RNN sequence padding with batch_first #1176
Description
I noticed that using batch_first with pack_padded_Sequence and pad_packed_sequence with an LSTM/GRU doesn't give me the expected output shape.
import torch.nn as nn
from torch.autograd import Variable
from torch.nn.utils.rnn import pad_packed_sequence, pack_padded_sequence
x = Variable(torch.randn(10, 20, 30))
lens = range(20)
lens = [z + 1 for z in lens]
lens = lens[::-1][:10]
x = pack_padded_sequence(x, lens, batch_first=True)
lstm = nn.LSTM(30, 50, batch_first=True)
h0 = Variable(torch.zeros(1, 10, 50))
c0 = Variable(torch.zeros(1, 10, 50))
packed_h, (packed_h_t, packed_c_t) = lstm(x, (h0, c0))
h, _ = pad_packed_sequence(packed_h)
print h.size() # Size 20 x 10 x 50 instead of 10 x 20 x 50