2016 08 08 - HenglinShi/LSTM_LIP_READING GitHub Wiki

#Task of the day

• Adding layers: one convolution, one relu, one pooling, and one dropOut
• Change input data, to light flow image

#Results ##Adding layers samples-->Convolution-->ReLU-->Pooling-->InnerProduct-->ReLU-->DroupOut----> ----> LSTM--------> clip_markers-------------->Reshape-----------------------------------------> - Loss and accuracy - labels-------------------->Reshape-------------------------------------------------------->

• Result does not looks good, average accuracy is 13%

#Before practicing the second task of the data, sample normalization is tried

• Using batch normalization
• Reference: https://arxiv.org/abs/1510.01378