[project] upload new_self_critical project page.

[Minor] clean up the code.

Update readme

Update ADVANCED.md

Allow dbs with temperature.

Fix use_ft and use_att for topdown model.

Fix use_ft and use_att for topdown model.

Merge branch 'self_critical_bottom_up' into self-critical

* self_critical_bottom_up: (42 commits)
  Add advanced. (Still nothing in it.)
  Update readme.
  Sort the features in the forwarding instead of dataloader.
  Add compatibility to resnet features.
  Add comments in Attmodel.
  Make image_root an optional option when prepro_label.
  Add options and verbose for make_bu_data.
  Add cider submodule
  Simplify resnet code.
  Update more to 0.4 version.
  Update to pytorch 0.4
  Fix some in evals.
  Simplify AttModel.
  Update FC Model to the compatible version (previously FC Model is depreacated and not adapted to new structure.)
  Move set_lr to the right place in train.py
  Add max ppl option (beam search sorted by perplexity instead of logprob) (it doens't seem changing too much)
  Fix a bug in ensemble sample.
  Add logit layers option. (haven't reigourously tested if it works or not)
  Allow new ways of computing (using pack sequence) capable of using dataparallel.
  Add batch normalization layer in att_embed.
  ...

# Conflicts:
#	misc/rewards.py
#	train.py

Fix #26.

The code was using the learning_rate from optimizer.pth after starting self critical training.