Python model_utils.init_from_checkpoint方法代码示例

# 需要导入模块: import model_utils [as 别名]
# 或者: from model_utils import init_from_checkpoint [as 别名]
def get_model_fn(self,
                     model_config,
                     run_config,
                     init_checkpoint,
                     model_type):
        """Returns `model_fn` closure for TPUEstimator."""
        def model_fn(features,
                     labels,
                     mode,
                     params):  # pylint: disable=unused-argument
            """The `model_fn` for TPUEstimator."""
            tf.logging.info("*** Features ***")
            for name in sorted(features.keys()):
                tf.logging.info("  name = %s, shape = %s" % (name, features[name].shape))

            input_ids = features["input_ids"]
            input_mask = features["input_mask"]
            segment_ids = features["segment_ids"]

            embeddings = self._create_model(model_config, run_config, input_ids, input_mask, segment_ids, model_type)
            scaffold_fn = model_utils.init_from_checkpoint(FLAGS)
            
            output_spec = None
            if mode == tf.estimator.ModeKeys.TRAIN:
                loss = tf.Variable(0.0, name="loss", dtype=tf.float32)
                train_op, _, _ = model_utils.get_train_op(FLAGS, loss)
                output_spec = tf.contrib.tpu.TPUEstimatorSpec(
                    mode=mode,
                    loss=loss,
                    train_op=train_op,
                    scaffold_fn=scaffold_fn)
            else:
                output_spec = tf.contrib.tpu.TPUEstimatorSpec(
                    mode=mode,
                    predictions={ "embeddings": embeddings },
                    scaffold_fn=scaffold_fn)
            
            return output_spec
        
        return model_fn 

# 需要导入模块: import model_utils [as 别名]
# 或者: from model_utils import init_from_checkpoint [as 别名]
def get_model_fn():
  """doc."""
  def model_fn(features, labels, mode, params):
    """doc."""
    #### Training or Evaluation
    is_training = (mode == tf.estimator.ModeKeys.TRAIN)
    assert is_training

    #### Retrieve `mems` from `params["cache"]`
    mems = {}
    idx = 0
    if FLAGS.mem_len > 0:
      mems["mems"] = params["cache"]

    #### Get loss from inputs
    total_loss, new_mems, monitor_dict = function_builder.get_loss(
        FLAGS, features, labels, mems, is_training)

    #### Turn `new_mems` into `new_cache`
    new_cache = []
    if FLAGS.mem_len > 0:
      new_cache += new_mems["mems"]

    #### Check model parameters
    num_params = sum([np.prod(v.shape) for v in tf.trainable_variables()])
    tf.logging.info("#params: {}".format(num_params))

    #### Configuring the optimizer
    train_op, learning_rate, gnorm = model_utils.get_train_op(
        FLAGS, total_loss)
    monitor_dict["lr"] = learning_rate
    monitor_dict["gnorm"] = gnorm

    #### Customized initial checkpoint
    scaffold_fn = model_utils.init_from_checkpoint(FLAGS, global_vars=True)

    #### Creating host calls
    host_call = function_builder.construct_scalar_host_call(
        monitor_dict=monitor_dict,
        model_dir=FLAGS.model_dir,
        prefix="train/",
        reduce_fn=tf.reduce_mean)

    #### Constucting training TPUEstimatorSpec with new cache.
    train_spec = tf.contrib.tpu.TPUEstimatorSpec(
        mode=mode, loss=total_loss, train_op=train_op, host_call=host_call,
        scaffold_fn=scaffold_fn)

    train_spec.cache = new_cache

    return train_spec

  return model_fn 

# 需要导入模块: import model_utils [as 别名]
# 或者: from model_utils import init_from_checkpoint [as 别名]
def get_model_fn(self):
        """Returns `model_fn` closure for TPUEstimator."""
        def model_fn(features,
                     labels,
                     mode,
                     params):  # pylint: disable=unused-argument
            """The `model_fn` for TPUEstimator."""
            tf.logging.info("*** Features ***")
            for name in sorted(features.keys()):
                tf.logging.info("  name = %s, shape = %s" % (name, features[name].shape))
            
            is_training = (mode == tf.estimator.ModeKeys.TRAIN)
            
            unique_id = features["unique_id"]
            input_ids = features["input_ids"]
            input_mask = features["input_mask"]
            p_mask = features["p_mask"]
            segment_ids = features["segment_ids"]
            cls_index = features["cls_index"]
            
            if is_training:
                start_position = features["start_position"]
                end_position = features["end_position"]
                is_impossible = features["is_impossible"]
            else:
                start_position = None
                end_position = None
                is_impossible = None

            loss, predicts = self._create_model(is_training, input_ids, input_mask,
                p_mask, segment_ids, cls_index, start_position, end_position, is_impossible)
            
            scaffold_fn = model_utils.init_from_checkpoint(FLAGS)
            
            output_spec = None
            if is_training:
                train_op, _, _ = model_utils.get_train_op(FLAGS, loss)
                output_spec = tf.contrib.tpu.TPUEstimatorSpec(
                    mode=mode,
                    loss=loss,
                    train_op=train_op,
                    scaffold_fn=scaffold_fn)
            else:
                output_spec = tf.contrib.tpu.TPUEstimatorSpec(
                    mode=mode,
                    predictions={
                        "unique_id": unique_id,
                        "answer_prob": predicts["answer_prob"],
                        "start_prob": predicts["start_prob"],
                        "start_index": predicts["start_index"],
                        "end_prob": predicts["end_prob"],
                        "end_index": predicts["end_index"]
                    },
                    scaffold_fn=scaffold_fn)
            
