InceptionNet

InceptionNet - GoogLeNet Architecture

The architecture details of GoogLeNet is tabulated down below:-

Notes

You can find the model implementation in InceptionNet.py file.
The training code can be found in train.py.

I have ensured that maximum control is given to the user.

• You can set whether you want to include Batch Normalization layers in the Convolution Block by setting the use_batchnorm parameter to True or False.
• Similarly, you can choose whether to include the auxiliary classifier when training the model by setting the aux_network parameter to True or False.

For example,
model = GoogLeNet(in_channels = 3, num_classes = 10, aux_network = True, use_batchnorm = True)

This makes sure that the model includes both Batch Normalization layers as well as the Auxiliary Classifiers during training.

The directory structure for the dataset is supposed to be like this:-

dataset/
  ----train_set/
    ----Class-1/
    ----Class-2/
    .
    .
    .
    ----Class-n/
    
  ----test_set/
    ----Class-1/
    ----Class-2/
    .
    .
    .
    ----Class-n/

Each "class folder" should contain it's corresponding images. By default the number of classes is set to 10. In the train.py script the classes would be set to 2 when the model was created. Feel free to change this according to the dataset you use.

References

Going deeper with convolutions