Chris Fregly, Founder @ PipelineAI, will walk you through a real-world, complete end-to-end Pipeline-optimization example. We highlight hyper-parameters - and model pipeline phases - that have never been exposed until now.
While most Hyperparameter Optimizers stop at the training phase (ie. learning rate, tree depth, ec2 instance type, etc), we extend model validation and tuning into a new post-training optimization phase including 8-bit reduced precision weight quantization and neural network layer fusing - among many other framework and hardware-specific optimizations.
Next, we introduce hyperparameters at the prediction phase including request-batch sizing and chipset (CPU v. GPU v. TPU).
Lastly, we determine a PipelineAI Efficiency Score of our overall Pipeline including Cost, Accuracy, and Time. We show techniques to maximize this PipelineAI Efficiency Score using our massive PipelineDB along with the Pipeline-wide hyper-parameter tuning techniques mentioned in this talk.
Bio
Chris Fregly is Founder and Applied AI Engineer at PipelineAI, a Real-Time Machine Learning and Artificial Intelligence Startup based in San Francisco.
He is also an Apache Spark Contributor, a Netflix Open Source Committer, founder of the Global Advanced Spark and TensorFlow Meetup, author of the O’Reilly Training and Video Series titled, "High Performance TensorFlow in Production with Kubernetes and GPUs."
Previously, Chris was a Distributed Systems Engineer at Netflix, a Data Solutions Engineer at Databricks, and a Founding Member and Principal Engineer at the IBM Spark Technology Center in San Francisco.
Hyper-Parameter Tuning Across the Entire AI Pipeline GPU Tech Conference San Jose March 2018
1. HYPER-PARAMETER TUNING ACROSS THE ENTIRE
AI PIPELINE: MODEL TRAINING TO PREDICTING
GPU TECH CONFERENCE -- SAN JOSE, MARCH 2018
CHRIS FREGLY
FOUNDER @ PIPELINEAI
2. KEY TAKE-AWAYS
With PipelineAI, You Can…
§ Hyper-Parameter Tuning From Training to Inference
§ Generate Hardware-Specific Pipeline Optimizations
§ Deploy & Compare Optimizations in Live Production
§ Perform Continuous Model Training & Data Labeling
3. AGENDA
Part 0: Introductions and Setup
Part 1: Optimize TensorFlow Training
Part 2: Optimize TensorFlow Serving
Part 3: Advanced Model Serving + Routing
4. INTRODUCTIONS: ME
§ Chris Fregly, Founder & Engineer @ PipelineAI
§ Formerly Netflix, Databricks, IBM Spark Tech
§ Founder @ Advanced Spark TensorFlow Meetup
§ Please Join Our 75,000+ Global Members!!
Contact Me
chris@pipeline.ai
@cfregly
Global Locations
* San Francisco
* Chicago
* Austin
* Washington DC
* Dusseldorf
* London
5. INTRODUCTIONS: YOU
You Want To …
§ Perform Hyper-Parameter Tuning Across *Entire* Pipeline
§ Measure Results of Tuning Both Offline *and* Online
§ Deploy Models Rapidly, Safely, *Directly* in Production
§ Trace and Explain *Live* Model Predictions
6. PIPELINEAI IS OPEN SOURCE
§ https://github.com/PipelineAI/pipeline/
§ Please Star this GitHub Repo!
§ “Each Star is Worth $1,500 in Seed Money”
- A Prominent Venture Capitalist in Silicon Valley
http://jrvis.com/red-dwarf/
9. PIPELINEAI TERMINOLOGY
§ “Flask-App Falacy”: Flask is Not Enough for Production-izing ML/AI Models
§ “Pipeline”: All Phases Including Train, Validate, Optimize, Deploy, and Predict
§ “Experiment”: Across All Environments from Research Lab to Live Production
§ “Turning Knobs”: Hyper-Parameter Tuning Across All Phases of the Pipeline
§ “Model Serving”: Models Serving Predictions in Live Production
§ “Runtime”: Execution Environment for Any Phase of Pipeline (TensorRT, Caffe)
§ “Train-to-Serve”: Training with Intent to Serve Predictions
§ “Train-Serving Skew”: Model Performs Poorly on Live Data
§ “Post-Training Optimization”: Prepare Model and Runtime for Fast Inference
http://NoFlaskApp.com
10. Any Runtime
Any Device CPU, GPU, TPU, IoT
Any Network and System Configuration
Any Clouud and On-Premise Environment
AnyModel
AnyLanguage
AnyFramework
AnyHyper-
Parameter
1,000,000’s of
Model + Runtime Pipeline
Combinations
We Find the Best Combinations
For Your Model and Workload!
