add monot5 tpu train doc (#108)

* add monot5 tpu train doc * Update pygaggle/data/create_msmarco_t5_training_pairs.py Co-authored-by: Rodrigo Frassetto Nogueira <[email protected]> * update tpu experiment link to README and add trainning time info in tpu doc Co-authored-by: Rodrigo Frassetto Nogueira <[email protected]>
castorini · Nov 11, 2020 · da0fb61 · da0fb61
1 parent 5e1e0dd
commit da0fb61
Show file tree

Hide file tree

Showing 3 changed files with 81 additions and 0 deletions.
diff --git a/README.md b/README.md
@@ -90,3 +90,4 @@ The following documents describe how to use Pygaggle on various IR test collecti
 + [Experiments on MS MARCO Document Retrieval](https://github.com/castorini/pygaggle/blob/master/docs/experiments-msmarco-document.md)
 + [Experiments on MS MARCO Passage Retrieval - Dev Subset](https://github.com/castorini/pygaggle/blob/master/docs/experiments-msmarco-passage-subset.md)
 + [Experiments on MS MARCO Passage Retrieval - Entire Dev Set](https://github.com/castorini/pygaggle/blob/master/docs/experiments-msmarco-passage-entire.md)
++ [Experiments on MS MARCO Passage Retrieval - with TPU](https://github.com/castorini/pygaggle/blob/master/docs/experiments-monot5-tpu.md)
diff --git a/docs/experiments-monot5-tpu.md b/docs/experiments-monot5-tpu.md
@@ -211,4 +211,63 @@ You should see the same result.
 
 If you were able to replicate these results, please submit a PR adding to the replication log! Please mention in your PR if you find any difference!
 
+## Train monoT5
+
+First, download the MS MARCO train triples:
+```
+wget https://storage.googleapis.com/duobert_git/triples.train.small.tar.gz
+tar -xvf triples.train.small.tar.gz
+rm triples.train.small.tar.gz
+```
+
+Then convert the train triples to t5 input format:
+```
+python -m pygaggle.data.create_msmarco_t5_training_pairs --triples_train triples.train.small.tsv --output_to_t5 query_doc_pairs.train.tsv
+```
+
+Next, copy the input file to Google Storage. TPU training will read data directly from `gs`
+```
+gsutil cp query_doc_pairs.train.tsv ${GS_FOLDER}/query_doc_pairs.train.tsv
+```
+
+Recall the environment variables
+```
+export MODEL=<t5 pretrain model, e.g. base, large, 3B>
+export GS_FOLDER=<gs folder to store checkpoints>
+export PROJECT_NAME=<gcloud project name>
+export TPU_NAME=<name of tpu to create>
+export BASE_CKPT=<initial model checkpoint, e.g. 999900>
+```
+
+Copy pre-trained checkpoint to our target model
+```
+echo "model_checkpoint_path: \"model.ckpt-${BASE_CKPT}\"" > checkpoint
+gsutil cp checkpoint ${GS_FOLDER}
+gsutil cp gs://t5-data/pretrained_models/${MODEL}/model.ckpt-${BASE_CKPT}* ${GS_FOLDER}
+```
+
+```
+nohup t5_mesh_transformer  \
+  --tpu="${TPU_NAME}" \
+  --gcp_project="${PROJECT_NAME}" \
+  --tpu_zone="europe-west4-a" \
+  --model_dir="${GS_FOLDER}" \
+  --gin_param="init_checkpoint = 'gs://t5-data/pretrained_models/${MODEL}/model.ckpt-${BASE_CKPT}'" \
+  --gin_file="dataset.gin" \
+  --gin_file="models/bi_v1.gin" \
+  --gin_file="gs://t5-data/pretrained_models/${MODEL}/operative_config.gin" \
+  --gin_param="utils.tpu_mesh_shape.model_parallelism = 1" \
+  --gin_param="utils.tpu_mesh_shape.tpu_topology = '2x2'" \
+  --gin_param="utils.run.train_dataset_fn = @t5.models.mesh_transformer.tsv_dataset_fn" \
+  --gin_param="tsv_dataset_fn.filename = '${GS_FOLDER}/query_doc_pairs.train.tsv'" \
+  --gin_file="learning_rate_schedules/constant_0_001.gin" \
+  --gin_param="run.train_steps = 1100000" \
+  --gin_param="tokens_per_batch = 65536" \
+  >> out.log_exp 2>&1 &
+
+tail -100f out.log_exp
+```
+
+Training T5 base, large, and 3B take approximately 12, 48, and 160 hours overall, respectively, on a single TPU.
+
 ## Replication Log
diff --git a/pygaggle/data/create_msmarco_t5_training_pairs.py b/pygaggle/data/create_msmarco_t5_training_pairs.py
@@ -0,0 +1,21 @@
+"""
+This script creates monoT5 input files for training,
+Each line in the monoT5 input file follows the format:
+    f'Query: {query} Document: {document} Relevant:\t{label}\n')
+"""
+import argparse
+from tqdm import tqdm
+
+parser = argparse.ArgumentParser()
+parser.add_argument("--triples_train", type=str, required=True,
+                    help="tsv file <query>, <positive_document>, <negative_document>")
+parser.add_argument("--output_to_t5", type=str, required=True,
+                    help="t5 train input file")
+args = parser.parse_args()
+
+with open(args.output_to_t5, 'w') as fout_t5:
+    for line_num, line in enumerate(tqdm(open(args.triples_train))):
+        query, positive_document, negative_document = line.strip().split('\t')
+        fout_t5.write(f'Query: {query} Document: {positive_document} Relevant:\ttrue\n')
+        fout_t5.write(f'Query: {query} Document: {negative_document} Relevant:\tfalse\n')
+print('Done!')