Pack_padded_sequence with unsorted question length

nlp
1

Hi, I am the first time using the pack_padded_sequence , and it seems to input the sequence length in a decreasing order. Here is the default pack_padded_sequence package from official website

torch.nn.utils.rnn.pack_padded_sequence(input, lengths, batch_first=False, enforce_sorted=True)

  • enforce_sorted: if True , the input is expected to contain sequences sorted by length in a decreasing order. If False , the input will get sorted unconditionally. Default: True

It also mentions that enforce_sorted = True is only necessary for ONNX export., what is this mean? If I want to input the length without sorting, can I simply set it to False?

below is my idea of implement

# in dataset class; 
#return the question idx after padding and the original question length
class VQADataset(Dataset):
    def __getitem__(self, idx):
        return sample{'question': question, 'length':len(qst_token)}
    
# in model class
class model(nn.Module):
    def forward(self, qst, qst_length):
        word = self.embedding(qst) 
        qst_length = qst_length.tolist()             
        packed = rnn.pack_padded_sequence(word , qst_length, batch_first=True)

So, the length of question will be different every batch. can I simply set the enforce_sorted into False ?

2

Yes, if you are not planning to export the model via ONNX, you could use enforce_sorted=False and pass the unsorted padded sequence to this method.

3

This is the explaination from the internet. ONNX (Open Neural Network Exchange) is an open container format for the exchange of neural network models between different frameworks.

So, if I only use pytorch and not convert into tensorflow. I can simply set it to False, right?

4

Yes, if you use PyTorch without any intention to use ONNX, you can set this argument to False.