WHOLE-PIPELINE HYPER-PARAMETER TUNING
12. WHOLE-PIPELINE HYPER-PARAMETERS
Training: Hyperparameters
pipelinedb.add("learning_rate", 0.025)
pipelinedb.add(”batch_size", 8192)
pipelinedb.add(”num_epochs", 100)
^^ THIS IS WHERE MOST DATA SCIENTISTS END BECAUSE ^^
^^ THEY HAVE NO WAY OF COLLECTING ANYTHING MORE ^^
^^ UNTIL NOW! ^^
pipelinedb.add("ec2_instance_type", "g3.4xlarge”)
pipelinedb.add("utilized_memory_gigabyte", 20)
pipelinedb.add(“network_speed_gigabit”, 10)
pipelinedb.add("training_precision_bits", 16)
pipelinedb.add("accelerator_type", "nvidia_gpu_v100") # google_tpu
pipelinedb.add(“cpu_to_accelerator_network_type", “pcie”) # nvlink
pipelinedb.add(“cpu_to_accelerator_network_bandwidth_gigabit”, 100)
Training: Results
pipelinedb.add("training_accuracy_percent", 95)
pipelinedb.add(“validation_accuracy_percent", 94)
pipelinedb.add("training_auc", 0.70)
pipelinedb.add(“validation_auc", 0.69)
pipelinedb.add(”time_to_train_seconds", 0.69)
Optimization: Hyperparameters
pipelinedb.add(”batch_norm_fusing", True)
pipelinedb.add("weight_quantization_bits", 8) # 2-bit, 7-bit
Optimization: Results (Collected At End of Optimization)
pipelinedb.add("weight_quantization_reduction_percent", 50)
Inference: Hyperparameters
pipelinedb.add("runtime_type", ”tfserving") # python,tensorrt
Pipelinedb.add(“runtime_chip”, “gpu”)
pipelinedb.add("model_type", "tensorflow") # caffe, scikit
pipelinedb.add("request_batch_window_ms", 10)
pipelinedb.add("request_batch_size", 1000)
Inference: Results (Every ~15 Mins Inside PipelineAI Runtime)
pipelinedb.add("latency_99_percentile_ms", 5)
pipelinedb.add("cost_per_prediction_usd", 0.000001)
pipelinedb.add("24_hr_auc", 0.70)
pipelinedb.add("48_hr_auc", 0.30)
Training Optimizing
Inferencing
13. WHY EMPHASIS ON MODEL INFERENCE?
Model Training
Batch & Boring
Offline in Research Lab
Pipeline Ends at Training
No Insight into Live Production
Small Number of Data Scientists
Optimizations Are Very Well-Known
Real-Time & Exciting!!
Online in Live Production
No Ability To Turn Inference Knobs (Yet)
Extend Model Validation Into Production
Huuuuuuge Number of Application Users
Inference Optimizations Not Yet Explored
<<<
Model Inference
100’s Training Jobs per Day 1,000,000’s Predictions per Sec
14. GROWTH IN ML/AI MODELS
2017 2026
Data
Scientists
44,000
11,500,000
$39 Billion in 2017
$2 Trillion by 2026
2017 2026
Models
Trained
50,000,000
200,000
2017 2026
Model
Predictions
250,000,000,000
4,000,000
2016 2026 2016 2026 2016 2026
15. MODEL DEPLOYMENT OPTIONS
§ AWS SageMaker
§ Released Nov 2017 @ Re-invent
§ Custom Docker Images for Training/Serving (ie. PipelineAI Images)
§ Distributed TensorFlow Training through Estimator API
§ Traffic Splitting for A/B Model Testing
§ Google Cloud ML Engine
§ Mostly Command-Line Based
§ Driving TensorFlow Open Source API (ie. Estimator API)
§ Azure ML
§ On-Premise Docker, Docker Swarm, Kubernetes, Mesos
PipelineAI Supports All
Hybrid-Cloud, On-Prem,
and Air-Gap Deployments!