            return output_spec
        
        return model_fn 

# 需要导入模块: import model_utils [as 别名]
# 或者: from model_utils import init_from_checkpoint [as 别名]
def get_model_fn(self,
                     label_list):
        """Returns `model_fn` closure for TPUEstimator."""
        def model_fn(features,
                     labels,
                     mode,
                     params):  # pylint: disable=unused-argument
            """The `model_fn` for TPUEstimator."""
            def metric_fn(label_ids,
                          predict_ids):
                precision = tf.metrics.precision(labels=label_ids, predictions=predict_ids)
                recall = tf.metrics.recall(labels=label_ids, predictions=predict_ids)

                metric = {
                    "precision": precision,
                    "recall": recall,
                }

                return metric
            
            tf.logging.info("*** Features ***")
            for name in sorted(features.keys()):
                tf.logging.info("  name = %s, shape = %s" % (name, features[name].shape))

            input_ids = features["input_ids"]
            input_masks = features["input_masks"]
            segment_ids = features["segment_ids"]
            label_ids = features["label_ids"] if mode in [tf.estimator.ModeKeys.TRAIN, tf.estimator.ModeKeys.EVAL] else None

            loss, predict_ids = self._create_model(input_ids, input_masks, segment_ids, label_ids, label_list, mode)
            
            scaffold_fn = model_utils.init_from_checkpoint(FLAGS)
            
            output_spec = None
            if mode == tf.estimator.ModeKeys.TRAIN:
                train_op, _, _ = model_utils.get_train_op(FLAGS, loss)
                output_spec = tf.contrib.tpu.TPUEstimatorSpec(
                    mode=mode,
                    loss=loss,
                    train_op=train_op,
                    scaffold_fn=scaffold_fn)
            elif mode == tf.estimator.ModeKeys.EVAL:
                masked_label_ids = self._get_masked_data(label_ids, label_list)
                masked_predict_ids = self._get_masked_data(predict_ids, label_list)
                eval_metrics = (metric_fn, [masked_label_ids, masked_predict_ids])
                output_spec = tf.contrib.tpu.TPUEstimatorSpec(
                    mode=mode,
                    loss=loss,
                    eval_metrics=eval_metrics,
                    scaffold_fn=scaffold_fn)
            else:
                output_spec = tf.contrib.tpu.TPUEstimatorSpec(
                    mode=mode,
                    predictions={ "predict": predict_ids },
                    scaffold_fn=scaffold_fn)
            
            return output_spec
        
        return model_fn 

# 需要导入模块: import model_utils [as 别名]
# 或者: from model_utils import init_from_checkpoint [as 别名]
def get_model_fn(self,
                     sent_label_list):
        """Returns `model_fn` closure for TPUEstimator."""
        def model_fn(features,
                     labels,
                     mode,
                     params):  # pylint: disable=unused-argument
            """The `model_fn` for TPUEstimator."""
            def metric_fn(sent_label_ids,
                          sent_predict_ids):
                sent_accuracy = tf.metrics.accuracy(labels=sent_label_ids, predictions=sent_predict_ids)

                metric = {
                    "sent_accuracy": sent_accuracy,
                }

                return metric
            
            tf.logging.info("*** Features ***")
            for name in sorted(features.keys()):
                tf.logging.info("  name = %s, shape = %s" % (name, features[name].shape))

            input_ids = features["input_ids"]
            input_masks = features["input_masks"]
            segment_ids = features["segment_ids"]
            sent_label_ids = features["sent_label_ids"] if mode in [tf.estimator.ModeKeys.TRAIN, tf.estimator.ModeKeys.EVAL] else None
            
            loss, sent_predict_ids, sent_predict_scores, sent_predict_probs = self._create_model(input_ids, input_masks, segment_ids, sent_label_ids, sent_label_list, mode)
            
            scaffold_fn = model_utils.init_from_checkpoint(FLAGS)
            
            output_spec = None
            if mode == tf.estimator.ModeKeys.TRAIN:
                train_op, _, _ = model_utils.get_train_op(FLAGS, loss)
                output_spec = tf.contrib.tpu.TPUEstimatorSpec(
                    mode=mode,
                    loss=loss,
                    train_op=train_op,
                    scaffold_fn=scaffold_fn)
            elif mode == tf.estimator.ModeKeys.EVAL:
                eval_metrics = (metric_fn, [sent_label_ids, sent_predict_ids])
                output_spec = tf.contrib.tpu.TPUEstimatorSpec(
                    mode=mode,
                    loss=loss,
                    eval_metrics=eval_metrics,
                    scaffold_fn=scaffold_fn)
            else:
                output_spec = tf.contrib.tpu.TPUEstimatorSpec(
                    mode=mode,
                    predictions={
                        "sent_predict_id": sent_predict_ids,
                        "sent_predict_score": sent_predict_scores,
                        "sent_predict_prob": sent_predict_probs
                    },
                    scaffold_fn=scaffold_fn)
            
            return output_spec
        
        return model_fn