def test_model_creation(self):
N, C, H, W = 2, 3, 5, 5
input1 = keras.layers.Input(shape=(H, W, C))
x1 = keras.layers.Dense(8, activation='relu')(input1)
input2 = keras.layers.Input(shape=(H, W, C))
x2 = keras.layers.Dense(8, activation='relu')(input2)
maximum_layer = keras.layers.Maximum()([x1, x2])
out = keras.layers.Dense(8)(maximum_layer)
model = keras.models.Model(inputs=[input1, input2], outputs=out)
trial1 = np.random.rand(N, H, W, C).astype(np.float32, copy=False)
trial2 = np.random.rand(N, H, W, C).astype(np.float32, copy=False)
predicted = model.predict([trial1, trial2])
self.assertIsNotNone(predicted)
converted_model_7 = onnxmltools.convert_keras(model, target_opset=7)
converted_model_5 = onnxmltools.convert_keras(model, target_opset=5)
self.assertIsNotNone(converted_model_7)
self.assertIsNotNone(converted_model_5)
opset_comparison = converted_model_7.opset_import[0].version > converted_model_5.opset_import[0].version
self.assertTrue(opset_comparison)
def _test_one_to_one_operator_coreml(self, keras_model, x):
# Verify Keras-to-CoreML-to-ONNX path
coreml_model = None
try:
coreml_model = coremltools.converters.keras.convert(keras_model)
except (AttributeError, ImportError) as e:
warnings.warn(
"Unable to test due to an error in coremltools '{0}'".format(
e))
onnx_model = None if coreml_model is None else onnxmltools.convert_coreml(
coreml_model)
self.assertTrue(onnx_model or coreml_model is None)
if self._no_available_inference_engine():
return
y_reference = keras_model.predict(x)
y_produced = evaluate_deep_model(onnx_model, x)
self.assertTrue(np.allclose(y_reference, y_produced))
# Verify Keras-to-ONNX path
onnx_model = onnxmltools.convert_keras(keras_model)
y_produced = evaluate_deep_model(onnx_model, x)
self.assertTrue(np.allclose(y_reference, y_produced))
def convert_keras2onnx(model, output):
import onnxmltools
from keras.models import load_model
keras_model = load_model(model)
onnx_model = onnxmltools.convert_keras(keras_model)
onnxmltools.utils.save_model(onnx_model, output)
def save(self, using='dnn'):
if using == 'sklearn':
filename = 'scikit_model'
pickle.dump(self.model, open(filename, 'wb'))
else:
onnx_model = onnxmltools.convert_keras(self.model)
return onnx_model
def test_custom_op(self):
N, C, H, W = 2, 3, 5, 5
x = np.random.rand(N, H, W, C).astype(np.float32, copy=False)
model = Sequential()
model.add(
Conv2D(2,
kernel_size=(1, 2),
strides=(1, 1),
padding='valid',
input_shape=(H, W, C),
data_format='channels_last'))
model.add(ScaledTanh(0.9, 2.0))
model.add(
MaxPooling2D((2, 2), strides=(2, 2), data_format='channels_last'))
model.compile(optimizer='sgd', loss='mse')
converted_model = onnxmltools.convert_keras(
model, custom_conversion_functions={ScaledTanh: custom_activation})
actual = model.predict(x)
self.assertIsNotNone(actual)
self.assertIsNotNone(converted_model)
dump_data_and_model(x.astype(np.float32),
model,
converted_model,
basename="KerasCustomOp-Out0",
context=dict(ScaledTanh=ScaledTanh))
def _test_one_to_one_operator_core_keras(self, keras_model, x):
y_reference = keras_model.predict(x)
onnx_model = onnxmltools.convert_keras(keras_model)
y_produced = _evaluate(onnx_model, x)
self.assertTrue(np.allclose(y_reference, y_produced))
def test_channel_last(self):
