def sorted_rnn(self, sequences, sequence_lengths, rnn):
sorted_inputs, sorted_sequence_lengths, restoration_indices = sort_batch_by_length(sequences, sequence_lengths)
# print(sorted_inputs)
# print(sorted_sequence_lengths.data.tolist())
packed_sequence_input = pack_padded_sequence(sorted_inputs,
list(map(int, sorted_sequence_lengths.data.tolist())),
batch_first=True)
packed_sequence_output, _ = rnn(packed_sequence_input, None)
unpacked_sequence_tensor, _ = pad_packed_sequence(packed_sequence_output, batch_first=True)
return unpacked_sequence_tensor.index_select(0, restoration_indices)
def sorted_rnn(self, sequences, sequence_lengths, rnn):
sorted_inputs, sorted_sequence_lengths, restoration_indices = sort_batch_by_length(
sequences, sequence_lengths)
packed_sequence_input = pack_padded_sequence(
sorted_inputs,
sorted_sequence_lengths.data.tolist(),
batch_first=True) #压缩：去除多余的PAD的影响
packed_sequence_output, _ = rnn(packed_sequence_input, None)
unpacked_sequence_tensor, _ = pad_packed_sequence(
packed_sequence_output, batch_first=True) #解压
return unpacked_sequence_tensor.index_select(
0, restoration_indices) #返回复原后的经过LSTM的embed
def sorted_rnn(self, sequences, sequence_lengths, rnn):
sorted_inputs, sorted_sequence_lengths, restoration_indices = sort_batch_by_length(
sequences, sequence_lengths)
packed_sequence_input = pack_padded_sequence(
sorted_inputs,
sorted_sequence_lengths.data.long().tolist(),
batch_first=True)
# except RuntimeError:
# print(sequences)
# print(sequence_lengths)
# exit()
packed_sequence_output, _ = rnn(packed_sequence_input, None)
unpacked_sequence_tensor, _ = pad_packed_sequence(
packed_sequence_output, batch_first=True)
return unpacked_sequence_tensor.index_select(0, restoration_indices)
