def testInvalidModelOrigin(self): # Test that both uri and model is given task_args = self.BATCH_PREDICTION_DEFAULT_ARGS.copy() task_args['uri'] = 'gs://fake-uri/saved_model' task_args['model_name'] = 'fake_model' with self.assertRaises(ValueError) as context: CloudMLBatchPredictionOperator(**task_args).execute(None) self.assertEquals('Ambiguous model origin.', str(context.exception)) # Test that both uri and model/version is given task_args = self.BATCH_PREDICTION_DEFAULT_ARGS.copy() task_args['uri'] = 'gs://fake-uri/saved_model' task_args['model_name'] = 'fake_model' task_args['version_name'] = 'fake_version' with self.assertRaises(ValueError) as context: CloudMLBatchPredictionOperator(**task_args).execute(None) self.assertEquals('Ambiguous model origin.', str(context.exception)) # Test that a version is given without a model task_args = self.BATCH_PREDICTION_DEFAULT_ARGS.copy() task_args['version_name'] = 'bare_version' with self.assertRaises(ValueError) as context: CloudMLBatchPredictionOperator(**task_args).execute(None) self.assertEquals('Missing model origin.', str(context.exception)) # Test that none of uri, model, model/version is given task_args = self.BATCH_PREDICTION_DEFAULT_ARGS.copy() with self.assertRaises(ValueError) as context: CloudMLBatchPredictionOperator(**task_args).execute(None) self.assertEquals('Missing model origin.', str(context.exception))
def testSuccessWithModel(self): with patch('airflow.contrib.operators.cloudml_operator.CloudMLHook') \ as mock_hook: input_with_model = self.INPUT_MISSING_ORIGIN.copy() input_with_model['modelName'] = \ 'projects/test-project/models/test_model' success_message = self.SUCCESS_MESSAGE_MISSING_INPUT.copy() success_message['predictionInput'] = input_with_model hook_instance = mock_hook.return_value hook_instance.get_job.side_effect = errors.HttpError( resp=httplib2.Response({'status': 404}), content=b'some bytes') hook_instance.create_job.return_value = success_message prediction_task = CloudMLBatchPredictionOperator( job_id='test_prediction', project_id='test-project', region=input_with_model['region'], data_format=input_with_model['dataFormat'], input_paths=input_with_model['inputPaths'], output_path=input_with_model['outputPath'], model_name=input_with_model['modelName'].split('/')[-1], dag=self.dag, task_id='test-prediction') prediction_output = prediction_task.execute(None) mock_hook.assert_called_with('google_cloud_default', None) hook_instance.create_job.assert_called_once_with( 'test-project', { 'jobId': 'test_prediction', 'predictionInput': input_with_model }, ANY) self.assertEquals(success_message['predictionOutput'], prediction_output)
def testHttpError(self): http_error_code = 403 with patch('airflow.contrib.operators.cloudml_operator.CloudMLHook') \ as mock_hook: input_with_model = self.INPUT_MISSING_ORIGIN.copy() input_with_model['modelName'] = \ 'projects/experimental/models/test_model' hook_instance = mock_hook.return_value hook_instance.create_job.side_effect = errors.HttpError( resp=httplib2.Response({'status': http_error_code}), content=b'Forbidden') with self.assertRaises(errors.HttpError) as context: prediction_task = CloudMLBatchPredictionOperator( job_id='test_prediction', project_id='test-project', region=input_with_model['region'], data_format=input_with_model['dataFormat'], input_paths=input_with_model['inputPaths'], output_path=input_with_model['outputPath'], model_name=input_with_model['modelName'].split('/')[-1], dag=self.dag, task_id='test-prediction') prediction_task.execute(None) mock_hook.assert_called_with('google_cloud_default', None) hook_instance.create_job.assert_called_with( 'test-project', { 'jobId': 'test_prediction', 'predictionInput': input_with_model }, ANY) self.assertEquals(http_error_code, context.exception.resp.status)
def testFailedJobError(self): with patch('airflow.contrib.operators.cloudml_operator.CloudMLHook') \ as mock_hook: hook_instance = mock_hook.return_value hook_instance.create_job.return_value = { 'state': 'FAILED', 'errorMessage': 'A failure message' } task_args = self.BATCH_PREDICTION_DEFAULT_ARGS.copy() task_args['uri'] = 'a uri' with self.assertRaises(RuntimeError) as context: CloudMLBatchPredictionOperator(**task_args).execute(None) self.assertEquals('A failure message', str(context.exception))
def create_evaluate_ops(task_prefix, project_id, job_id, region, data_format, input_paths, prediction_path, metric_fn_and_keys, validate_fn, dataflow_options, model_uri=None, model_name=None, version_name=None, dag=None): """ Creates Operators needed for model evaluation and returns. It gets prediction over inputs via Cloud ML Engine BatchPrediction API by calling CloudMLBatchPredictionOperator, then summarize and validate the result via Cloud Dataflow using DataFlowPythonOperator. For details and pricing about Batch prediction, please refer to the website https://cloud.google.com/ml-engine/docs/how-tos/batch-predict and for Cloud Dataflow, https://cloud.google.com/dataflow/docs/ It returns three chained operators for prediction, summary, and validation, named as <prefix>-prediction, <prefix>-summary, and <prefix>-validation, respectively. (<prefix> should contain only alphanumeric characters or hyphen.) The upstream and downstream can be set accordingly like: pred, _, val = create_evaluate_ops(...) pred.set_upstream(upstream_op) ... downstream_op.set_upstream(val) Callers will provide two python callables, metric_fn and validate_fn, in order to customize the evaluation behavior as they wish. - metric_fn receives a dictionary per instance derived from json in the batch prediction result. The keys might vary depending on the model. It should return a tuple of