N, C, H, W = 2, 3, 5, 5
x = np.random.rand(N, H, W, C).astype(np.float32, copy=False)
model = Sequential()
model.add(
Conv2D(2,
kernel_size=(1, 2),
strides=(1, 1),
padding='valid',
input_shape=(H, W, C),
data_format='channels_last')) # , activation='softmax')
model.add(
MaxPooling2D((2, 2), strides=(2, 2), data_format='channels_last'))
model.compile(optimizer='sgd', loss='mse')
converted_model = onnxmltools.convert_keras(
model, channel_first_inputs=[model.inputs[0].name])
expected = model.predict(x)
self.assertIsNotNone(expected)
self.assertIsNotNone(converted_model)
try:
import onnxruntime
sess = onnxruntime.InferenceSession(
converted_model.SerializeToString())
actual = sess.run(
[], {
sess.get_inputs()[0].name:
np.transpose(x.astype(np.float32), [0, 3, 1, 2])
})
self.assertTrue(np.allclose(expected, actual, rtol=1.e-3))
except ImportError:
pass
def from_keras(
model: keras.models.Model,
opset: int = DEFAULT_OPSET,
):
onnx_model = onnxmltools.convert_keras(model, target_opset=opset)
logger.info('keras to onnx converted successfully')
return onnx_model
def convertToKerasAndBack(fileName, outputFileName, inputs):
ret = None
onnxModel = onnx.load(fileName)
kwArgs = dict()
if "shuffle" in fileName:
kwArgs["input_shapes"] = [1, 3, 224, 224]
import shufflenet
import keras
kerasModel = shufflenet.ShuffleNet(groups=3)
kerasModel.load_weights(
"keras-shufflenet/weights/ShuffleNet_1X_g3_br_0.25_373.hdf5")
kerasModel.compile(optimizer=keras.optimizers.SGD(lr=.05,
decay=5e-4,
momentum=0.9),
metrics=['accuracy'],
loss='categorical_crossentropy')
ret = kerasModel.predict(inputs[0][0].transpose(0, 2, 3, 1))
else:
kerasModel = onnx_to_keras(onnxModel,
getOnnxInputNames(fileName),
verbose=False,
**kwArgs)
#tf.keras.utils.plot_model(kerasModel, show_shapes=True)
backconvOnnxModel = onnxmltools.convert_keras(kerasModel)
onnxmltools.utils.save_model(backconvOnnxModel, outputFileName)
return ret
def testRunModelConv(self):
# keras model
N, C, H, W = 2, 3, 5, 5
x = np.random.rand(N, H, W, C).astype(np.float32, copy=False)
model = Sequential()
model.add(Conv2D(2, kernel_size=(1, 2), strides=(1, 1), padding='valid', input_shape=(H, W, C),
data_format='channels_last'))
model.add(ScaledTanh(0.9, 2.0))
model.add(MaxPooling2D((2, 2), strides=(2, 2), data_format='channels_last'))
model.compile(optimizer='sgd', loss='mse')
actual = model.predict(x)
self.assertIsNotNone(actual)
# conversion
converted_model = onnxmltools.convert_keras(model, custom_conversion_functions={ScaledTanh: custom_activation})
self.assertIsNotNone(converted_model)
# runtime
content = converted_model.SerializeToString()
rt = onnxrt.InferenceSession(content)
input = {rt.get_inputs()[0].name: x}
actual_rt = rt.run(None, input)
self.assertEqual(len(actual_rt), 1)
np.testing.assert_allclose(actual, actual_rt[0], rtol=1e-05, atol=1e-08)
def main(args):
spec = importlib.util.spec_from_file_location("model", args.dave_path)
model = importlib.util.module_from_spec(spec)
spec.loader.exec_module(model)
onnx_model = onnxmltools.convert_keras(model.DAVE())
onnxmltools.utils.save_model(onnx_model, args.output_path)
def _test_one_to_one_operator_coreml_channels_last(self, keras_model, x):
'''
There are two test paths. One is Keras-->CoreML-->ONNX and the other one is Keras-->ONNX.
