[Image Captioning] states in lstm() about pytorch-tutorial HOT 2 CLOSED

Training phase

If we don't assign hidden state, nn.LSTM assigns zero state as initial state in default. It is desirable to set the initial state value of nn.LSTM to zero.

Test phase

In hiddens, states = self.lstm(inputs, states), you should note that returned states are used for the next time step. If you do not assign the states such as self.lstm(inputs), it assigns zero states internally and it performs wrong operation.

In test phase, the shape of inputs is (batch_size, 1, embed_size) so that only 1 time step forward propagation is performed when self.lstm(inputs, states) is called.

from pytorch-tutorial.

It makes much sense to me. Thanks for the answer!

from pytorch-tutorial.