What kinds of input take a pytorch model?

Hello everyone!

I’m trying to do the pytorch tutorials which start here:

Until now the examples given were quite simple. All models I encountered in the tutorial until reaching this page:
took only one input and it was always a Variable containing a tensor.

But then I tried to do the exercise called “Augmenting the LSTM part-of-speech tagger with character-level features” presented on this page:
and I realized that I needed three different inputs.

To put more context, the model takes one sentence of english at the time. I would like to have the following inputs

  1. a list containing the length of the words of the sentence
  2. a one dimensional LongTensor containing the indexes corresponding to the words of my sentence (I have a dictionary encoding all the words for the vocabulary into indexes).
  3. a two dimensional LongTensor containing for each word the indexes corresponding to the letters of the word (with padding to have a vector of same length for every word).

For now the first argument is given as a Python list, and the two others as Variable containing torch.LongTensor. I know that I could also put my list in a LongTensor, but I would like to know if it is really required.
I suspect that I have to because I ran my model and I obtain the following error:
"forward() missing 1 required positional argument: ‘input’ "
and my forward method has been defined to take as second argument (after ‘self’), the list of length (and in particular no argument called ‘input’).

You can give your model any data type you like so long as your .forward method knows what to do with it.

The error you are seeing is likely related to an error inside your forward method.

Could you post the code for the forward method and the full stack trace of the error?

1 Like

Thank you very much for your answer jpeg729!
I didn’t post my code in my first message because it is a bit long. You answered my main question anyway, and I will try to find out the reason of this error by myself. But if you want to look at my code I put it here:

class LSTMTaggerCharLevelFeature(nn.Module):
    def __init__(self, w_emb_dim, t_hidden_dim, c_emb_dim,
                c_hidden_dim, n_word, n_char, tagset_size):
        super(LSTMTaggerCharLevelFeature, self).__init__()
        self.w_emb_dim = w_emb_dim
        self.t_hidden_dim = t_hidden_dim
        self.c_emb_dim = c_emb_dim
        self.c_hidden_dim = c_hidden_dim
        self.w_embeddings = nn.Embedding(n_word, w_emb_dim)
        self.c_embeddings = nn.Embedding(n_char, c_emb_dim)
        # The LSTM for embedded characters.
        self.c_lstm = nn.LSTM(c_emb_dim, c_hidden_dim)
        # The LSTM for the concatenation of the embedded
        # word and the output of (the last run) of the
        # first LSTM.
        self.t_lstm = nn.LSTM(w_emb_dim+c_hidden_dim, t_hidden_dim)
        # The linear layer that maps the total output
        # of the second LSTM to the tag space
        self.hidden2tag = nn.Linear(t_hidden_dim, tagset_size)
        self.c_hidden = self.init_hidden(1, c_hidden_dim)
        self.t_hidden = self.init_hidden(1, t_hidden_dim)

    def init_hidden(self, batch, hidden_dim):
        # For initialization of an LSTM
        return(autograd.Variable(torch.zeros(1, batch, hidden_dim)),
               autograd.Variable(torch.zeros(1, batch, hidden_dim)))
    def forward(self, list_len, w_ix, c_ix):
        self.c_hidden = self.init_hidden(len(list_len), c_hidden_dim)
        w_embeds = self.w_embeddings(w_ix)
        c_embeds = self.c_embeddings(c_ix)
        ## First LSTM on embedded characters
        out, self.c_hidden = self.c_lstm(c_embeds.permute(1, 0, 2), self.c_hidden)
        ## Selecting the output corresponding
        ## to the last real character
        last_char_ix = [list_len[i]-1 for i in range(len(list_len))]
        kept_lstm_output_ix = [[[last_char_ix[i] for j in range(self.c_hidden_dim)]
                              for i in range(len(list_len))]]
        kept_lstm_output_ix = autograd.Variable(
        out_c_lstm = out.gather(dim=0, 
        ## Creation of the input of the second LSTM
        in_t_lstm = torch.cat((w_embeds, out_c_lstm), 1)
        in_t_lstm = in_t_lstm.view(list(in_t_lstm.size())[0],
                                   1, list(in_t_lstm.size())[1])
        ## Second LSTM on aggregation of output of the first LSTM
        ## and embeddings of the words
        out_t_lstm, self.t_hidden = self.t_lstm(in_t_lstm, self.t_hidden)
        tag_space = self.hidden2tag()
        tag_scores = F.log_softmax(tag_space, dim=1)

And here is the full stack trace of the error:

TypeError                                 Traceback (most recent call last)
<ipython-input-29-287341512899> in <module>()
     17         # Step 3. Run our forward pass.
---> 18         tag_scores = model(list_len, w_ix, c_ix)
     20         # Step 4. Compute the loss, gradients, and update the parameters by

~/anaconda3/envs/mypy36/lib/python3.6/site-packages/torch/nn/modules/module.py in __call__(self, *input, **kwargs)
    355             result = self._slow_forward(*input, **kwargs)
    356         else:
--> 357             result = self.forward(*input, **kwargs)
    358         for hook in self._forward_hooks.values():
    359             hook_result = hook(self, input, result)

<ipython-input-25-afab1327140d> in forward(self, list_len, w_ix, c_ix)
     50         ## and embeddings of the words
     51         out_t_lstm, self.t_hidden = self.t_lstm(in_t_lstm, self.t_hidden)
---> 52         tag_space = self.hidden2tag()
     53         tag_scores = F.log_softmax(tag_space, dim=1)
     54         return(tag_scores)

~/anaconda3/envs/mypy36/lib/python3.6/site-packages/torch/nn/modules/module.py in __call__(self, *input, **kwargs)
    355             result = self._slow_forward(*input, **kwargs)
    356         else:
--> 357             result = self.forward(*input, **kwargs)
    358         for hook in self._forward_hooks.values():
    359             hook_result = hook(self, input, result)

TypeError: forward() missing 1 required positional argument: 'input'

self.hidden2tag() has no input.

1 Like

Thank you so much!
I corrected this and the errors disappeared!