SyrineKrichene
commented
2 years ago Hi apurvak, The table.tsv contains: table_id and the embedding extracted by the dual encoder.
Example of dual encoder to use:
The dual encoder model
! gsutil cp "gs://tapas_models/2021_04_27/tapas_nq_hn_retriever_medium.zip" "tapas_retriever.zip" && unzip tapas_retriever.zip
You can run:
! python -m tapas.experiments.table_retriever_experiment \
--do_predict \
--eval_name="dual_encoder_tables" \
--minutes_to_sleep_before_predictions=0 \
--num_eval_steps=0 \
--model_dir="
The output will be written in "output_dir/predict_results_0.tsv"
In case you didn't generate the table.tfrecord yet, you can use functions similar to the following ones:
1- This function is to get the interactions: def get_table(document_title, table_data): """Extracts the interaction for an str table. Args: table_data: str table where the columns are separated by '|' and rows by ' \n' document_title: str title of the page containing the table or a table title it also could be empty str.""" table = [list(map(lambda s: s.strip(), row.split("|"))) for row in table_data.split("\n") if row.strip()] table_interaction = interaction_pb2.Table() table_interaction.document_title = document_title
Please add a meaningful, unique table_id: I put as a default the
docment_title table_interaction.table_id = document_title if not table: return table_interaction for header in table[0]: table_interaction.columns.add().text = header for line in table[1:]: row = table_interaction.rows.add() for cell in line: row.cells.add().text = cell return table_interaction
2- To write the interactions as tf_records. You can use: def write_tfrecord(filename, examples): """From interactions examples to tfrecord.""" with tf.io.TFRecordWriter(filename) as writer: for example in examples: writer.write(example.SerializeToString())
3- You need to extract the tf_examples. You can use: def extract_tables_tf_examples(path_to_input_tables): """Extracts tf_examples from interactions.tfrecord."""
PLease use the max_seq_length used by the loaded dual encoder model. Most
of the models use 512.
config=tf_example_utils.RetrievalConversionConfig(
vocab_file="
On Sun, Nov 21, 2021 at 7:12 PM apurvak @.***> wrote:
I am trying to evaluate the NQ Table model for my own data. To that end, I ran tapas/retrieval/create_retrieval_data_main.py script to generate the tables.tfrecord. However, the main notebook uses the following code to get the TSV file (that looks like embeddings of table) for retrieval model to get nearest neighbours:
def get_nearest_neighbors(num_neighbors):
….
tables = eval_table_retriever_utils.read_tables("results/nq_retrieval/model/tables.tsv", make_tables_unique=False)
….
How do I generate the tables.tsv file for my data? - the file for prepared data looks like this:
— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/google-research/tapas/issues/148, or unsubscribe https://github.com/notifications/unsubscribe-auth/APARZOPT2O5Q24S3LATN22DUNEY75ANCNFSM5IPNIS3Q . Triage notifications on the go with GitHub Mobile for iOS https://apps.apple.com/app/apple-store/id1477376905?ct=notification-email&mt=8&pt=524675 or Android https://play.google.com/store/apps/details?id=com.github.android&referrer=utm_campaign%3Dnotification-email%26utm_medium%3Demail%26utm_source%3Dgithub.
apurvak
I am trying to evaluate the NQ Table model for my own data. To that end, I ran
tapas/retrieval/create_retrieval_data_main.pyscript to generate the tables.tfrecord. However, the main notebook uses the following code to get the TSV file (that looks like embeddings of table) for retrieval model to get nearest neighbours:How do I generate the tables.tsv file for my data? - the file for prepared data looks like this: