* BUG: Fix loss and metric calculation

* Sync MSELoss with MXNet L2Loss

* Sync MSELoss with MXNet L2Loss undefit-overfit

* Sync MSELoss with MXNet L2Loss softmax-regression-scratch

* Update d2l lib for train_epoch_ch3

* Use reduction=none minibatch sgd

* Update d2l lib for train_concise_ch11

* Fix loss computation in weight decay train_concise

* drop 1/2 factor attention

* add eqmse label

* drop 1/2 underfit-overfit

* drop 1/2 softmax-scratch

* drop 1/2

* drop 1/2 tf

* drop 1/2 tf minibatch sgd

* 0.17.1

* translate minibatch sgd

* translate sequence
Co-authored-by: NAston Zhang <asv325@gmail.com>

Co-authored-by: NUbuntu <ubuntu@ip-172-31-12-66.us-west-2.compute.internal>

* align exercise with the english version

* fix

Co-authored-by: NUbuntu <ubuntu@ip-172-31-12-66.us-west-2.compute.internal>

* fix typo

* fix typo 我妈 > 我们

* in-place

* in-place

* Fix: Epoch instead of batch

* Epoch -> epoch

* index

* finetuning-bert

* sentiment

* revert sentiment rnn

* revert sentiment rnn

* sentiment-analysis-rnn

* natural-language-inference

* sentiment-analysis-rnn

This reverts commit 63d6597f.

* appendix tool soft polish

* appendix tool soft polish

* contribute image translate

* chapter 14 polish half

* chapter 14 soft polish done

* chapter12 0-3 polish

* soft polish chapter 12 done