Keras-->CoreML-->ONNX:
Keras computation path:
[N, C, H, W] ---> numpy transpose ---> [N, H, W, C] ---> keras convolution --->
[N, H, W, C] ---> numpy transpose ---> [N, C, H, W]
ONNX computation path:
[N, C, H, W] ---> ONNX convolution ---> [N, C, H, W]
The reason for having extra transpose's in the Keras path is that CoreMLTools doesn't not handle channels_last
flag properly. Precisely, oreMLTools always converts Conv2D under channels_first mode.
Keras-->ONNX
Keras computation path:
[N, C, H, W] ---> numpy transpose ---> [N, H, W, C] ---> keras convolution --->
[N, H, W, C]
ONNX computation path:
[N, C, H, W] ---> numpy transpose ---> [N, H, W, C] ---> ONNX convolution ---> [N, H, W, C]
'''
# Verify Keras-to-CoreML-to-ONNX path
coreml_model = None
try:
coreml_model = coremltools.converters.keras.convert(keras_model)
except (AttributeError, ImportError) as e:
warnings.warn(
"Unable to test due to an error in coremltools '{0}'.".format(
e))
onnx_model_p1 = None if coreml_model is None else onnxmltools.convert_coreml(
coreml_model)
onnx_model_p2 = onnxmltools.convert_keras(keras_model)
self.assertTrue(onnx_model_p1 or coreml_model is None)
self.assertTrue(onnx_model_p2)
if self._no_available_inference_engine():
return
if isinstance(x, list):
x_t = [np.transpose(_, [0, 2, 3, 1]) for _ in x]
else:
x_t = np.transpose(x, [0, 2, 3, 1])
y_reference = np.transpose(keras_model.predict(x_t), [0, 3, 1, 2])
y_produced = evaluate_deep_model(onnx_model_p1, x)
self.assertTrue(np.allclose(y_reference, y_produced))
# Verify Keras-to-ONNX path
y_reference = np.transpose(y_reference, [0, 2, 3, 1])
y_produced = evaluate_deep_model(onnx_model_p2, x_t)
self.assertTrue(np.allclose(y_reference, y_produced, atol=1e-6))
def export_onnx(keras_model, output_path):
"""Export the model to the ONNX format.
Args:
keras_model: the loaded yolo/tiny-yolo model.
output_path: output path of the ONNX model
"""
onnx_model = onnxmltools.convert_keras(keras_model)
onnxmltools.utils.save_model(onnx_model, output_path)
def _test_one_to_one_operator_core_keras(self, keras_model, x):
y_reference = keras_model.predict(x)
onnx_model = onnxmltools.convert_keras(keras_model)
if find_keras_backend():
y_produced = evaluate_deep_model(onnx_model, x)
self.assertTrue(np.allclose(y_reference, y_produced))
else:
warnings.warn("cntk or caffe2 are not available")
def main(args):
os.makedirs(os.path.dirname(args.output_path), exist_ok=True)
with open(args.model_path) as json_file:
keras_model = model_from_json(json_file.read())
keras_model.load_weights(args.weights_path)
onnx_model = onnxmltools.convert_keras(keras_model)
optimized_onnx_model = concat_outputs(onnx_model)
onnxmltools.utils.save_model(optimized_onnx_model, args.output_path)
def get_onnx_model(model_format,
model,
initial_types: list = None,
final_types: list = None):
if model_format == ModelFormat.KERAS:
return onnxmltools.convert_keras(model)
if model_format == ModelFormat.SK_LEARN:
return onnxmltools.convert_sklearn(model, initial_types=initial_types)
if model_format == ModelFormat.TENSORFLOW:
return onnxmltools.convert_tensorflow(model)
def log_model(model, artifact_path, model_name=None):
import mlflow
import mlflow.onnx
import onnx
import onnxmltools
onnx_model = onnxmltools.convert_keras(model, artifact_path)
print("onnx_model.type:",type(onnx_model))
mlflow.onnx.log_model(onnx_model, artifact_path, registered_model_name=model_name)
mlflow.set_tag("version.onnx",onnx.__version__)
mlflow.set_tag("version.onnxtools",onnxmltools.__version__)
def save(self, filename=None):
if filename == None:
raise ValueError(
'Name Error: to save the model you need to specify the filename')
pickle.dump(self.final_model, open(filename + '.pkl', 'wb'))
self.final_model.save(filename + '.h5')
