I was going over through the masking code in the chatbot tutorial and noticed that it masks with a zero on indices that are 0 but are NOT padding tokens (e.g. the first token). Is that a bug? Is the fix to use the lengths of the sequences to pad?
# Returns padded target sequence tensor, padding mask, and max target length
def outputVar(l, voc):
'''
padVar = padded (transposed) list of sentences with the batches
tensor([[1391, 188, 122, 53, 5091],
[ 4, 53, 12, 154, 7708],
[ 2, 3026, 1048, 747, 4],
[ 0, 4, 115, 5747, 2],
[ 0, 2, 12, 2281, 0],
[ 0, 0, 1048, 4, 0],
[ 0, 0, 4, 2, 0],
[ 0, 0, 2, 0, 0]])
mask = mask indicating where words occur and what isn't a word (i.e 0 for padding)
tensor([[1, 1, 1, 1, 1],
[1, 1, 1, 1, 1],
[1, 1, 1, 1, 1],
[0, 1, 1, 1, 1],
[0, 1, 1, 1, 0],
[0, 0, 1, 1, 0],
[0, 0, 1, 1, 0],
[0, 0, 1, 0, 0]], dtype=torch.uint8)
max_target_len = length of longest target sentence
max_target_len = 8
'''
# list of index representation of sentence [[124, 101, 102, 4401, 98, 382, 4, 2], ..., [67, 188, 38, 4, 2]]
indexes_batch = [indexesFromSentence(voc, sentence) for sentence in l]
# get length of the largest (target) sentence
max_target_len = max([len(indexes) for indexes in indexes_batch])
# (transposed) list of index represtations of sentences with padded zeros at the end [(124, 25, 25, 218, 67), ..., (4, 2, 0, 0, 0), (2, 0, 0, 0, 0)]
padList = zeroPadding(indexes_batch) # padds with zeros sentences that are too long
# returns the mask indicating which position are words locations and marks with 0 which ones are simply zeros
st()
mask = binaryMatrix(padList)
#mask = torch.Tensor(padList) != PAD_token ## ALSO BUGGY?!
mask = torch.ByteTensor(mask)
# tensorfy (transposed) list of index represtations of sentences with padded zeros at the end. The last list is now a tensor/matrix
padVar = torch.LongTensor(padList)
return padVar, mask, max_target_len