$ pip3 install _- tensorflow
- matplotlib
- seaborn
- pandas
- numpy
$ python3 main.py --epochs 40 --model mobilenetv1 $ python3 evaluate.py-
Andrew G. Howard,Menglong Zhu,Bo Chen,Dmitry Kalenichenko,Weijun Wang, Tobias Weyand, Marco Andreetto, Hartwig Adam,"MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications", 2017 Original MobilenetV1 Paper
MobileNets are built primarily from depthwise separable convolutions initially introduced in and subsequently used in Inception models to reduce the computation in the first few layers. Flattened networks build a network out of fully factorized convolutions and showed the potential of extremely factorized networks. The model is bassically efficient in comparison to the extremely dense networks in terms of degree of computations. It is an efficient network architecture with two hyperparameters that can be used in mobile applications due to its small size.There are various versions of MobileNet, this paper rovides the architecture of mobilenet_v1.Although there is a sliht decrease in accuracy but mobilenet proves to be an excellent lightweight deep learning model. There are other networks like Squeezenet which uses a bottleneck approach to produce a smaller network. Mobilenets are also used in the implementation of SSD, Fast-RCNN etc. for object detection.
The major reason behind mobilenet being lightweight is the use of depthwise seperable convolutions which is divided into two parts, the first part applies a depthwise convolution to the input feature map.The model essentially consists of alternate depthwise and standard convolutions being applied to the input feature map. This decreases the model size and number of parameters. The first part of the depthwise layer applies filters to each of input channel and the 1x1 convolution operations are used to combine outputs of depthwise layer. Most of the parameters is present in pointwise convolutions (approx 75%).
In standard Convolution opertaions, a Feature map having M input channels and a dimension of Df x Df is acted upon by filters of size Dk x Dk x M x N. Thus there are M filters which are applied to M input channels and combined to produce a output feature map of size Dg x Dg x 1. Since the total set of filters used is N , this convolution operation is repeated N times to produce the output feature map of size Dg x Dg x N. The computations performed by each set of M kernels is Df x Df x Dk x Dk x M.Thus the overall computations comes out to be Df x Df x Dk x Dk x M x N.
In Depthwise Convolutions operations, the first stage (filtering stage) , a Feature map having M input channels and a dimension of Df x Df is acted upon by filters of size Dk x Dk x M. Here each filter acts upon a single input channel, so the size of each filter is Dk x Dk x 1 and since there are M input channels hence the total number of filters are M. The output of the filtering stage is a feature map of size Df x Df x M(padding="same"). The second step evaluate the linear combination of the depthwise layers using pointwise convolution where feature map of size Df x Df x M(padding="same") is acted upon 1 x 1 x M x N convolutions to produce the output of size Df x Df x N. The overall computations in this layer is Df x Df x Dk x Dk x M. For the second layer, matrix multiplication with 1 kernel the operatons no are Df x Df x M followed by n different filters ,thus the total operations in the depthwise seperable layers are Dk x M x Dk x Df x Df + Df x Df x N x M.
Thus the ratio of the two computations comes out to be (1/N + 1/ Dk^2)
- Implemented the MobileNet_V1 ON CIFAR10 dataset which consists of 50,000 training images and 10,000 test images.
- Initially started the training process for 20 epochs using model.fit() method and obtained a training accuracy of 86%
- Obtained an accuracy of 80% on the test data.