metrics. - validation_fn receives a dictionary of the averaged metrics that metric_fn generated over all instances. The key/value of the dictionary matches to what's given by metric_fn_and_keys arg. The dictionary contains an additional metric, 'count' to represent the total number of instances received for evaluation. The function would raise an exception to mark the task as failed, in a case the validation result is not okay to proceed (i.e. to set the trained version as default). Typical examples are like this: def get_metric_fn_and_keys(): import math # imports should be outside of the metric_fn below. def error_and_squared_error(inst): label = float(inst['input_label']) classes = float(inst['classes']) # 0 or 1 err = abs(classes-label) squared_err = math.pow(classes-label, 2) return (err, squared_err) # returns a tuple. return error_and_squared_error, ['err', 'mse'] # key order must match. def validate_err_and_count(summary): if summary['err'] > 0.2: raise ValueError('Too high err>0.2; summary=%s' % summary) if summary['mse'] > 0.05: raise ValueError('Too high mse>0.05; summary=%s' % summary) if summary['count'] < 1000: raise ValueError('Too few instances<1000; summary=%s' % summary) return summary For the details on the other BatchPrediction-related arguments (project_id, job_id, region, data_format, input_paths, prediction_path, model_uri), please refer to CloudMLBatchPredictionOperator too. :param task_prefix: a prefix for the tasks. Only alphanumeric characters and hyphen are allowed (no underscores), since this will be used as dataflow job name, which doesn't allow other characters. :type task_prefix: string :param model_uri: GCS path of the model exported by Tensorflow using tensorflow.estimator.export_savedmodel(). It cannot be used with model_name or version_name below. See CloudMLBatchPredictionOperator for more detail. :type model_uri: string :param model_name: Used to indicate a model to use for prediction. Can be used in combination with version_name, but cannot be used together with model_uri. See CloudMLBatchPredictionOperator for more detail. :type model_name: string :param version_name: Used to indicate a model version to use for prediciton, in combination with model_name. Cannot be used together with model_uri. See CloudMLBatchPredictionOperator for more detail. :type version_name: string :param data_format: either of 'TEXT', 'TF_RECORD', 'TF_RECORD_GZIP' :type data_format: string :param input_paths: a list of input paths to be sent to BatchPrediction. :type input_paths: list of strings :param prediction_path: GCS path to put the prediction results in. :type prediction_path: string :param metric_fn_and_keys: a tuple of metric_fn and metric_keys: - metric_fn is a function that accepts a dictionary (for an instance), and returns a tuple of metric(s) that it calculates. - metric_keys is a list of strings to denote the key of each metric. :type metric_fn_and_keys: tuple of a function and a list of strings :param validate_fn: a function to validate whether the averaged metric(s) is good enough to push the model. :type validate_fn: function :param dataflow_options: options to run Dataflow jobs. :type dataflow_options: dictionary :returns: a tuple of three operators, (prediction, summary, validation) :rtype: tuple(DataFlowPythonOperator, DataFlowPythonOperator, PythonOperator) """ # Verify that task_prefix doesn't have any special characters except hyphen # '-', which is the only allowed non-alphanumeric character by Dataflow. if not re.match(r"^[a-zA-Z][-A-Za-z0-9]*$", task_prefix): raise AirflowException( "Malformed task_id for DataFlowPythonOperator (only alphanumeric " "and hyphens are allowed but got: " + task_prefix) metric_fn, metric_keys = metric_fn_and_keys if not callable(metric_fn): raise AirflowException("`metric_fn` param must be callable.") if not callable(validate_fn): raise AirflowException("`validate_fn` param must be callable.") evaluate_prediction = CloudMLBatchPredictionOperator( task_id=(task_prefix + "-prediction"), project_id=project_id, job_id=_normalize_cloudml_job_id(job_id), region=region, data_format=data_format, input_paths=input_paths, output_path=prediction_path, uri=model_uri, model_name=model_name, version_name=version_name, dag=dag) metric_fn_encoded = base64.b64encode(dill.dumps(metric_fn, recurse=True)) evaluate_summary = DataFlowPythonOperator( task_id=(task_prefix + "-summary"), py_options=["-m"], py_file="airflow.contrib.operators.cloudml_prediction_summary", dataflow_default_options=dataflow_options, options={ "prediction_path": prediction_path, "metric_fn_encoded": metric_fn_encoded, "metric_keys": ','.join(metric_keys) }, dag=dag) # TODO: "options" is not template_field of DataFlowPythonOperator (not sure # if intended or by mistake); consider fixing in the DataFlowPythonOperator. evaluate_summary.template_fields.append("options") evaluate_summary.set_upstream(evaluate_prediction) def apply_validate_fn(*args, **kwargs): prediction_path = kwargs["templates_dict"]["prediction_path"] scheme, bucket, obj, _, _ = urlsplit(prediction_path) if scheme != "gs" or not bucket or not obj: raise ValueError("Wrong format prediction_path: %s", prediction_path) summary = os.path.join(obj.strip("/"), "prediction.summary.json") gcs_hook = GoogleCloudStorageHook() summary = json.loads(gcs_hook.download(bucket, summary)) return validate_fn(summary) evaluate_validation = PythonOperator( task_id=(task_prefix + "-validation"), python_callable=apply_validate_fn, provide_context=True, templates_dict={"prediction_path": prediction_path}, dag=dag) evaluate_validation.set_upstream(evaluate_summary) return evaluate_prediction, evaluate_summary, evaluate_validation