ONNX tutorial with TensorFlow

(opens in a new tab)

import turboml as tb

!pip install -U tensorflow tf2onnx onnx==1.14.1 scikit-learn

Tensorflow - Standard Model Training

The following blocks of code define a standard tensorflow training code. This is completely independent of TurboML.

import pandas as pd
import tf2onnx.convert
import matplotlib.pyplot as plt
import tensorflow as tf
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Dense

transactions = tb.datasets.FraudDetectionDatasetFeatures()
labels = tb.datasets.FraudDetectionDatasetLabels()

joined_df = pd.merge(transactions.df, labels.df, on="transactionID", how="right")
joined_df

X = joined_df.drop("is_fraud", axis=1)
numerical_fields = [
    "transactionAmount",
    "localHour",
    "isProxyIP",
    "digitalItemCount",
    "physicalItemCount",
]
 
feats = X[numerical_fields]
targets = joined_df["is_fraud"].astype(int)

from sklearn.model_selection import train_test_split
 
X_train, X_test, y_train, y_test = train_test_split(
    feats, targets, test_size=0.2, random_state=42
)

model = Sequential()
model.add(Dense(64, activation="relu", input_shape=(X_train.shape[1],)))
model.add(Dense(64, activation="relu"))
model.add(Dense(2, activation="softmax"))

model.summary()

model.compile(
    optimizer="adam", loss="sparse_categorical_crossentropy", metrics=["accuracy"]
)
 
num_epochs = 10
model.fit(X_train, y_train, epochs=num_epochs, batch_size=64, verbose=1)

_, accuracy = model.evaluate(X_test, y_test)
print("Accuracy:", accuracy)

Export model to ONNX format

Exporting a model to ONNX format depends on the framework. Tutorials for different frameworks can be found at https://github.com/onnx/tutorials#converting-to-onnx-format (opens in a new tab)

onnx_model_path = "tensorflow_model.onnx"
input_signature = [
    tf.TensorSpec([1, len(numerical_fields)], tf.float32, name="keras_tensor")
]
model.output_names = ["output"]
onnx_model, _ = tf2onnx.convert.from_keras(model, input_signature, opset=13)
 
onnx_model_bytes = onnx_model.SerializeToString()

Create an ONNX model with TurboML

Now that we've converted the model to ONNX format, we can deploy it with TurboML.

transactions = transactions.to_online(id="transactions", load_if_exists=True)
labels = labels.to_online(id="transaction_labels", load_if_exists=True)
 
features = transactions.get_model_inputs(numerical_fields=numerical_fields)
label = labels.get_model_labels(label_field="is_fraud")

tb.set_onnx_model("tensorflowmodel", onnx_model_bytes)
onnx_model = tb.ONNX(model_save_name="tensorflowmodel")

deployed_model = onnx_model.deploy("onnx_model_tf", input=features, labels=label)

deployed_model.add_metric("WindowedAUC")

model_auc_scores = deployed_model.get_evaluation("WindowedAUC")
plt.plot([model_auc_score.metric for model_auc_score in model_auc_scores])

ONNX - Scikit-Learn TF-IDF Example