from __gin__ import dynamic_registration
import tasks

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/byt5/large.gin"
include "t5x/configs/runs/finetune.gin"

MIXTURE_OR_TASK_NAME = "sentencefix"
TASK_FEATURE_LENGTHS = {"inputs": 256, "targets": 256}
TRAIN_STEPS = 1_200_000  # 1000000 pre-trained steps + 100000 fine-tuning steps.
USE_CACHED_TASKS = False
DROPOUT_RATE = 0.0
RANDOM_SEED = 0
BATCH_SIZE = 64

# `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 =193536
INITIAL_CHECKPOINT_PATH = "gs://t5-data/pretrained_models/byt5/large/model.ckpt-1000000"

train_script.train:
  eval_period = 500
  partitioner = @partitioning.ModelBasedPjitPartitioner()

# `num_decodes` is equivalent to a beam size in a beam search decoding.
models.EncoderDecoderModel.predict_batch_with_aux.num_decodes = 4

partitioning.ModelBasedPjitPartitioner.num_partitions = 2


#from t5.models import mesh_transformer
#import t5.models
#mesh_transformer.learning_rate_schedules.constant_learning_rate.learning_rate = 0.0005
#run.learning_rate_schedule = @learning_rate_schedules.constant_learning_rate