the last ouput of Bi-RNN in TextRNN

[longbowking]

text_classification/a03_TextRNN/p8_TextRNN_model.py

Line 67 in 68e2fcf

self.output_rnn_last=tf.reduce_mean(output_rnn,axis=1) #[batch_size,hidden_size*2] #output_rnn_last=output_rnn[:,-1,:] ##[batch_size,hidden_size*2] #TODO

In the implementation, the final outputs of Bi-RNN are calculated as the reduce mean among all time stamps. Compared with output_rnn_last=output_rnn[:,-1,:], what is the difference between these two strategies on the impact of the final classification results?

[brightmart]

hi,
you post a good question. the difference is whether you use mean average of all timestamp or use last timestamp to make a prediction; the other way is use reduce_max to make a prediction.

so you can choose the way that the best performance among these three methods.