[Poly-encoder] Fixes for DSTC7 Task

[klshuster]

Patch description
The DSTC7 Augmented Sampled teacher was using a different train set than the one we used in the Poly-encoder paper; this PR fixes that.

Also a small fix to use the data in shared.

Fix for #2306

Testing steps
I trained a poly-encoder model and got similar results to those in the paper.

Logs

$ python examples/train_model.py -m transformer/poly_encoder -t dstc7:UbuntuDSTC7TeacherAugmentedSampled --init-model zoo:pretrained_transformers/poly_model_huge_reddit/model  --batchsize 256   --model transformer/polyencoder  --warmup_updates 100  --lr-scheduler-patience 0  --lr-scheduler-decay 0.4  -lr 5e-05  --data-parallel True  --history-size 20  --label-truncate 72  --text-truncate 360  -vp 5  -veps 0.5  --validation-metric accuracy  --validation-metric-mode max  --save-after-valid True  --log_every_n_secs 20  --candidates batch  --dict-tokenizer bpe   --dict-lower True  --optimizer adamax  --output-scaling 0.06  --variant xlm  --reduction_type mean  --share-encoders False  --learn-positional-embeddings True  --n-layers 12  --n-heads 12  --ffn-size 3072  --attention-dropout 0.1  --relu-dropout 0.0  --dropout 0.1  --n-positions 1024  --embedding-size 768  --activation gelu  --embeddings-scale False  --n-segments 2  --learn-embeddings True  --share-word-embeddings False  --dict-endtoken __start__  --fp16 True  --polyencoder-type codes  --codes-attention-type basic  --poly-n-codes 64  --poly-attention-type basic  --polyencoder-attention-keys context --eval-batchsize 10 --model-file /tmp/test_model --dict-file data/models/pretrained_transformers/poly_model_huge_reddit/model.dict

.
.
.
[creating task(s): dstc7:DSTC7TeacherAugmentedSampled]
[ running eval: valid ]
[ eval completed in 282.11s ]
valid:{'exs': 5000, 'accuracy': 0.6068, 'f1': 0.6216, 'hits@1': 0.607, 'hits@5': 0.81, 'hits@10': 0.883, 'hits@100': 1.0, 'bleu-4': 0.5299, 'lr': 2e-05, 'total_train_updates': 3537, 'gpu_mem_percent': 0.419, 'examples': 5000, 'loss': 9024.0, 'mean_loss': 1.805, 'mean_rank': 5.001, 'mrr': 0.6998}
[creating task(s): dstc7:DSTC7TeacherAugmentedSampled]
[ running eval: test ]
[ eval completed in 56.18s ]
test:{'exs': 1000, 'accuracy': 0.696, 'f1': 0.7125, 'hits@1': 0.696, 'hits@5': 0.864, 'hits@10': 0.905, 'hits@100': 1.0, 'bleu-4': 0.666, 'lr': 2e-05, 'total_train_updates': 3537, 'gpu_mem_percent': 0.419, 'examples': 1000, 'loss': 1358.0, 'mean_loss': 1.358, 'mean_rank': 3.936, 'mrr': 0.7736}

[codecov]

Codecov Report

Merging #2314 into master will decrease coverage by 7.28%.
The diff coverage is 0%.

@@            Coverage Diff             @@
##           master    #2314      +/-   ##
==========================================
- Coverage   30.64%   23.35%   -7.29%     
==========================================
  Files         595      578      -17     
  Lines       62959    62529     -430     
  Branches    10070    10023      -47     
==========================================
- Hits        19296    14606    -4690     
- Misses      42262    46984    +4722     
+ Partials     1401      939     -462

Flag	Coverage Δ
#mturk_tests	8.53% <0%> (ø)	⬆️
#nightly_gpu_tests	?
#unittests_36	14.88% <0%> (-0.18%)	⬇️
#unittests_37	14.89% <0%> (-0.22%)	⬇️
#unittests_gpu11	13.96% <0%> (-0.19%)	⬇️
#unittests_gpu12	13.96% <0%> (-0.19%)	⬇️

Impacted Files	Coverage Δ
parlai/tasks/dstc7/agents.py	0% <0%> (ø)	⬆️
parlai/tasks/dstc7/build.py	0% <0%> (ø)	⬆️
...zard_of_wikipedia/full_dialogue_retrieval_model.py	0% <0%> (-100%)	⬇️
parlai/zoo/drqa/squad.py	0% <0%> (-100%)	⬇️
parlai/zoo/fasttext_cc_vectors/build.py	0% <0%> (-100%)	⬇️
...arlai/zoo/wizard_of_wikipedia/end2end_generator.py	0% <0%> (-100%)	⬇️
parlai/zoo/convai2/kvmemnn.py	0% <0%> (-100%)	⬇️
projects/controllable_dialogue/tasks/agents.py	0% <0%> (-100%)	⬇️
...ollable_dialogue/controllable_seq2seq/stopwords.py	0% <0%> (-100%)	⬇️
parlai/zoo/controllable_dialogue/build.py	0% <0%> (-100%)	⬇️
... and 97 more

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 0e2d49c...8570557. Read the comment docs.

[emilydinan]

nice!

[emilydinan]

can you add more detail in this docstring? what does augmented mean? what is the difference between augmented and augmented sampled?

[klshuster]

to be honest, @samhumeau would have a better idea of what exactly this means. I will consult him offline (unless he wants to comment here 😄)

[stephenroller]

yeah expand that docstring like emily says

[klshuster]

Updated the docstring after consulting @samhumeau, will merge when tests pass.