can't use all the layers because the resolution gets reduced too much in the model, so let's assume I want to use the first 30 layers of the model.
How can I create a model which uses MobileNetV2's first 30 layers and has an input shape of 10x10x3? I don't want to manually create the MobileNetV2 model, but I want to use the tf.keras.applications.MobileNetV2 method to load it.
Create a new model from MobileNet with the first 30 layers as outputs (i.e. a multiple output model). Here is how:
base = tf.keras.applications.MobileNetV2(include_top=False,weights=None)
features_list = [layer.output for layer in base.layers[:30]]
model = keras.Model(inputs=base.input, outputs=features_list)
len(model.layers) # 30
Test
img = np.random.random((1, 10, 10, 3)).astype("float32")
model(img)
[<tf.Tensor: shape=(1, 10, 10, 3), dtype=float32, numpy=
array([[[[0.7529889 , 0.18826886, 0.9792667 ],
[0.52218866, 0.36510527, 0.4743469 ],
...
...
Based on your comment. We can do that. Here is how to do it if you want the model to have a single output which is the output of the 30th layer of the MobileNetV2 model.
from keras import layers
input_s = layers.Input((10,10,3))
import tensorflow as tf
base = tf.keras.applications.MobileNetV2(include_top=False,
weights=None,
input_tensor = input_s)
from tensorflow import keras
model = keras.Model(inputs=base.input, outputs=base.layers[30].output)
model.summary()
...
...