def get_detext_inference_job(self):
""" Get detext inference job. For LR model the inference job is included in train
job, this job is for DeText model inference.
Return: an inference job inferencing training and validation data
(job_type, job_name, "", job_params)
"""
params = replace(
self.gdmix_params,
action=ACTION_INFERENCE), DetextArg(**self.fixed_effect_config,
out_dir=self.output_model_dir)
return GDMIX_TFJOB, f"{self.fixed_effect_name}-tf-inference", "", params
def get_train_job(self):
""" Get tfjob training job.
:return (job_type, job_name, "", job_params) where job_params are typed param containers supported by smart-arg
"""
if self.model_type == LOGISTIC_REGRESSION:
params = FixedLRParams(**self.fixed_effect_config,
output_model_dir=self.output_model_dir)
elif self.model_type == DETEXT:
params = DetextArg(**self.fixed_effect_config,
out_dir=self.output_model_dir)
else:
raise ValueError(f'unsupported model_type: {self.model_type}')
return GDMIX_TFJOB, f"{self.fixed_effect_name}-tf-train", "", (
self.gdmix_params, params)
def test_detext_model_fixed_effect_workflow_generator(self):
# return # skip test for now
fe_workflow = FixedEffectWorkflowGenerator(self.detext_config_obj)
# check sequence
seq = fe_workflow.get_job_sequence()
self.assertEqual(len(seq), 3)
# check job properties
actual_train_job = seq[0]
actual_inference_job = seq[1]
actual_compute_metric_job = seq[2]
expected_train_job_param = (
Params(uid_column_name='uid', weight_column_name='weight', label_column_name='response', prediction_score_column_name='predictionScore',
prediction_score_per_coordinate_column_name='predictionScorePerCoordinate', action='train', stage='fixed_effect', model_type='detext',
training_score_dir='detext-training/global/training_scores', validation_score_dir='detext-training/global/validation_scores'),
DetextArg(feature_names=['label', 'doc_query', 'uid', 'wide_ftrs_sp_idx', 'wide_ftrs_sp_val'], ftr_ext='cnn', num_units=64,
sp_emb_size=1, num_hidden=[0], num_wide=0, num_wide_sp=45, use_deep=True, elem_rescale=True, use_doc_projection=False,
use_usr_projection=False, ltr_loss_fn='pointwise', emb_sim_func=['inner'], num_classes=1, filter_window_sizes=[3], num_filters=50,
explicit_empty=False, use_bert_dropout=False, unit_type='lstm', num_layers=1,
num_residual_layers=0, forget_bias=1.0, rnn_dropout=0.0, bidirectional=False, normalized_lm=False, optimizer='bert_adam',
max_gradient_norm=1.0, learning_rate=0.002, num_train_steps=1000, num_warmup_steps=0, train_batch_size=64,
test_batch_size=64, l1=None, l2=None, train_file='movieLens/detext/trainingData/train_data.tfrecord',
dev_file='movieLens/detext/validationData/test_data.tfrecord', test_file='movieLens/detext/validationData/test_data.tfrecord',
out_dir='detext-training/global/models', max_len=16, min_len=3, vocab_file='movieLens/detext/vocab.txt', we_file=None,
we_trainable=True, PAD='[PAD]', SEP='[SEP]', CLS='[CLS]', UNK='[UNK]', MASK='[MASK]', we_file_for_id_ftr=None,
we_trainable_for_id_ftr=True, PAD_FOR_ID_FTR='[PAD]', UNK_FOR_ID_FTR='[UNK]', random_seed=1234, steps_per_stats=10, num_eval_rounds=None,
steps_per_eval=100, keep_checkpoint_max=1, init_weight=0.1, pmetric='auc', all_metrics=['auc'], score_rescale=None,
add_first_dim_for_query_placeholder=False, add_first_dim_for_usr_placeholder=False, tokenization='punct', resume_training=False,
use_tfr_loss=False, tfr_loss_fn='softmax_loss'))
expected_train_job = ('gdmix_tfjob', 'global-tf-train', '', expected_train_job_param)
self.assertEqual(expected_train_job, actual_train_job)
expected_inference_job_param = (replace(expected_train_job_param[0],
action=ACTION_INFERENCE,
training_score_dir="detext-training/global/training_scores",
validation_score_dir="detext-training/global/validation_scores"), expected_train_job_param[1])
expected_inference_job = 'gdmix_tfjob', 'global-tf-inference', '', expected_inference_job_param
self.assertEqual(expected_inference_job, actual_inference_job)
expected_compute_metric_job = (
'gdmix_sparkjob',
'global-compute-metric',
'com.linkedin.gdmix.evaluation.AreaUnderROCCurveEvaluator',
{'\\--metricsInputDir': 'detext-training/global/validation_scores',
'--outputMetricFile': 'detext-training/global/metric',
'--labelColumnName': 'response',
'--predictionColumnName': 'predictionScore'})
self.assertEqual(actual_compute_metric_job, expected_compute_metric_job)
def test_run_detext_bert(self):
"""
This method test run_detext with BERT models
"""
output = os.path.join(out_dir, "bert")
argument = DetextArg(ftr_ext='bert',
num_units=4,
num_wide=3,
ltr_loss_fn='softmax',
emb_sim_func=['inner', 'concat', 'diff'],
num_filters=50,
bert_config_file=bert_config,
use_bert_dropout=True,
optimizer='bert_adam',
learning_rate=0.002,
num_train_steps=4,
train_batch_size=2,
test_batch_size=2,
train_file=data,
dev_file=data,
test_file=data,
out_dir=output,
max_len=16,
vocab_file=vocab,
vocab_file_for_id_ftr=vocab,
steps_per_stats=1,
steps_per_eval=2,
keep_checkpoint_max=5,
feature_names=[
'label', 'query', 'doc_completedQuery',
'usr_headline', 'usr_skills',
'usr_currTitles', 'usrId_currTitles',
'docId_completedQuery', 'wide_ftrs', 'weight'
],
pmetric='ndcg@10',
all_metrics=['precision@1', 'ndcg@10'])
run_detext(argument)
self._cleanUp(output)
def _parse_parameters(self, raw_model_parameters):
return DetextArg.__from_argv__(raw_model_parameters,
error_on_unknown=False)