gin_config.gin

from __gin__ import dynamic_registration
import t5x
import seqio
import t5.data.mixtures
from t5x import utils
from t5x import models
import t5x.infer
from t5x import partitioning
import rephrase_inference

include 't5x/configs/runs/infer.gin'
include 't5x/examples/t5/t5_1_1/base.gin'

MIXTURE_OR_TASK_NAME = "hf_inference_task"
TASK_FEATURE_LENGTHS = {'inputs': 1024, 'targets': 200}
CHECKPOINT_PATH = "./checkpoint_1020000"
USE_CACHED_TASKS = False
DROPOUT_RATE = 0.0

# TPU Configuration
partitioning.PjitPartitioner:
  num_partitions = 4

models.EncoderDecoderModel.predict_batch_with_aux.num_decodes = 1

infer_eval/utils.DatasetConfig:
  mixture_or_task_name = %MIXTURE_OR_TASK_NAME
  task_feature_lengths = %TASK_FEATURE_LENGTHS
  split = %FILE_PATH
  batch_size = 32
  shuffle = False
  seed = 42
  pack = False
  use_cached = False
  module = None

utils.DatasetConfig:
  split = %FILE_PATH

seqio.Evaluator:
  logger_cls = [@seqio.PyLoggingLogger, @seqio.TensorBoardLogger, @seqio.JSONLogger]
  num_examples = None
  use_memory_cache = True

FILE_PATH = %gin.REQUIRED