takes input and output both as sequence (for machine transaltion)

why seq2seq is difficult?

1. input sentence (eng) of variable length
2. output sentence(hindi) of variable length
3. we have studied to handle these variable length only for input in (lstm and gru’s)

encoder decoder came to solve this variable length problem.

Training the arch using backpropagation ➖

encoder → context vector → decoder

this is what inside encoder(a set of lstm cell)

mostly used are lstm and gru’s

(used to produce timestamp)

• here params(ht&ct) are intialized and then
• each sentence is encoded (by one hot encoding ) which gives tokens [1 0 0 0 0 ]
• these tokens are then added to lstm where they update the value of params