onnx_model = onnxmltools.convert_keras(self.final_model)
onnxmltools.utils.save_model(onnx_model, filename + '.onnx')
def save_onnx(model_name):
keras_path = "tmp.h5"
onnx_path = f'{model_name}.onnx'
keras_model = load_model(keras_path)
onnx_model = onnxmltools.convert_keras(keras_model)
onnx_model.graph.input[0].type.tensor_type.shape.dim[0].dim_value = 1
onnx_model.graph.output[0].type.tensor_type.shape.dim[0].dim_value = 1
onnx.checker.check_model(onnx_model)
onnx.save(onnx_model, onnx_path)
def log_model(model, artifact_path):
import mlflow
import mlflow.onnx
import onnx
import onnxmltools
onnx_model = onnxmltools.convert_keras(model, artifact_path)
print("onnx_model.type:", type(onnx_model))
mlflow.onnx.log_model(onnx_model, artifact_path)
mlflow.set_tag("onnx_version", onnx.__version__)
mlflow.set_tag("onnxtools_version", onnxmltools.__version__)
def create_onnx_model(self):
self.construct_matfile()
W, b, a = self.matDict["W"], self.matDict["b"], self.matDict[
"act_fcns"]
model = create_nn(W, b, a)
self.final_output_path = os.path.join(self.outputFilePath,
self.originalFilename) + '.onnx'
onnx_model = onnxmltools.convert_keras(model)
onnxmltools.utils.save_model(onnx_model, self.final_output_path)
self.originalFile.close()
def evaluate_model(trainX, trainy):
verbose, epochs, batch_size = 0, 10, 32
# random seed for reproducibility
seed = 7
np.random.seed(seed)
# k-fold validation
X = trainX
Y = trainy
n_split = 10
# evaluation metrics
accuracylist = list()
precisionlist = list()
recalllist = list()
f1scorelist = list()
# repeat calculator
number = 0
for train_index, test_index in KFold(n_split).split(X, Y):
x_train, x_test = X[train_index], X[test_index]
y_train, y_test = Y[train_index], Y[test_index]
# create model
model = create_model(trainX, trainy)
# fit network
model.fit(x_train, y_train, epochs=epochs, batch_size=batch_size, verbose=verbose)
# save model
model.save('./models/model' + str(number+1) + '.h5')
keras_model = load_model('./models/model' + str(number+1) + '.h5')
onnx_model = onnxmltools.convert_keras(keras_model)
onnxmltools.utils.save_model(onnx_model, './models/model' + str(number+1) + '.onnx')
# evaluate model
y_predict = model.predict(x_test, batch_size=batch_size, verbose=verbose)
print('#' + str(number+1) + '#' + 'y_predict:', y_predict)
y_predict = np.argmax(y_predict, axis=1)
y_test = np.argmax(y_test, axis=1)
y_true = np.reshape(y_test, [-1])
y_pred = np.reshape(y_predict, [-1])
accuracy = accuracy_score(y_true, y_pred)
accuracylist.append(accuracy)
precision = precision_score(y_true, y_pred, average='macro')
precisionlist.append(precision)
recall = recall_score(y_true, y_pred, average='macro')
recalllist.append(recall)
f1score = f1_score(y_true, y_pred, average='macro')
f1scorelist.append(f1score)
number = number + 1
# summarize results
accuracy_mean, accuracy_std = summarize_results(accuracylist)
print('K-fold Accuracy: %.2f%% (+/-%.2f)' % (accuracy_mean*100, accuracy_std))
precision_mean, precision_std = summarize_results(precisionlist)
print('K-fold Precision: %.2f%% (+/-%.2f)' % (precision_mean*100, precision_std))
recall_mean, recall_std = summarize_results(recalllist)
print('K-fold Recall: %.2f%% (+/-%.2f)' % (recall_mean*100, recall_std))
f1score_mean, f1score_std = summarize_results(f1scorelist)
print('K-fold F1 Score: %.2f%% (+/-%.2f)' % (f1score_mean*100, f1score_std))
def _test_one_to_one_operator_keras(self, keras_model, x):
y_reference = keras_model.predict(x)
onnx_model = onnxmltools.convert_keras(keras_model)
if not self._no_available_inference_engine():
y_produced = evaluate_deep_model(onnx_model, x)
self.assertTrue(np.allclose(y_reference, y_produced))
else:
self.assertIsNotNone(onnx_model)
warnings.warn("None of onnx inference engine is available.")