16. WHOLE-PIPELINE OPTIMIZATION OPTIONS
§ Model Training Optimizations
§ Model Hyper-Parameters (ie. Learning Rate)
§ Reduced Precision (ie. FP16 Half Precision)
§ Model Optimizations to Prepare for Inference
§ Quantize Model Weights + Activations From 32-bit to 8-bit
§ Fuse Neural Network Layers Together
§ Model Inference Runtime Optimizations
§ Runtime Config: Request Batch Size, etc
§ Different Runtime: TensorFlow Serving CPU/GPU, Nvidia TensorRT
17. NVIDIA TENSOR-RT RUNTIME
§ Post-Training Model Optimizations
§ Specific to Nvidia GPUs
§ GPU-Optimized Prediction Runtime
§ Alternative to TensorFlow Serving
§ PipelineAI Supports TensorRT!
18. TENSORFLOW LITE OPTIMIZING CONVERTER
§ Post-Training Model Optimizations
§ Currently Supports iOS and Android
§ On-Device Prediction Runtime
§ Low-Latency, Fast Startup
§ Selective Operator Loading
§ 70KB Min - 300KB Max Runtime Footprint
§ Supports Accelerators (GPU, TPU)
§ Falls Back to CPU without Accelerator
§ Java and C++ APIs
bazel build tensorflow/contrib/lite/toco:toco &&
./bazel-bin/third_party/tensorflow/contrib/lite/toco/toco
--input_file=frozen_eval_graph.pb
--output_file=tflite_model.tflite
--input_format=TENSORFLOW_GRAPHDEF --output_format=TFLITE
--inference_type=QUANTIZED_UINT8
--input_shape="1,224, 224,3"
--input_array=input
--output_array=outputs
--std_value=127.5 --mean_value=127.5
19. PIPELINEAI QUICK START
§ http://quickstart.pipeline.ai
§ Any Model, Any Training Runtime, Any Prediction Runtime
§ Support for Docker, Docker Swarm, Kubernetes, Mesos
§ Package Model+Runtime into a Docker Image
§ Emphasizes Immutable Deployment and Infrastructure
§ Same Image Across All Environments
§ No Library or Dependency Surprises from Laptop to Production
§ Allows Tuning Offline and Online Model+Runtime Together
20. STEP 1: BUILD MODEL+TRAINING SERVER
§ Train Model with Specific Hyper-Parameters
§ Monitor and Compare Validation Accuracy
§ Tune Hyper-Parameters to Improve Accuracy
pipeline train-server-build --model-name=mnist
--model-tag=A
--model-type=tensorflow
--model-path=./tensorflow/mnist/0.025/model
Build Model
Training Server A
(Learning Rate 0.025)
pipeline train-server-build --model-name=mnist
--model-tag=B
--model-type=tensorflow
--model-path=./tensorflow/mnist/0.050/model
Build Model
Training Server B
(Learning Rate 0.050)
21. STEP 2: TRAIN, MEASURE, TUNE
§ Train Model with Specific Hyper-Parameters
§ Monitor abnd Compare Validation Accuracy
§ Tune Hyper-Parameters to Improve Accuracy
pipeline train-server-start --model-name=mnist
--model-tag=A
--input-host-path=./tensorflow/mnist/input
--output-host-path=./tensorflow/mnist/output
--train-args= "--learning-rate=0.025 --batch-size=128"
Train
Model A
(Learning Rate 0.025)
pipeline train-server-start --model-name=mnist
--model-tag=B
--input-host-path=./tensorflow/mnist/input
