liuzh91 commented on issue #16930: Plug and Play fit_batch() for the estimator class URL: https://github.com/apache/incubator-mxnet/issues/16930#issuecomment-559403015 > Why don't we just pack the input and unpack it when we feed it into the model? The order of input will be determined by the user, i.e, how they construct the dataloader. The RNN model specific input `begin_state` will always serve as a keyword argument. The problem is even we can use pack/unpack workaround to bypass the input/output issue. The computation of loss function may still be diverse. For example, for `standardRNN`, one has: ``` output, h, encoder_hs, dropped_encoder_hs = model(X, h) l = joint_loss(output, y, encoder_hs, dropped_encoder_hs) Ls.append(l / (len(context) * X.size)) hiddens[j] = h ``` On the other hand, for `BigRNN`, one has ``` output, hidden, new_target = self._model(X, y, h, s) output = output.reshape((-3, -1)) new_target = new_target.reshape((-1,)) ls = self._loss(output, new_target) * m.reshape((-1,)) ls = ls / args.batch_size ``` In this case, we may resort to if/else clause to judge which computation routine we shall use. The code will be cumbersome in the future.
---------------------------------------------------------------- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: [email protected] With regards, Apache Git Services
