Cuda 11 pytorch compatability?

[markobogoevski]

How can I run this on a Cuda11 graphics card which cant be downgraded to cuda10? There is no binary yet for torch 1.5 for cuda11 and the code isn't compatible for torch1.7 (earliest compatible torch version for cuda11). Will there be an updated version of the code for torch 1.7 anytime soon?

[risto-trajanov]

Migration from 1.5.1 to 1.7.1 torch in PICK:

Error: RuntimeError: 'lengths' argument should be a 1D CPU int64 tensor

nn.utils.rnn.pack_padded_sequence-> expects 1D cpu int64 tensor but we give him LongTensor in previous version 1.5.1 lenghts argument was converted to cpu() in the function but now we need to convert it outside of it. That fixes the problem of migration.

In decoder.py in forward function

 def forward(self, x_seq: torch.Tensor,
            lenghts: torch.Tensor,
            initial: Tuple[torch.Tensor, torch.Tensor]):
    '''

    :param x_seq: (B, N*T, D)
    :param lenghts: (B,)
    :param initial: (num_layers * directions, batch, D)
    :return: (B, N*T, out_dim)
    '''

    # B*N, T, hidden_size
    x_seq, sorted_lengths, invert_order, h_0, c_0 = self.sort_tensor(x_seq, lenghts, initial[0], initial[0])
    packed_x = nn.utils.rnn.pack_padded_sequence(x_seq, lengths=sorted_lengths.cpu(), batch_first=True)
    self.lstm.flatten_parameters()
    output, _ = self.lstm(packed_x)
    output, _ = nn.utils.rnn.pad_packed_sequence(output, batch_first=True,
                                                 padding_value=keys_vocab_cls.stoi['<pad>'])
    # total_length=documents.MAX_BOXES_NUM * documents.MAX_TRANSCRIPT_LEN
    output = output[invert_order]
    logits = self.mlp(output)
    # (B, N*T, out_dim)
    return logits