Using BigDL on Apache Spark to Improve the MLS Real Estate Search Experience at Scale with Sergey Ermolin and Dave Wetzel
0 likes493 views
The document outlines the implementation of a personalized real estate search ranking system using the BigDL framework on Microsoft Azure, focusing on image similarity as a search parameter. It describes the machine learning processes involved in training a deep learning model using a labeled dataset of real estate images, along with various technical components and infrastructure roles. The project aims to enhance real estate search results by ranking listings based on visual similarities of attached images.
Using BigDL on Apache Spark to Improve the MLS Real Estate Search Experience at Scale with Sergey Ermolin and Dave Wetzel
- 1. Dave Wetzel, COO and CTO, MLS Listings Sergey Ermolin, Solutions Architect, Intel Real Estate Search Ranking with BigDL Framework on Microsoft Azure Platform #ExpSAIS16
- 2. MLSListings Inc, Sunnyvale, California
- 3. MLSListings Business Use-Case: Personalized Visual Search Ranking Image similarity is an extra search parameter along with area, location, size, price, etc. If you looked at this house….. You will want to look at this one, too Business need: real estate search results need to be sorted based on image similarities of attached photos
- 4. 4 Implementation Example - 1 #ExpSAIS16
- 5. 5 Implementation Example - 2 #ExpSAIS16
- 6. LIVE DEMO 6
- 7. High-Level Data Workflow 7#ExpSAIS16 OData Media API Machine Learning Process Property Details WebPage API Ranked Images Top 10
- 8. 8 Labeled Dataset of Real Estate Images (Bing Search) BigDL VGG Arch BigDL Trained Model + = Data Engineering Deep Learning ● Long Compute. 8500 images ● 2 nodes, 28 cores/node. 3 minutes for a one single pass ● Model parameters are changing. ● Repeat until convergence ● But: only do once ! Note-1: Images are *not* stored in the model Note-2: you can trade compute resources for time. Compute TrainingDeep Learning Data Flow #ExpSAIS16
- 9. Deep Learning Data Flow 9 Inference Feature Vector BigDL Trained Model Image Class (Front, Bdr, Bath,…) House Style Tag (Ranch, Victorian,…) House Levels (1, 2..) Latent Features (25k entries) + = ● Short Compute. Real Time ● 1 node, 1 core/node ● Model parameters unchanged. ● Only run once per image ● But: need to do for every image in the searched dataset ! Compute Only #ExpSAIS16
- 10. 10 Labeled Dataset of Real Esate Images (Bing Search) BigDL VGG CNN BigDL Trained Model + = Feature Vector BigDL Trained Model Image Class (Front, Bdr, Bath,…) House Style Tag (Ranch, Victorian,…) House Levels (1, 2..) Latent Features (25k entries) + = TrainingInference Deep Learning Data Flow – putting it together #ExpSAIS16
- 11. 11 Deep Learning Data Flow Listing images BigDL Trained Model Feature Vectors MLS Query Landing Page BigDL Trained Model Cosine Similarity + Tag + Class Rank Top 10 MLS Listings Front End Real-time Ranking #ExpSAIS16 (Weighed)
- 12. 12 Deep Learning Data Flow Labeled Image Dataset (Bing Search) Big DL Model+ = Listing Images Big DL + Model Big DL + Model Cosine Similarity + Tag + Class Rank • Room Type (Class) • House Style (Tag) • House Levels (Tag) • Features (25k vec) Top 10 MLS Listings Front End #ExpSAIS16
- 13. 13 Implementation Example - 3 #ExpSAIS16
- 14. Implementation - BigDL Building BigDL Graph • Prepare Training/Validation data. • Image Transformer: – Image scale/crop – Channel color normalizing • Caffe Model Import • Render BigDL as SparkML Transformer • Create BigDL Linear SoftMax model • Define Classifier, SparkML Transformer • Set up SparkML Pipeline Executing BigDL Graph . Referethub https://github.com/intel-analytics/analytics-zoo
