def test_apogee_cnn(self):
"""
Test ApogeeCNN models
- training, testing, evaluation
- basic astroNN model method
"""
print("======ApogeeCNN======")
# Data preparation
random_xdata = np.random.normal(0, 1, (200, 1024))
random_ydata = np.random.normal(0, 1, (200, 2))
# setup model instance
neuralnet = ApogeeCNN()
print(neuralnet)
# assert no model before training
self.assertEqual(neuralnet.has_model, False)
neuralnet.max_epochs = 1 # for quick result
neuralnet.callbacks = ErrorOnNaN() # Raise error and fail the test if Nan
neuralnet.train(random_xdata, random_ydata) # training
self.assertEqual(neuralnet.uses_learning_phase, True) # Assert ApogeeCNN uses learning phase (bc of Dropout)
# test basic astroNN model method
neuralnet.get_weights()
neuralnet.summary()
output_shape = neuralnet.output_shape
input_shape = neuralnet.input_shape
neuralnet.get_config()
neuralnet.save_weights('save_weights_test.h5') # save astroNN weight only
neuralnet.plot_dense_stats()
neuralnet.plot_model()
prediction = neuralnet.test(random_xdata)
jacobian = neuralnet.jacobian(random_xdata[:2])
hessian = neuralnet.hessian_diag(random_xdata[:2])
hessian_full_approx = neuralnet.hessian(random_xdata[:2], method='approx')
hessian_full_exact = neuralnet.hessian(random_xdata[:2], method='exact')
# make sure raised if data dimension not as expected
self.assertRaises(ValueError, neuralnet.jacobian, np.atleast_3d(random_xdata[:3]))
# make sure evaluate run in testing phase instead of learning phase
# ie no Dropout which makes model deterministic
self.assertEqual(
np.all(neuralnet.evaluate(random_xdata, random_ydata) == neuralnet.evaluate(random_xdata, random_ydata)),
True)
# assert shape correct as expected
np.testing.assert_array_equal(prediction.shape, random_ydata.shape)
np.testing.assert_array_equal(jacobian.shape, [random_xdata[:2].shape[0], random_ydata.shape[1],
random_xdata.shape[1]])
np.testing.assert_array_equal(hessian.shape, [random_xdata[:2].shape[0], random_ydata.shape[1],
random_xdata.shape[1]])
# hessian approx and exact result should have the same shape
np.testing.assert_array_equal(hessian_full_approx.shape, hessian_full_exact.shape)
# save weight and model again
neuralnet.save(name='apogee_cnn')
neuralnet.save_weights('save_weights_test.h5')
# load the model again
neuralnet_loaded = load_folder("apogee_cnn")
neuralnet_loaded.plot_dense_stats()
# assert has model without training because this is a trained model
self.assertEqual(neuralnet_loaded.has_model, True)
# fine tune test
prediction_loaded = neuralnet_loaded.test(random_xdata)
# ApogeeCNN is deterministic check again
np.testing.assert_array_equal(prediction, prediction_loaded)
# Fine tuning test
neuralnet_loaded.max_epochs = 1
neuralnet_loaded.callbacks = ErrorOnNaN()
neuralnet_loaded.train(random_xdata, random_ydata)
prediction_loaded = neuralnet_loaded.test(random_xdata)
# prediction should not be equal after fine-tuning
self.assertRaises(AssertionError, np.testing.assert_array_equal, prediction, prediction_loaded)
def test_apogee_cnn(self):
# Data preparation, keep the data size large (>800 data points to prevent issues)
random_xdata = np.random.normal(0, 1, (200, 1024))
random_ydata = np.random.normal(0, 1, (200, 2))
# ApogeeCNN
print("======ApogeeCNN======")
neuralnet = ApogeeCNN()
self.assertEqual(neuralnet.has_model, False)
neuralnet.max_epochs = 1
neuralnet.callbacks = ErrorOnNaN()
neuralnet.train(random_xdata, random_ydata)
self.assertEqual(neuralnet.uses_learning_phase, True)
neuralnet.get_weights()
neuralnet.get_config()
neuralnet.save_weights('save_weights_test.h5')
neuralnet.summary()
output_shape = neuralnet.output_shape
input_shape = neuralnet.input_shape
prediction = neuralnet.test(random_xdata)
jacobian = neuralnet.jacobian(random_xdata[:3])
hessian = neuralnet.hessian_diag(random_xdata[:2])
hessian_full_approx = neuralnet.hessian(random_xdata[:2],
method='approx')
hessian_full_exact = neuralnet.hessian(random_xdata[:2],
method='exact')
self.assertRaises(ValueError, neuralnet.jacobian,
np.atleast_3d(random_xdata[:3]))
# make sure evaluate run in testing phase instead of learning phase
# ie no Dropout which makes model deterministic
self.assertEqual(
np.all(
neuralnet.evaluate(random_xdata, random_ydata) ==
neuralnet.evaluate(random_xdata, random_ydata)), True)
np.testing.assert_array_equal(prediction.shape, random_ydata.shape)
np.testing.assert_array_equal(jacobian.shape, [
random_xdata[:3].shape[0], random_ydata.shape[1],
random_xdata.shape[1]
])
np.testing.assert_array_equal(hessian.shape, [
random_xdata[:2].shape[0], random_ydata.shape[1],
random_xdata.shape[1]
])
neuralnet.save(name='apogee_cnn')
neuralnet.save_weights('save_weights_test.h5')
neuralnet_loaded = load_folder("apogee_cnn")
self.assertEqual(neuralnet_loaded.has_model, True)
neuralnet_loaded.max_epochs = 1
neuralnet_loaded.callbacks = ErrorOnNaN()
prediction_loaded = neuralnet_loaded.test(random_xdata)
# Apogee_CNN is deterministic
np.testing.assert_array_equal(prediction, prediction_loaded)
# Fine tuning test
neuralnet_loaded.train(random_xdata, random_ydata)
prediction_loaded = neuralnet_loaded.test(random_xdata)
# prediction should not be equal after fine-tuning
self.assertRaises(AssertionError, np.testing.assert_array_equal,
prediction, prediction_loaded)