This document discusses TensorFlow Wide & Deep models and their applications. It covers feature columns, model estimators, exporting models for TensorFlow Serving, and tools for working with models. It also provides examples of how Alibaba's food delivery service Ele.me has used machine learning techniques like predicting order times and recommending discounts. The document promotes Google Cloud ML Engine for distributed training and serving models at scale.
5. Confidential & Proprietary
Wide & Deep Model
• Submitted on 2016-06-24
• Jointly train a wide linear model and a deep feed-forward neural network
• Productionized and evaluated the system on Google Play
https://arxiv.org/abs/1606.07792
17. Confidential & Proprietary
Model ⼯工具
tensorflow/python/tools/saved_model_cli.py
https://youtu.be/sqYdlSF0BI8?t=30m4s
# What meta_graphs are in a model?
saved_model_cli show --dir /tmp/model_dir
# What signatures are in a meta_graph?
saved_model_cli show --dir /tmp/model_dir --tag_set serve
# What input & output tensors are in a signature?
saved_model_cli show --dir /tmp/model_dir --tag_set serve --signature_def serving_default
# Run a graph
saved_model_cli show --dir /tmp/model_dir --tag_set serve --signature_def xxx --inputs x1=/xxx/
xxx.npy --input_exprs 'x2=np.ones((3,1))'
20. Confidential & Proprietary
a (gRPC) client to call TensorFlow Serving
pip install tensorflow-serving-api
# Build the input data as an Example object
# tf.Example:
# https://github.com/tensorflow/tensorflow/blob/master/tensorflow/core/example/example.proto
# tf.Features:
# https://github.com/tensorflow/tensorflow/blob/master/tensorflow/core/example/feature.proto
python2.7