- 15. BigDL is an open-source distributed deep learning library for Apache Spark* that can run directly on top of existing Spark or Apache Hadoop* clusters Feature Parity & Model Exchange with TensorFlow*, Caffe*, Keras, Torch* Lower TCO and improved ease of use with existing infrastructure Deep Learning on Big Data Platform, Enabling Efficient Scale-Out BigDL Spark Core High BigDL: Performance Deep Learning for Apache Spark* on CPU Infrastructure No need to deploy costly GPUs, duplicate data, or suffer through scaling headaches! Designed and Optimized for Intel® Xeon® Ideal for DL Models TRAINING and INFERENCE Powered by Intel® MKL and multi-threaded programming #ExpSAIS16 15 https://github.com/intel-analytics/analytics-zoo
- 16. ModelsInteroperability Support • Model Snapshot • Long training work checkpoint • Model deployment and sharing • Fine-tune • Caffe/Torch/Tensorflow Model Support • Model file load • Easy to migrate your Caffe/Torch/Tensorflow code base to Spark • NEW - BigDL supports loading pre-defined Keras models (Keras 1.2.2) Caffe Model File Torch Model File Storage BigDL BigDL Model File Load Save Tensorflow Model File 16 #ExpSAIS16 #ExpSAIS16 https://github.com/intel-analytics/analytics-zoo
- 17. Visualization for Learning BigDL integration with TensorBoard • TensorBoard is a suite of web applications from Google for visualizing and understanding deep learning applications 17 #ExpSAIS16https://github.com/intel-analytics/analytics-zoo
- 18. 2018-BigDLAnalyticsZooStack Reference Use Cases Anomaly detection, sentiment analysis, fraud detection, chatbot, sequence prediction, etc. Built-In Algorithms and Models Image classification, object detection, text classification, recommendations, GAN, etc. Feature Engineering and Transformations Image, text, speech, 3D imaging, time series, etc. High-Level Pipeline APIs DataFrames, ML Pipelines, Autograd, Transfer Learning, etc. Runtime Environment Spark, BigDL, Python, etc. Making it easier to build end-to-end analytics + AI applications https://github.com/intel-analytics/analytics-zoo
- 19. Engineering Team 19 ● Data scientist, proficient in Machine Learning / Deep Learning ● Software Engineer, experience with Apache Spark. ● Technical project manager Domain Expertise: ● Machine Learning / Deep Learning, ● Python, Scala ● Software Engineer, Web API ● Software Engineer, Web UI Domain Expertise: ● OData, .net Core MSSQL ● C#, HTML, JavaScript #ExpSAIS16
- 20. “ROAD AHEAD” 20 “Fireplace in the living room” 1. Feature extraction and tagging. 2. Image-based listings search 3. Feature verification based on listing images. 4. Image-based compliance and quality check
- 21. 21
- 22. LIVE DEMO 22
- 24. MLS Listings Apps and Services 24#ExpSAIS16 Microsoft Azure RESO ODATA API homes.mlslistings.com MLSL Azure Web App Service Azure SQL DBData Feed Customers Full PayloadVOW PayloadIDX Payload OPENIDCONNECT Big DL
- 25. Roles and Responsibilities • Microsoft: Microsoft's data science team in Mountain View, CA participated in project discussions and provided Azure Data Science VM to deploy and train the deep learning model. • • Microsoft - Apache Spark Cloud Service Provider • • Intel - BigDL distributed Deep Learning Library • • MLSListings - RESO Web API Provider • Intel: Team members worked to integrate MLSListings’s OData Media Services to deploy a custom real estate image similarity comparison solution on Azure using Big DL. • MLSListings : MLSListings’s team working on new web portal provided Media API and worked on the user interface to integrate with Big DL API. 25#ExpSAIS16
- 26. BigDL: Python API • Support deep learning model training, evaluation, inference • Support Spark v1.6 - 2.2 • Support Python 2.7/3.5/3.6 • Based on PySpark, Python API in BigDL allows use of existing Python libs (Numpy, Scipy, Pandas, Scikit- learn, NLTK, Matplotlib, etc) 26#ExpSAIS16