--output-host-path=./tensorflow/mnist/output
--train-args= "--learning-rate=0.025 --batch-size=128"
Train
Model B
(Learning Rate 0.050)
23. STEP 4: BUILD MODEL+PREDICTION SERVER
pipeline predict-server-build --model-name=mnist
--model-tag=C
--model-type=tensorflow
--model-runtime=tensorrt
--model-chip=gpu
--model-path=./tensorflow/mnist/
Build Local
Model Server C
TensorRT GPU
pipeline predict-server-build --model-name=mnist
--model-tag=A
--model-type=tensorflow
--model-runtime=tfserving
--model-chip=cpu
--model-path=./tensorflow/mnist/
Build Local
Model Server A
TF Serving CPU
pipeline predict-server-build --model-name=mnist
--model-tag=B
--model-type=tensorflow
--model-runtime=tfserving
--model-chip=gpu
--model-path=./tensorflow/mnist/
Build Local
Model Server B
TF Serving GPU
Same Model,
3 Different
Prediction
Runtimes
24. STEP 5: PREDICT, MEASURE, TUNE (LOCAL)
§ Perform Mini-Load Test on Local Model Server
§ Immediate Feedback on Prediction Performance
§ Compare to Previous Model+Runtime Variations
§ Gain Intuition Before Pushing to Prod
pipeline predict-server-start --model-name=mnist
--model-tag=A
--memory-limit=2G
pipeline predict-http-test --model-endpoint-url=http://localhost:8080
--test-request-path=test_request.json
--test-request-concurrency=1000
Start Local
Predict Load Test
Start Local
Model Server
25. STEP 6: DEPLOY, MEASURE, TUNE (IN PROD)
§ Deploy from CLI or Jupyter Notebook
§ Tear-Down and Rollback Models Quickly
§ Shadow Canary: Deploy to 20% Live Traffic
§ Split Canary: Deploy to 97-2-1% Live Traffic
pipeline predict-kube-start --model-name=mnist
--model-tag=BStart Cluster B
pipeline predict-kube-start --model-name=mnist
--model-tag=CStart Cluster C
pipeline predict-kube-start --model-name=mnist
--model-tag=AStart Cluster A
pipeline predict-kube-route --model-name=mnist
--model-split-tag-and-weight-dict='{"A":97, "B":2, "C”:1}'
--model-shadow-tag-list='[]'
Route Live Traffic
26. STEP 7: OPTIMIZE, MEASURE, RE-DEPLOY
§ Prepare Model for Predicting
§ Simplify Network, Reduce Size
§ Reduce Precision -> Fast Math
§ Some Tools
§ Graph Transform Tool (GTT)
§ tfcompile
After Training
After
Optimizing!
pipeline optimize --optimization-list=[‘quantize_weights’,‘tfcompile’]
--model-name=mnist
--model-tag=A
--model-path=./tensorflow/mnist/model
--model-inputs=[‘x’]
--model-outputs=[‘add’]
--output-path=./tensorflow/mnist/optimized_model
Linear
Regression
Model Size: 70MB –> 70K (!)
27. STEP 8: EVALUATE MODEL+RUNTIME VARIANT
§ Offline, Batch Metrics
§ Validation + Training Accuracy
§ CPU + GPU Utilization
§ Online, Live Prediction Values
§ Compare Relative Precision
§ Newly-Seen, Streaming Data
§ Online, Real-Time Metrics
§ Response Time, Throughput
§ Cost ($) Per Prediction
29. STEP 10: SHIFT TRAFFIC TO BEST VARIANT
§ A/B Tests
§ Inflexible and Boring
§ Multi-Armed Bandits
§ Adaptive and Exciting!
