from __gin__ import dynamic_registration
import tasks
import seqio

import __main__ as train_script
from t5.data import mixtures
from t5x import models
from t5x import partitioning
from t5x import utils

include "t5x/examples/t5/t5_1_1/large.gin"
include "t5x/configs/runs/finetune.gin"

MIXTURE_OR_TASK_NAME = %gin.REQUIRED
TASK_FEATURE_LENGTHS = {"inputs": 512, "targets": 64}
INITIAL_CHECKPOINT_PATH = %gin.REQUIRED
TRAIN_STEPS = %gin.REQUIRED  # 1000000 pre-trained steps + 10000 fine-tuning steps.
USE_CACHED_TASKS = False
DROPOUT_RATE = 0.1
RANDOM_SEED = 0
#BATCH_SIZE = 32

#Fixing a small error
infer_eval/utils.DatasetConfig:
  task_feature_lengths = %TASK_FEATURE_LENGTHS

#Saving every 1000 steps
utils.SaveCheckpointConfig:
  period = 1000


# Pere: Only necessary if we load a t5 model. We can start with an t5x model here
# `LOSS_NORMALIZING_FACTOR`: When fine-tuning a model that was pre-trained
# using Mesh Tensorflow (e.g. the public T5 / mT5 / ByT5 models), this should be
# set to `pretraining batch_size` * `target_token_length`. For T5 and T5.1.1:
# `2048 * 114`. For mT5: `1024 * 229`. For ByT5: `1024 * 189`.
# LOSS_NORMALIZING_FACTOR = 234496

# Might have to ba changed based on architecture
# partitioning.PjitPartitioner.num_partitions = 1

VOCABULARY = @seqio.SentencePieceVocabulary()
seqio.SentencePieceVocabulary.sentencepiece_model_file = "gs://nb-t5/t5/vocabs/wikipedia/no-da-en-sv-nn-is_32000_unigram.sp.model"
seqio.SentencePieceVocabulary.extra_ids = 100