Batch is None when I use the GPU

Hi everyone.
My model runs okay on the CPUs. But when I use the GPU, the validation step is not getting any batch of tensors. The validation step has None in the batch parameter. Have you encountered this problem before? Any pointers would be helpful


Answer from slack (thanks @awaelchli):

Your code was the following:

import torch.nn as nn
import torch
from torch.nn.utils.rnn import pack_padded_sequence, pad_packed_sequence
class LSTMEncoder(nn.Module):
    """ simple wrapper for a bi-lstm """
    def __init__(
        emb_dim: int,
        hidden_dim: int,
        layers: int,
        bidirectional: bool,
        dropout: float,
        super(LSTMEncoder, self).__init__()
        self.num_directions = 2 if bidirectional else 1
        self.lstm = nn.LSTM(
            hidden_dim // self.num_directions,
        self.pack = pack
    def init_state(self, input):
        batch_size = input.size(0)  # retrieve dynamically for decoding
        h0 = torch.zeros(
            self.lstm.num_layers * self.num_directions,
        c0 = torch.zeros(
            self.lstm.num_layers * self.num_directions,
        return h0, c0
    def forward(self, src_embedding, srclens, srcmask, temp=1):
        h0, c0 = self.init_state(src_embedding)
        if self.pack:
            inputs = pack_padded_sequence(src_embedding, srclens, batch_first=True)
            inputs = src_embedding
        outputs, (h_final, c_final) = self.lstm(inputs, (h0, c0))
        if self.pack:
            outputs, _ = pad_packed_sequence(outputs, batch_first=True)
        # outputs: batch, seq_len num_directions * hidden_size
        # h_n: num_layers * num_directions, batch, hidden_size
        # c_n: num_layers * num_directions, batch, hidden_size
        return outputs, (h_final, c_final)

Lightning will automatically move parameter weights from the model and data from the DataLoaders to the appropriate device, but it cannot move tensors created in the forward() function. The fix here is easy:

h0 = torch.zeros(..., device=input.device)
c0 = torch.zeros(..., device=input.device)

By specifying a the appropriate device for the new tensors, you won’t run into this issue!

1 Like

Correct! However, I could not solve the original problem. I was using a DataModule before and it still returns None for the batch . I changed it back to use DataLoader in the function. Moved the tensors to GPU. (Tensors created within a model will have to be moved by the code and Lightning wont do it.
). Thanks to @awaelchli for working with me to resolve this. Cheers!

1 Like

Update: The DataLoader has a function called transfer_batch_to_device which is an abstract method. Do not implement this, if you do not need to. I got None because I included the function in the DataModule, but returned None from it.