pipeline predict-kube-route --model-name=mnist
--model-split-tag-and-weight-dict='{"A":1, "B":2, "C”:97}’
--model-shadow-tag-list='[]'
Dynamically Route
Traffic to Winning
Model+Runtime
30. PIPELINE PROFILING AND TUNING
§ Instrument Code to Generate “Timelines” for Any Metric
§ Analyze with Google Web
Tracing Framework (WTF)
§ Can Also Monitor CPU with top, GPU with nvidia-smi
http://google.github.io/tracing-framework/
from tensorflow.python.client import timeline
trace =
timeline.Timeline(step_stats=run_metadata.step_stats)
with open('timeline.json', 'w') as trace_file:
trace_file.write(
trace.generate_chrome_trace_format(show_memory=True))
31. MODEL AND ENSEMBLE TRACING/AUDITING
§ Necessary for Model Explain-ability
§ Fine-Grained Request Tracing
§ Used for Model Ensembles
32. VIEW REAL-TIME PREDICTION STREAMS
§ Visually Compare Real-time Predictions
Features and
Inputs
Predictions and
Confidences
Model B Model CModel A
33. CONTINUOUS DATA LABELING AND FIXING
§ Identify and Fix Borderline (Unconfident) Predictions
§ Fix Predictions Along Class Boundaries
§ Facilitate ”Human in the Loop”
§ Path to Crowd-Sourced Labeling
§ Retrain with Newly-Labeled Data
§ Game-ify the Labeling Process
34. CONTINUOUS MODEL TRAINING
§ The Holy Grail of Machine Learning
§ Kafka, Kinesis, Spark Streaming, Flink, Storm, Heron
PipelineAI Supports
Continuous Model Training
35. AGENDA
Part 0: Introductions and Setup
Part 1: Optimize TensorFlow Training
Part 2: Optimize TensorFlow Serving
Part 3: Advanced Model Serving + Traffic Routing
36. AGENDA
Part 1: Optimize TensorFlow Training
§ GPUs and TensorFlow
§ Feed, Train, and Debug TensorFlow Models
§ TensorFlow Distributed Cluster Model Training
§ Optimize Training with JIT XLA Compiler
37. SETTING UP TENSORFLOW WITH GPUS
§ Very Painful!
§ Especially inside Docker
§ Use nvidia-docker
§ Especially on Kubernetes!
§ Use the Latest Kubernetes (with Init Script Support)
§ http://pipeline.ai for GitHub + DockerHub Links
39. VOLTA V100 AND TENSOR CORES
§ 84 Streaming Multiprocessors (SM’s)
§ 5,376 GPU Cores
§ 640 Tensor Cores (ie. Google TPU)
§ Can Perform 640 FP16 4x4 Matrix Multiplies
§ 120 TFLOPS = 4x FP32 and 10x FP64
§ Allows Mixed FP16/FP32 Precision Operations
§ Matrix Dims Should be Multiples of 8
§ More Shared Memory
§ New L0 Instruction Cache
§ Faster L1 Data Cache
40. GPU HALF-PRECISION SUPPORT
§ FP32: “Full Precision”, FP16: “Half Precision”
§ Two(2) FP16’s in 1 FP32 GPU Core
§ 2x Throughput!
§ Lower Precision is OK
§ Deep learning is approximate
§ The Network Matters Most
§ Not individual neuron accuracy
41. MORE ON HALF-PRECISION
§ 1997: Related Work by SGI
§ Commercial Request from ILM in 2002
§ Implemented in Silicon by Nvidia in 2002
§ Supported by Pascal P100 and Volta V100
42. MORE ON REDUCED-PRECISION
§ Less Precision => Less Memory & Bandwidth
=> Faster Math & Less Energy
§ Fits into Smaller Places Close to ALU’s
§ 4-bit, 2-bit, 1-bit (?!) Quantization
§ More Layers Help Maintain Accuracy at Reduced Precision
§ Tip: Scale and Center Dynamic Range at Each Layer
§ Otherwise, FP16’s become 0 - model may not converge!
43. GPU: 4-WAY DOT PRODUCT OF 8-BIT INTS
§ GPU Hardware and CUDA Support
§ Compute Capability (CC) >= 6.1
44. FP16 VS. INT8
§ FP16 Has Larger Dynamic Range Than INT8
§ Larger Dynamic Range Allows Higher Precision
§ Truncated FP32 Dynamic Range Higher Than FP16
§ Not IEEE 754 Standard, But Worth Exploring
45. ENABLING FP16 IN TENSORFLOW
§ Harder Than You Think!
