cuda_functional is required

[vinhqdang]

Hi there

When I tried to follow the example, it seems that the package "cuda_functional" is required. However, this package requires GPU and CUDA.

Is there a way that I can perform your code on CPU only?

Thanks

[chickenbestlover]

Hi vinhqdang,

cuda_functional is only required when you want to use SRU instead of LSTM. I modified model.py so that you don't have to from cuda_functional import SRU when you use LSTM. (I simply moved the location of from cuda_functional import SRU like this: .

class RNNPredictor(nn.Module):
    """Container module with an encoder, a recurrent module, and a decoder."""

    def __init__(self, rnn_type, enc_inp_size, rnn_inp_size, rnn_hid_size, dec_out_size, nlayers, dropout=0.5,
                 tie_weights=False,res_connection=False):
        super(RNNPredictor, self).__init__()
        self.enc_input_size = enc_inp_size

        self.drop = nn.Dropout(dropout)
        self.encoder = nn.Linear(enc_inp_size, rnn_inp_size)
        if rnn_type in ['LSTM', 'GRU']:
            self.rnn = getattr(nn, rnn_type)(rnn_inp_size, rnn_hid_size, nlayers, dropout=dropout)
        elif rnn_type == 'SRU':
            from cuda_functional import SRU
            self.rnn = SRU(input_size=rnn_inp_size,hidden_size=rnn_hid_size,num_layers=nlayers,dropout=dropout,
                           use_tanh=False,use_selu=True,layer_norm=True)
        else:
          '''

[drillep]

Attempting to run the code on cpu with LSTM default; python 1_train_predictor.py --data nyc_taxi --filename nyc_taxi.pkl --device cpu

I've run this model successfully on the GPU but when switching to CPU I have the following error.

RuntimeError: invalid argument 2: view size is not compatible with input tensor's size and stride (at least one dimension spans across two contiguous subspaces). Call .contiguous() before .view(). at /pytorch/aten/src/TH/generic/THTensor.cpp:237

[chickenbestlover]

@drillep See #2.