PyTorch: nn.LSTM output different results for the same inputs in the same batch

Question

I try to implement a two-layer bidirectional LSTM with torch.nn.LSTM .

I made a toy example: a batch of 3 tensors, which are exactly the same (see my code below). And I expected the outputs of the BiLSTM to be the the same along the batch dimension, ie out[:,0,:] == out[:,1,:] == out[:, 2, :] .

But it seemed not to be the case. According to my experiments, 20%~40% of the time, the output were not the same. So I wonder where I got it wrong.

# Python 3.6.6, Pytorch 0.4.1
import torch

def test(hidden_size, in_size):
    seq_len, batch = 4, 3
    bilstm = torch.nn.LSTM(input_size=in_size, hidden_size=hidden_size, 
                            num_layers=2, bidirectional=True)

    # create a batch with 3 exactly the same tensors
    a = torch.rand(seq_len, 1, in_size)  # (seq_len, 1, in_size)
    x = torch.cat((a, a, a), dim=1)

    out, _ = bilstm(x)  # (seq_len, batch, n_direction * hidden_size)

    # expect the output should be the same along the batch dimension
    assert torch.equal(out[:, 0, :], out[:, 1, :])  
    assert torch.equal(out[:, 1, :], out[:, 2, :])

if __name__ == '__main__':
    count, total = 0, 0
    for h_size in range(1, 51):
        for in_size in range(1, 51):
            total += 1
            try:
                test(h_size, in_size)
            except AssertionError:
                count += 1
    print('percentage of assertion error:', count / total)

Answer 1

What is confusing you is the float precision. Floating point operations are slightly inaccurate and can differ by very small amounts Use this instead:

torch.set_default_dtype(torch.float64) 

Then you will see they should be the same along the batch dim.

Thanks for correcting some English grammar mistakes.

Answer 2

I had the same issue with GRU , and the following solved it for me.
Set a manual seed and set your model in evaluation mode before testing:

torch.manual_seed(42)
bilstm.eval()  # or: bilstm.train(false)

Source: LSTMcell and LSTM returning different outputs

In addition, I had to set the same seed before each call to the model (during testing). In your case:

torch.manual_seed(42)
out, _ = bilstm(x)  # (seq_len, batch, n_direction * hidden_size)