Same LSTM(GRU) implementation different results (pytorch& keras)

[ptrblck]

You can use the torch.nn.init methods to initialize the parameters as shown in e.g. this post.