def test_save_load_fitted_inerror(self):
import numpy as np
om = self.om
# create classifier and save untrained
estmdl = TFEstimatorModel(estimator_fn=make_estimator_fn())
om.models.put(estmdl, 'estimator-model')
# create dataasets
train_x, train_y, test_x, test_y = make_data()
estmdl.fit(train_x, train_y)
predict = [v for v in estmdl.predict(test_x)]
# this effectively resets the model to untrained state
# i.e. it can be reinstantiated but will not be trained
estmdl._estimator = None
meta = om.models.put(estmdl, 'estimator-model')
# restore and use
estmdl_r = om.models.get('estimator-model')
self.assertIsInstance(estmdl_r, estmdl.__class__)
predict_r = [v for v in estmdl_r.predict(test_x)]
self.assertIsInstance(predict_r, list)
# make sure the model was actually reloaded from a new model_dir
self.assertNotEqual(estmdl.estimator.model_dir,
estmdl_r.estimator.model_dir)
# since we get back an unfitted model, predictions should be way off
self.assertFalse(
np.allclose(predict_r[0]['logits'], predict[0]['logits']))
self.assertFalse(
np.allclose(predict_r[0]['probabilities'],
predict[0]['probabilities']))
def test_fit_predict_from_numpy(self):
import tensorflow as tf
from omegaml.backends.tensorflow.tfestimatormodel import TFEstimatorModel
om = self.om
# note we use a custom input_fn
estmdl = TFEstimatorModel(estimator_fn=make_estimator_fn(),
input_fn=make_input_fn())
train_x, train_y, test_x, test_y = make_data()
# create a feature dict from a numpy array
train_x = train_x.values # numpy
train_y = train_y.values
test_x = test_x.values
classifier = estmdl.fit(train_x, train_y)
self.assertIsInstance(classifier, tf.estimator.LinearClassifier)
# score
score = estmdl.score(test_x, test_y)
self.assertIsInstance(score, dict)
self.assertIn('accuracy', score)
self.assertIn('loss', score)
# predict
predict = [v for v in estmdl.predict(test_x)]
self.assertIsInstance(predict, list)
self.assertIn('logits', predict[0])
self.assertIn('probabilities', predict[0])
self.assertIn('classes', predict[0])
def test_save_load_fitted_estimator(self):
from omegaml.backends.tensorflow.tfestimatormodel import TFEstimatorModel
import numpy as np
om = self.om
# create classifier and save
estmdl = TFEstimatorModel(estimator_fn=make_estimator_fn())
train_x, train_y, test_x, test_y = make_data()
estmdl.fit(train_x, train_y)
predict = [v for v in estmdl.predict(test_x)]
meta = om.models.put(
TFEstimatorModel(estimator_fn=make_estimator_fn(),
model=estmdl.estimator), 'estimator-model')
# restore and use
estmdl_r = om.models.get('estimator-model')
self.assertIsInstance(estmdl_r, estmdl.__class__)
predict_r = [v for v in estmdl_r.predict(test_x)]
self.assertIsInstance(predict_r, list)
# make sure the model was actually reloaded from a new model_dir
self.assertNotEqual(estmdl.estimator.model_dir,
estmdl_r.estimator.model_dir)
self.assertTrue(
np.allclose(predict_r[0]['logits'], predict[0]['logits']))
self.assertTrue(
np.allclose(predict_r[0]['probabilities'],
predict[0]['probabilities']))
def test_predict_from_objecthandler(self):
import tensorflow as tf
from omegaml.backends.tensorflow import _tffn
from omegaml.backends.tensorflow.tfestimatormodel import TFEstimatorModel
om = self.om
@virtualobj
def train_xy_fn(Xname=None, Yname=None, **kwargs):
import omegaml as om
X = om.datasets.get(Xname)
Y = om.datasets.get(Yname)
dataset = _tffn('pandas_input_fn')(X, Y, shuffle=True)
return dataset
@virtualobj
def test_x_fn(Xname=None, **kwargs):
import omegaml as om
X = om.datasets.get(Xname)
dataset = _tffn('pandas_input_fn')(X, shuffle=False)
return dataset
# create classifier
estmdl = TFEstimatorModel(estimator_fn=make_estimator_fn())
train_x, train_y, test_x, test_y = make_data()
# store actual data
om.datasets.put(train_x, 'train_x')
om.datasets.put(train_y, 'train_y')
om.datasets.put(test_x, 'test_x')
# make these virtual objects to return (the input fn that returns a) tf.data.Dataset
om.datasets.put(train_xy_fn, 'train_data')
om.datasets.put(test_x_fn, 'test_dataset')
# use these datasets as an input fn
# -- why can't we pass the Dataset directly? All objects must be executed on the same
# graph wich is only created by estimator.fit() before calling the input_fn.
dataset = om.datasets.get('train_data',
Xname='train_x',
Yname='train_y')
classifier = estmdl.fit(input_fn=dataset)
self.assertIsInstance(classifier, tf.estimator.LinearClassifier)
# predict directly
dataset = om.datasets.get('test_dataset', Xname='test_x')
predict = [v for v in estmdl.predict(input_fn=dataset)]
self.assertIsInstance(predict, list)
self.assertIn('logits', predict[0])
self.assertIn('probabilities', predict[0])
self.assertIn('classes', predict[0])
# do the same thing on the runtime
om.models.put(estmdl, 'estimator-model')
meta = om.runtime.model('estimator-model').fit(
'train_data{Xname=train_x,Yname=train_y}').get()
self.assertTrue(meta.startswith('<Metadata:'))
def test_save_load_unfitted(self):
om = self.om
# create classifier and save
estmdl = TFEstimatorModel(estimator_fn=make_estimator_fn())
meta = om.models.put(estmdl, 'estimator-model')
# restore and use
estmdl_r = om.models.get('estimator-model')
self.assertIsInstance(estmdl_r, estmdl.__class__)
train_x, train_y, test_x, test_y = make_data()
estmdl_r.fit(train_x, train_y)
predict = [v for v in estmdl.predict(test_x)]
self.assertIsInstance(predict, list)
self.assertIn('logits', predict[0])
self.assertIn('probabilities', predict[0])
self.assertIn('classes', predict[0])
def test_fit_predict(self):
import tensorflow as tf
om = self.om
# create classifier
estmdl = TFEstimatorModel(estimator_fn=make_estimator_fn())
train_x, train_y, test_x, test_y = make_data()
classifier = estmdl.fit(train_x, train_y)
self.assertIsInstance(classifier, tf.estimator.LinearClassifier)
# score
score = estmdl.score(test_x, test_y)
self.assertIsInstance(score, dict)
self.assertIn('accuracy', score)
self.assertIn('loss', score)
# predict
predict = [v for v in estmdl.predict(test_x)]
self.assertIsInstance(predict, list)
self.assertIn('logits', predict[0])
self.assertIn('probabilities', predict[0])
self.assertIn('classes', predict[0])
