Layers MobileNet V1 weight analysis

Press ' - key to Hide/Show the right "Model" window. Set new layer number.

layer

Layer model from 'https://storage.googleapis.com/tfjs-models/tfjs/mobilenet_v1_1.0_224/model.json' is used. Channel weights are shown similar to Data-Free Quantization.

Comments

MobileNet V1 consists of a regular 3×3 convolution as the very first layer, followed by 13 times the depthwise separable convolution blocks (shown below). Batch Norm equations are shown below.

Layers model has too many data used for training. For efficient inference we shall fold convolution and batch normalization layers (and batch norm constants) next.

To Do

1. Merge BN-weights.
2. Implement "Data-Free Quantization" :)
TFjs notes     updated 31 Dec 2019