§ TPUs are 16-bit Native
GPU’s With CC 5.3+ (Only), Set the Following:
TF_FP16_MATMUL_USE_FP32_COMPUTE=0
TF_FP16_CONV_USE_FP32_COMPUTE=0
TF_XLA_FLAGS=--xla_enable_fast_math=1
Pascal P100 Volta V100
46. FP32 VS. FP16 ON AWS GPU INSTANCES
FP16 Half Precision
87.2 T ops/second for p3 Volta V100
4.1 T ops/second for g3 Tesla M60
1.6 T ops/second for p2 Tesla K80
FP32 Full Precision
15.4 T ops/second for p3 Volta V100
4.0 T ops/second for g3 Tesla M60
3.3 T ops/second for p2 Tesla K80
47. § Tesla K80
§ Pascal P100
§ Volta V100 (Beta)
§ TPU (Beta, Google Cloud Only)
GOOGLE CLOUD GPU + TPU
48. GOOGLE CLOUD TPUS
§ Attach/Detach As Needed
§ Scale In/Out As Needed
§ 180 TFlops per Device
§ TPU Pod = 64 TPUs
= 11.5 PetaFlops
§ $6.50 per TPU Hour
§ Supports 16-bit TensorFlow
49. V100 AND CUDA 9
§ Independent Thread Scheduling - Finally!!
§ Similar to CPU fine-grained thread synchronization semantics
§ Allows GPU to yield execution of any thread
§ Still Optimized for SIMT (Same Instruction Multi-Thread)
§ SIMT units automatically scheduled together
§ Explicit Synchronization
P100 V100
New CUDA
Thread Cooperative Groups
https://devblogs.nvidia.com/cooperative-groups/
50. GPU CUDA PROGRAMMING
§ Barbaric, But Fun Barbaric
§ Must Know Hardware Very Well
§ Hardware Changes are Painful
§ Use the Profilers & Debuggers
51. CUDA STREAMS
§ Asynchronous I/O Transfer
§ Overlap Compute and I/O
§ Keep GPUs Saturated!
§ Used Heavily by TensorFlow
Bad
Good
Bad
Good
53. NUMBA AND PYCUDA
§ Numba is Drop-In Replacement for Numpy
§ PyCuda is Python Binding for CUDA
54. AGENDA
Part 1: Optimize TensorFlow Training
§ GPUs and TensorFlow
§ Feed, Train, and Debug TensorFlow Models
§ TensorFlow Distributed Cluster Model Training
§ Optimize Training with JIT XLA Compiler
55. TRAINING TERMINOLOGY
§ Tensors: N-Dimensional Arrays
§ ie. Scalar, Vector, Matrix
§ Operations: MatMul, Add, SummaryLog,…
§ Graph: Graph of Operations (DAG)
§ Session: Contains Graph(s)
§ Feeds: Feed Inputs into Placeholder
§ Fetches: Fetch Output from Operation
§ Variables: What We Learn Through Training
§ aka “Weights”, “Parameters”
§ Devices: Hardware Device (GPU, CPU, TPU, ...)
-TensorFlow-
Trains
Variables
-User-
Fetches
Outputs
-User-
Feeds
Inputs
-TensorFlow-
Performs
Operations
-TensorFlow-
Flows
Tensors
with tf.device(“/cpu:0,/gpu:15”):
57. TENSORFLOW GRAPH EXECUTION
§ Lazy Execution by Default
§ Similar to Spark
§ Eager Execution
§ Similar to PyTorch
§ "Linearize” Execution Minimizes RAM
§ Useful on Single GPU with Limited RAM
§ May Need to Re-Compute (CPU/GPU) vs Store (RAM)
58. OPERATION PARALLELISM
§ Inter-Op (Between-Op) Parallelism
§ By default, TensorFlow runs multiple ops in parallel
§ Useful for low core and small memory/cache envs
§ Set to one (1)
§ Intra-Op (Within-Op) Parallelism
§ Different threads can use same set of data in RAM
§ Useful for compute-bound workloads (CNNs)
§ Set to # of cores (>=2)
59. TENSORFLOW MODEL
§ MetaGraph
§ Combines GraphDef and Metadata
§ GraphDef
§ Architecture of your model (nodes, edges)
§ Metadata
§ Asset: Accompanying assets to your model
§ SignatureDef: Maps external to internal tensors
§ Variables
§ Stored separately during training (checkpoint)
§ Allows training to continue from any checkpoint
§ Variables are “frozen” into Constants when preparing for inference
GraphDef
x
W
mul add
b
MetaGraph
Metadata
Assets
SignatureDef
Tags
Version
Variables:
“W” : 0.328
“b” : -1.407
60. STOCHASTIC GRADIENT DESCENT (SGD)
§ Or “Simply Go Down” J
§ Small Batch Sizes Are Ideal
§ But not too small!
