pytorch "RuntimeError: ~ inplace operation" for CUDA

def __init__(self, input_size, hidden_size, num_layers=2):

super().__init__()

self.lstm = nn.LSTM(input_size, hidden_size, num_layers)

def forward(self, features, init_hidden=None):

self.lstm.flatten_parameters()

RuntimeError: one of the variables needed for gradient computation has been modified by an inplace operation

Move the self.lstm.flatten_parameters() in def foward() to def __init__.

It has no effect on CPU, but on CUDA it will modify the LSTM's parameters part way through the computation which would break backward (leading to the exception). (reference: https://github.com/pytorch/pytorch/issues/7243)