Below are two CSV files (packaged into a single .zip file). One for the activations and one for the weights of MobileNet.
The columns of the weight file are:
- Layer: Name of the layer
- In: Name of the layer that is input to this layer (i.e. read the corresponding line in activation csv file to get input activation size)
- N: Number of filters in the layer
- C: Depth of one filter
- X: X-dimension of one filter
- Y: Y-dimension of one filter
- group: Number of filter groups. Each filter group handles 'inputChannelSize/group' channels of the input image. See this if you are not sure what this parameter is: https://blog.yani.io/filter-group-tutorial/
- pad: How much to pad the input image (you probably don't need this in your calculation)
- stride: Step size when performing the convolution
The columns of the activation file are:
- Layer: Name of the layer (Note the sizes are the output activation sizes for this layer)
- N: Batch Size (feel free to experiment with this value)
- C: Number of Channels (you can also think of it as depth of Activation)
- X: Width of Activation
- Y: Height of Activation
The idea is to process the network layer by layer. The weight and other parameters of each layer can be obtained by reading each line of the weight csv file. The input and output activation sizes of the corresponding layer can be obtained from the activation csv file.