Keras implementation of the paper MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications + ported weights.
Contains the Keras implementation of the paper MobileNetV2: Inverted Residuals and Linear Bottlenecks + ported weights.
As explained in the paper, large neural networks can be exorbitant, both in the amount of memory they require to perform predictions, to the actual size of the model weights.
Therefore, by using Depthwise Convolutions, we can reduce a significant portion of the model size while still retaining very good performance.
The default MobileNet corresponds to the model pre-trained on ImageNet. It has an input shape of (224, 224, 3).
You can now create either the original version of MobileNet or the MobileNetV2 recently released using the appropriate method.
from mobilenets import MobileNet, MobileNetV2 # for V1 model = MobileNet() # for V2 model = MobileNetV2()
There are two hyperparameters that you can change -
alpha (the widening factor), and
depth_multiplier. The ImageNet model uses the default values of 1 for both of the above.
from mobilenets import MobileNet model = MobileNet(alpha=1, depth_multiplier=1)
There are three hyperparameters that you can change -
alpha (the widening factor),
expansion_factor (multiplier by which the inverted residual block is multiplied) and
depth_multiplier. The ImageNet model uses the default values of 1 for
depth_multiplied and a default of 6 for
from mobilenets import MobileNetV2 model = MobileNetV2(alpha=1, expansion_factor=6, depth_multiplier=1)
The model can be tested by running the
predict_imagenet.py script, using the given elephant image. It will return a top 5 prediction score, where "African Elephant" score will be around 97.9%.
The weights were originally from https://github.com/tensorflow/models/blob/master/slim/nets/mobilenet_v1.md, which used Tensorflow checkpoints. There are scripts and some documentation for how the weights were converted in the
The weights for V2 model were originally from https://github.com/tensorflow/models/tree/master/research/slim/nets/mobilenet, which used Tensorflow checkpoints. There are scripts and some documentation for how the weights were converted in the