search

Tencent / PocketFlow

An Automatic Model Compression (AutoMC) framework for developing smaller and faster AI applications.
https://pocketflow.github.io
Other
2.78k stars 490 forks source link

UniformQuantLearner, if Net is full Conv net, no active layer, how this method work #296

Open tianylijun opened 5 years ago

tianylijun commented 5 years ago

class UniformQuantization:

TODO: add more types of activations and matmuls

self.support_act_types = ['Relu', 'Relu6', 'Crelu', 'Elu', 'Selu', 'Softplus',\
    'Softsign', 'Sigmoid', 'Tanh']
self.support_mul_types = ['Conv2D', 'MatMul', 'DepthwiseConv2dNative']

a common problem, we want do quantization on conv layer, not only on active layer. Worst case is a full Conv Net(no active layer), how this method work?