Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

Splade Encoding and Evaluation not working #150

Open
wants to merge 3 commits into
base: main
Choose a base branch
from
Open

Splade Encoding and Evaluation not working #150

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
19936ea
update encode_splade.py based on latest changes args and module paths
srikanthmalla Sep 4, 2024
4518808
updating splade_encode instructions in README.md
srikanthmalla Sep 4, 2024
65b0fd2
readme for splade query encoding updated
srikanthmalla Sep 4, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
15 changes: 8 additions & 7 deletions examples/splade/README.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -32,23 +32,24 @@ python encode_splade.py \
--model_name_or_path model_msmarco_splade \
--tokenizer_name bert-base-uncased \
--fp16 \
--passage_max_len 128 \
--per_device_eval_batch_size 512 \
--dataset_name Tevatron/msmarco-passage-corpus \
--encode_num_shard 10 \
--encode_shard_index ${i} \
--encoded_save_path encoding_splade/corpus/split${i}.jsonl
--dataset_number_of_shards 10 \
--dataset_shard_index ${i} \
--encode_output_path encoding_splade/corpus/split${i}.jsonl
done

python -m encode_splade.py \
python encode_splade.py \
--output_dir encoding_splade \
--model_name_or_path model_msmarco_splade \
--tokenizer_name bert-base-uncased \
--fp16 \
--q_max_len 128 \
--encode_is_qry \
--query_max_len 128 \
--encode_is_query \
--per_device_eval_batch_size 128 \
--dataset_name Tevatron/msmarco-passage/dev \
--encoded_save_path encoding_splade/query/dev.tsv
--encode_output_path encoding_splade/query/dev.tsv
```

## Index SPLADE with anserini
Expand Down
36 changes: 15 additions & 21 deletions examples/splade/encode_splade.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -16,11 +16,11 @@
HfArgumentParser,
)

from tevatron.arguments import ModelArguments, DataArguments, \
from tevatron.retriever.arguments import ModelArguments, DataArguments, \
TevatronTrainingArguments as TrainingArguments
from tevatron.data import EncodeDataset, EncodeCollator
from tevatron.modeling import EncoderOutput, SpladeModel
from tevatron.datasets import HFQueryDataset, HFCorpusDataset
from tevatron.retriever.dataset import EncodeDataset
from tevatron.retriever.collator import EncodeCollator
from tevatron.retriever.modeling import EncoderOutput, SpladeModel

logger = logging.getLogger(__name__)

Expand Down Expand Up @@ -62,25 +62,19 @@ def main():
config=config,
cache_dir=model_args.cache_dir,
)
encode_dataset = EncodeDataset(
data_args=data_args,
)

text_max_length = data_args.q_max_len if data_args.encode_is_qry else data_args.p_max_len
if data_args.encode_is_qry:
encode_dataset = HFQueryDataset(tokenizer=tokenizer, data_args=data_args,
cache_dir=data_args.data_cache_dir or model_args.cache_dir)
else:
encode_dataset = HFCorpusDataset(tokenizer=tokenizer, data_args=data_args,
cache_dir=data_args.data_cache_dir or model_args.cache_dir)
encode_dataset = EncodeDataset(encode_dataset.process(data_args.encode_num_shard, data_args.encode_shard_index),
tokenizer, max_len=text_max_length)
encode_collator = EncodeCollator(
data_args=data_args,
tokenizer=tokenizer,
)

encode_loader = DataLoader(
encode_dataset,
batch_size=training_args.per_device_eval_batch_size,
collate_fn=EncodeCollator(
tokenizer,
max_length=text_max_length,
padding='max_length'
),
collate_fn=encode_collator,
shuffle=False,
drop_last=False,
num_workers=training_args.dataloader_num_workers,
Expand All @@ -91,15 +85,15 @@ def main():
model.eval()
vocab_dict = tokenizer.get_vocab()
vocab_dict = {v: k for k, v in vocab_dict.items()}
collection_file = open(data_args.encoded_save_path, "w")
collection_file = open(data_args.encode_output_path, "w")

for (batch_ids, batch) in tqdm(encode_loader):
lookup_indices.extend(batch_ids)
with torch.cuda.amp.autocast() if training_args.fp16 else nullcontext():
with torch.no_grad():
for k, v in batch.items():
batch[k] = v.to(training_args.device)
if data_args.encode_is_qry:
if data_args.encode_is_query:
model_output: EncoderOutput = model(query=batch)
reps = model_output.q_reps.cpu().detach().numpy()
else:
Expand All @@ -119,7 +113,7 @@ def main():
print("empty input =>", id_)
dict_splade[vocab_dict[998]] = 1 # in case of empty doc we fill with "[unused993]" token (just to fill
# and avoid issues with anserini), in practice happens just a few times ...
if not data_args.encode_is_qry:
if not data_args.encode_is_query:
dict_ = dict(id=id_, content="", vector=dict_splade)
json_dict = json.dumps(dict_)
collection_file.write(json_dict + "\n")
Expand Down