* Update README.md

* minor fix

* Typo fix

* typo fix

…916)

* update boundry func with offsets

* update tasks that use indexes

* remove outdated temporary fix

* Initial QASRL

* Updated pred writing for QASRL

* Add validation shuffle to QASRL

* Remove tqdm, modify class check in preds

* qasrl rebase cleanup

* Update QA-SRL to new repo changes

* Removing src

* QASRL Cleanup

* updating to new model format

* csv to tsv

* QASRL update

* implemented data parallel

* black style

* Resolve last of merge marks

* deleting irrelevant logs

* adding new way to get attribute

* updating to master

* torch.Tensor -> torch.tensor for n_exs

* black style

* black style

* Merge master

* adapting other tasks to multiple GPU"

* adding helper function for model attributes

* adding get_model_attribute to main.py

* deleting unecessary n_inbput for span_module

* black style

* revert comment change

* fixing batch size keys

* opt_params -> optimizer_params

* Remove extraneous cahnges

* changed n_exs to one-liner

* adapting args.cuda to multi-GPU setting

* adding use_cuda variable

* Fixing parsing for case of args.cuda=subset

* fixing tests

* fixing nits, cleaning up parse_cuda function

* additional nit

* deleted extra space

* Revert nit

* refactoring into get_batch_size

* removing use_cuda

* adding options.py

* removing use_cuda in tests, deleting extra changes

* change cuda default

* change parse_cuda_list_args import

* jiant.options -> jiant.utils.options

* change demo.conf cuda setting

* fix bug -> make parse_cuda return int if only one gpu

* fix bug

* fixed tests

* revert test_retokenize change

* cleaning up code

* adding addiitonal jiant.options

* Separating cuda_device = int case with multiple cuda_device case

* deleting remains of uses_cuda

* remove time logging

* remove use_cuda from evaluate

* val_interval -> validation_interval

* adding cuda comment to tutorial

* fixed typo

* replace correct_sent_indexing with non inplace version

* Update modules.py

* Update modules.py

* anli

* anli fix

* Adding aNLI link, additional test/dev warning

* black style

* adding SocialQA

* black style

* black style

* fixed socialQA task

* black style

* Update citation

* Nit

* senteval

* socialIQA naming

* reverse unnecessary add

…rsing related code (#928)

* black style

* remove

* cleaning up code around cuda-parsing

* adding defaulting to -1 if there is no cuda devices detected

* fixing nits, throw error instead of log warning for cuda not found

* black style

* adding initial senteval, senteval preprocessing script

* black

* adding senteval to registry

* fixing bigram-shift

* adding label_namespace arg, fixing the ksenteval tasks

* revert extra changes

* black style

* change name -> senteval-probing

* fixing senteval-probing tasks

* renamed senteval -> sentevalprobing

* delete extra imports

* black style

* renaming files and cleaning up preprocessing code

* nit

* black

* deleting pdb

* Senteval -> SE shorthand

* fixing code style

* black style

* pre-loading tokenizer before retokenization function

* scitail

* Scitail

* Scitail

* update Scitail, removed config

* update Scitail, removed config

Thanks to #936, we've discovered errors in our data statistics reporting in the edge probing paper. This table contains the corrected values. As there is more space here, the full (unrounded) values are reported instead. This was generated by a script that read the stats.tsv file and the diff vs. the paper should match my comment on the issue yesterday.

* add commonsenseqa task

* add hellaswag task

* dabug

* from #928

* add special tokens to CommensenseQA input

* format

* revert irrelevant change

* Typo fix

* delete

* rename stuff

* Update qa.py

* black

* generalize ccg to other transformer models

* debug

* I don't know who broke this at what time, but let's just fix it here now

* Copying configs from superglue

* adding senteval probing config commands

* adding meta-script for transfer and probing exps

* Adding meta bash script fixed

* give_permissions script

* small fix transfer_analysis.sh (#946)

model_*.th might indicate several models; fixed to model_*.best.th

* lr_patience fix

* target_task training -> pretrain training

* adding edgeprobing configs and command

* adding edge probing conf

* fix load_target_train bug

* add hyperparameter sweeping

* val_interval change

* adding sweep function

* Task specific val_intervals

* add reload_vocab to hyperparameter sweep

* adding batch_size specification

* fixing senteval-word-content

* fixing senteval preprocess script

* revert extra delete

* remove extra files

* black format

* black formatting trainer.py

* remove load_data()

* removing extra changes

* Copying configs from superglue

* adding senteval probing config commands

* adding meta-script for transfer and probing exps

* Adding meta bash script fixed

* give_permissions script

* small fix transfer_analysis.sh (#946)

model_*.th might indicate several models; fixed to model_*.best.th

* lr_patience fix

* target_task training -> pretrain training

* adding edgeprobing configs and command

* adding edge probing conf

* fix load_target_train bug

* add hyperparameter sweeping

* val_interval change

* adding sweep function

* Task specific val_intervals

* add reload_vocab to hyperparameter sweep

* adding batch_size specification

* fixing senteval-word-content

* fixing senteval preprocess script

* revert extra delete

* remove extra files

* black format

* black formatting trainer.py

* remove load_data()

* removing extra changes

* adding alignment mapping function

* fix comment nits

* comment nit

* adding example of token_alignment

* add nli prob task template

* Create acceptablity_probing.py

* specify nli probing tasks

* port acceptablity probing tasks

* add directory name

* debug

* debug

* format

* black

* revert unintended change

* misc run scripts

* cosmosqa

* cosmosqa

* cosmosqa

* cosmosqa run

* cleaned up repo

* cleaned up repo

* reformatted

* qamr

* tokenization

* temp qamr

* qamr

* QASRL

* Undo slicing

* quick hack to bypass bad qasrl examples

* f1 em fix

* tokenization fixes

* average

* New tokenization aligner

* update example counts

* Cleanup

* Typography

Currently the `_unk_id` for Roberta is not set correctly, which triggers the assertion error on line 118.

* adding loading for mix

* black style

* generalize ccg to other transformer models

* debug

* I don't know who broke this at what time, but let's just fix it here now

* ccg lazy iterator

* debug

* clean up

* debug

* debug ccg, minor cleanup