(predicted_labels, log_probs) = create_model(
bert_config, is_training, is_predicting, input_ids, input_mask, segment_ids, label_ids, num_labels, use_one_hot_embeddings)
predictions = {
'probabilities': log_probs,
'labels': predicted_labels
}
return tf.estimator.EstimatorSpec(mode, predictions=predictions)
# Return the actual model function in the closure
return model_fn
if __name__ == "__main__":
label_list = [int(i) for i in hp.label_list.split(",")]
train_features, test_features = process_data(hp)
""" Start train """
# Compute # train and warmup steps from batch size
num_train_steps = int(len(train_features) / hp.BATCH_SIZE * hp.NUM_TRAIN_EPOCHS)
num_warmup_steps = int(num_train_steps * hp.WARMUP_PROPORTION)
# Specify output directory and number of checkpoint steps to save
run_config = tf.estimator.RunConfig(model_dir=hp.OUTPUT_DIR, save_summary_steps=hp.SAVE_SUMMARY_STEPS, save_checkpoints_steps=hp.SAVE_CHECKPOINTS_STEPS)
bert_config = modeling.BertConfig.from_json_file(hp.BERT_CONFIG)
model_fn = model_fn_builder(bert_config=bert_config, num_labels=len(label_list), learning_rate=hp.LEARNING_RATE, num_train_steps=num_train_steps,
num_warmup_steps=num_warmup_steps, use_one_hot_embeddings=False)
estimator = tf.estimator.Estimator(model_fn=model_fn, config=run_config, params={"batch_size": hp.BATCH_SIZE})
# Create an input function for training. drop_remainder = True for using TPUs.
global_step = tf.train.get_or_create_global_step()
tf.summary.scalar("loss", loss)
tf.summary.scalar("global_step", global_step)
accuracy = tf.metrics.accuracy(label_ids, predicted_labels)
eval_metrics = {"eval_accuracy": accuracy}
tf.summary.scalar("eval_accuracy", accuracy[1])
summaries = tf.summary.merge_all()
return loss, train_op, global_step, summaries, eval_metrics, label_ids, predicted_labels
if __name__ == "__main__":
label_list = [int(i) for i in hp.label_list.split(",")]
train_features, eval_features = process_data(hp)
# Compute # train and warmup steps from batch size
num_train_steps = int(
len(train_features) / hp.BATCH_SIZE * hp.NUM_TRAIN_EPOCHS)
num_warmup_steps = int(num_train_steps * hp.WARMUP_PROPORTION)
num_eval_batches = len(eval_features) // hp.BATCH_SIZE + int(
len(eval_features) % hp.BATCH_SIZE != 0)
# Create an input function for training. drop_remainder = True for using TPUs.
train_input_fn = run_classifier.input_fn_builder(
features=train_features,
seq_length=hp.MAX_SEQ_LENGTH,
is_training=True,
drop_remainder=False)