if self._no_available_inference_engine():
return
def create_onnx_model(self):
# Convert the Keras model into ONNX
if self.no_json:
model = models.load_model(self.pathToOriginalFile)
else:
model = self.load_files(self.jsonFile, self.pathToOriginalFile)
self.final_output_path = os.path.join(self.outputFilePath,
self.originalFilename) + '.onnx'
onnx_model = onnxmltools.convert_keras(model)
# Save as protobuf
onnxmltools.utils.save_model(onnx_model, self.final_output_path)
self.originalFile.close()
def keras2onnx(args):
import keras
# Load your Keras model
keras_model = keras.models.load_model(args.model)
# Convert the Keras model into ONNX
onnx_model = onnxmltools.convert_keras(keras_model,
initial_types=args.initial_types,
target_opset=int(args.target_opset))
# Save as protobuf
onnxmltools.utils.save_model(onnx_model, args.output_onnx_path)
def save(self, filename=None):
if filename == None:
raise ValueError(
'Name Error: to save the model you need to specify the filename')
if isinstance(self.proxy_model, (sklearn.DimensionalityReductionModel)):
raise AttributeError("'SklearnKerasRegressor' object has no attribute 'save'")
pickle.dump(self.final_model, open(filename + '.pkl', 'wb'))
self.final_model.save(filename + '.h5')
onnx_model = onnxmltools.convert_keras(self.final_model)
onnxmltools.utils.save_model(onnx_model, filename + '.onnx')
def convert_model_to_onnx(model: Any):
"""
Helper function to convert a ML model to onnx format
"""
if isinstance(model, model_classes_keras):
return onnxmltools.convert_keras(model)
if isinstance(model, model_classes_sklearn):
return onnxmltools.convert_sklearn(model)
if 'xgboost' in model.__repr__():
return onnxmltools.convert_sklearn(model)
if isinstance(model, model_classes_scipy):
raise Exception("Pytorch models not yet supported to onnx")
else:
raise Exception(
"Attempt to convert unsupported model to onnx: {model}")
def _test_one_to_one_operator_core(self, keras_model, x):
# Verify Keras-to-CoreML-to-ONNX path
coreml_model = coremltools.converters.keras.convert(keras_model)
onnx_model = onnxmltools.convert_coreml(coreml_model)
y_reference = keras_model.predict(x)
y_produced = _evaluate(onnx_model, x)
self.assertTrue(np.allclose(y_reference, y_produced))
# Verify Keras-to-ONNX path
onnx_model = onnxmltools.convert_keras(keras_model)
y_produced = _evaluate(onnx_model, x)
self.assertTrue(np.allclose(y_reference, y_produced))
def saveModel(model, label_persona_dict):
onnx_model = onnxmltools.convert_keras(model)
onnxmltools.utils.save_model(onnx_model, 'example.onnx')
model_json = model.to_json()
with open("model.json", "w") as json_file:
json_file.write(model_json)
# serialize weights to HDF5
model.save_weights("model.h5")
logging.info("Saved model to disk")
dict_writer = open("./label_persona_dict.json", 'w')
serialized_dict = json.dumps(label_persona_dict)
dict_writer.write(serialized_dict)
dict_writer.close()
logging.info("Saved dict to disk")
def training(model, train_set, test_set):
history = model.fit_generator(train_set,
steps_per_epoch=100,
epochs=100,
validation_data=test_set,
validation_steps=100)
model.save("saved_model.h5", True, True)
onnx = onnxmltools.convert_keras(model)
onnxmltools.save_model(onnx, "converted.onnx")
with open("history.json", "w+") as file:
file.write(str(history.history))
acc_plot(history)
loss_plot(history)
