def test_fit_and_predict_with_new_labels_set_via_method(self):
# Initialize model
invalid_entities = {"PAD": 0, "BACKGROUND": 1, "test3": 2}
cnn_model = CharacterLevelCnnModel(invalid_entities)
cnn_model._construct_model()
invalid_entities2 = {"PAD": 0, "BACKGROUND": 1}
cnn_model.set_label_mapping(invalid_entities2)
cnn_model._reconstruct_model()
cnn_model.set_label_mapping(self.label_mapping)
# data for model
data_gen = [[
np.array([['test']]), # x_data
np.zeros((1, 3400, max(self.label_mapping.values()) + 1))
]
# y_data
]
cv_gen = data_gen
cnn_model._construct_model()
# set different labels
cnn_model.set_label_mapping(self.label_mapping)
history, f1, f1_report = cnn_model.fit(data_gen, cv_gen)
# test predict on just the text
cnn_model.predict(data_gen[0][0])
def test_param_validation(self):
# Make sure all parameters can be altered. Make sure non-valid params
# are caught
parameters = {
'max_length': 10,
'max_char_encoding_id': 11,
'size_fc': [64, 64],
'dropout': 0.9,
'size_conv': 11,
'default_label': "BACKGROUND",
'num_fil': [48 for _ in range(2)]
}
invalid_parameters = {
'max_length': -1,
'max_char_encoding_id': "words",
'size_fc': 5,
'dropout': 0.9,
'size_conv': 11,
'optimizer': 6,
'num_fil': [48 for _ in range(2)],
'fake_extra_param': "fails"
}
cnn_model = CharacterLevelCnnModel(label_mapping=self.label_mapping,
parameters=parameters)
cnn_model._construct_model()
self.assertDictEqual(parameters, cnn_model._parameters)
with self.assertRaises(ValueError):
CharacterLevelCnnModel(label_mapping=self.label_mapping,
parameters=invalid_parameters)
def test_model_construct(self):
# Default Model Construct
cnn_model = CharacterLevelCnnModel(label_mapping=self.label_mapping)
cnn_model._construct_model()
# Test Details
cnn_model.details()
expected_layers = [
"input_1",
"lambda",
"embedding",
"conv1d",
"dropout",
"batch_normalization",
"conv1d_1",
"dropout_1",
"batch_normalization_1",
"conv1d_2",
"dropout_2",
"batch_normalization_2",
"conv1d_3",
"dropout_3",
"batch_normalization_3",
"dense",
"dropout_4",
"dense_1",
"dropout_5",
"dense_2",
"tf_op_layer_ArgMax",
"thresh_arg_max_layer",
]
model_layers = [layer.name for layer in cnn_model._model.layers]
self.assertEqual(len(expected_layers), len(model_layers))
self.assertEqual(17, cnn_model.num_labels)
def test_param_validation(self):
# Make sure all parameters can be altered. Make sure non-valid params
# are caught
parameters = {
"max_length": 10,
"max_char_encoding_id": 11,
"size_fc": [64, 64],
"dropout": 0.9,
"size_conv": 11,
"default_label": "UNKNOWN",
"num_fil": [48 for _ in range(2)],
}
invalid_parameters = {
"max_length": -1,
"max_char_encoding_id": "words",
"size_fc": 5,
"dropout": 0.9,
"size_conv": 11,
"optimizer": 6,
"num_fil": [48 for _ in range(2)],
"fake_extra_param": "fails",
}
cnn_model = CharacterLevelCnnModel(label_mapping=self.label_mapping,
parameters=parameters)
cnn_model._construct_model()
self.assertDictEqual(parameters, cnn_model._parameters)
with self.assertRaises(ValueError):
CharacterLevelCnnModel(label_mapping=self.label_mapping,
parameters=invalid_parameters)
def test_validation_evaluate_and_classification_report(self, *mocks):
cnn_model = CharacterLevelCnnModel(self.label_mapping)
cnn_model._construct_model()
# validation data
val_gen = [[
np.array([['123 fake st']]),
np.zeros((1, 3400, max(self.label_mapping.values()) + 1))
]]
val_gen[0][1][:, :11, self.label_mapping['ADDRESS']] = 1
f1, f1_report = cnn_model._validate_training(val_gen, 32, True, True)
self.assertIsNotNone(f1)
self.assertIsNotNone(f1_report)
self.assertEqual(11, f1_report['ADDRESS']['support'])
def test_validation(self):
# model
cnn_model = CharacterLevelCnnModel(label_mapping=self.label_mapping)
cnn_model._construct_model()
# data for model
cv_data_gen = [[
np.array([['test']]), # x_data
np.zeros((1, 3400, max(self.label_mapping.values()) + 1))
]
# y_data
]
# validation
cnn_model._validate_training(cv_data_gen,
batch_size_test=32,
verbose_log=True,
verbose_keras=False)
def test_model_construct(self):
# Default Model Construct
cnn_model = CharacterLevelCnnModel(label_mapping=self.label_mapping)
cnn_model._construct_model()
# Test Details
cnn_model.details()
expected_layers = [
'input_1', 'lambda', 'embedding', 'conv1d', 'dropout',
'batch_normalization', 'conv1d_1', 'dropout_1',
'batch_normalization_1', 'conv1d_2', 'dropout_2',
'batch_normalization_2', 'conv1d_3', 'dropout_3',
'batch_normalization_3', 'dense', 'dropout_4', 'dense_1',
'dropout_5', 'dense_2', 'tf_op_layer_ArgMax',
'thresh_arg_max_layer'
]
model_layers = [layer.name for layer in cnn_model._model.layers]
self.assertEqual(len(expected_layers), len(model_layers))
self.assertEqual(17, cnn_model.num_labels)
def test_fit_and_predict_with_new_labels(self):
# Initialize model
cnn_model = CharacterLevelCnnModel(self.label_mapping)
# data for model
data_gen = [[
np.array([['test']]), # x_data
np.zeros((1, 3400, max(self.label_mapping.values()) + 1))
]
# y_data
]
cv_gen = data_gen
cnn_model._construct_model()
# fit with new labels
history, f1, f1_report = cnn_model.fit(
data_gen, cv_gen, label_mapping=self.label_mapping)
# predict after fitting on just the text
cnn_model.predict(data_gen[0][0])
