updated training tasks for big corpus

finetune_large_mt5_sentencefix_v4_16.gin

@@ -12,7 +12,7 @@ include "t5x/configs/runs/finetune.gin"
 
 MIXTURE_OR_TASK_NAME = "sentencefix"
 TASK_FEATURE_LENGTHS = {"inputs": 256, "targets": 256}
-TRAIN_STEPS = 1_300_000  # 1000000 pre-trained steps + 20000 fine-tuning steps.
+TRAIN_STEPS = 1_200_000  # 1000000 pre-trained steps + 20000 fine-tuning steps.
 USE_CACHED_TASKS = False
 DROPOUT_RATE = 0.0
 RANDOM_SEED = 0

tasks.py

@@ -9,9 +9,9 @@ import t5
 import tensorflow.compat.v1 as tf
 
 tsv_path = {
-        "train": "gs://nb-t5x/corpus/train/train.tsv",
-        "validation": "gs://nb-t5x/corpus/eval/eval.tsv",
-        "test": "gs://nb-t5x/corpus/test/test.tsv"
+        "train": "gs://nb-t5x-us-central2/corpus_big/train.tsv",
+        "validation": "gs://nb-t5x-us-central2/corpus_big/eval.tsv",
+        "test": "gs://nb-t5x-us-central2/corpus_big/test.tsv"
 }
 
 vocabulary = seqio.SentencePieceVocabulary(

train_large_v4_16.sh

@@ -1,6 +1,6 @@
 PROJECT_DIR=${HOME}"/models/multi-sentencefix-mt5"
 T5X_DIR="../../t5x"  # directory where the t5x is cloned.
-MODEL_DIR="gs://nb-t5x-us-central2/model_mT5X_large_16_d"
+MODEL_DIR="gs://nb-t5x-us-central2/model_mT5X_large_16_e"
 export PYTHONPATH=${PROJECT_DIR}
 
 python3 ${T5X_DIR}/t5x/train.py \