§ Parallel, Distributed Training Across Devices
§ Each device calculates gradients on small batch
§ Gradients averaged across all devices
§ Training is Fast, Batches are Small
63. TENSORFLOW + SPARK OPTIONS
§ TensorFlow on Spark (Yahoo!)
§ TensorFrames <-Dead Project->
§ Separate Clusters for Spark and TensorFlow
§ Spark: Boring Batch ETL
§ TensorFlow: Exciting AI Model Training and Serving
§ Hand-Off Point is S3, HDFS, Google Cloud Storage
64. TENSORFLOW + KAFKA
§ TensorFlow Dataset API Now Supports Kafka!!
from tensorflow.contrib.kafka.python.ops import kafka_dataset_ops
repeat_dataset = kafka_dataset_ops.KafkaDataset(topics,
group="test",
eof=True)
.repeat(num_epochs)
batch_dataset = repeat_dataset.batch(batch_size)
…
65. TENSORFLOW I/O
§ TFRecord File Format
§ TensorFlow Python and C++ Dataset API
§ Python Module and Packaging
§ Comfort with Python’s Lack of Strong Typing
§ C++ Concurrency Constructs
§ Protocol Buffers
§ Old Queue API
§ GPU/CUDA Memory Tricks And a Lot of Coffee!
66. FEED TENSORFLOW TRAINING PIPELINE
§ Training is Limited by the Ingestion Pipeline
§ Number One Problem We See Today
§ Scaling GPUs Up / Out Doesn’t Help
§ GPUs are Heavily Under-Utilized
§ Use tf.dataset API for best perf
§ Efficient parallel async I/O (C++)
Tesla K80 Volta V100
67. DON’T USE FEED_DICT!!
§ feed_dict Requires Python <-> C++ Serialization
§ Not Optimized for Production Ingestion Pipelines
§ Retrieves Next Batch After Current Batch is Done
§ Single-Threaded, Synchronous
§ CPUs/GPUs Not Fully Utilized!
§ Use Queue or Dataset APIs
§ Queues are old & complex
sess.run(train_step, feed_dict={…}
68. DETECT UNDERUTILIZED CPUS, GPUS
§ Instrument Code to Generate “Timelines”
§ Analyze with Google Web
Tracing Framework (WTF)
§ Monitor CPU with top, GPU with nvidia-smi
http://google.github.io/tracing-framework/
from tensorflow.python.client import timeline
trace =
timeline.Timeline(step_stats=run_metadata.step_stats)
with open('timeline.json', 'w') as trace_file:
trace_file.write(
trace.generate_chrome_trace_format(show_memory=True))
69. QUEUES
§ More than Traditional Queue
§ Uses CUDA Streams
§ Perform I/O, Pre-processing, Cropping, Shuffling, …
§ Pull from HDFS, S3, Google Storage, Kafka, ...
§ Combine Many Small Files into Large TFRecord Files
§ Use CPUs to Free GPUs for Compute
§ Helps Saturate CPUs and GPUs
70. QUEUE CAPACITY PLANNING
§ batch_size
§ # examples / batch (ie. 64 jpg)
§ Limited by GPU RAM
§ num_processing_threads
§ CPU threads pull and pre-process batches of data
§ Limited by CPU Cores
§ queue_capacity
§ Limited by CPU RAM (ie. 5 * batch_size)
71. TF.DTYPE
§ tf.float32, tf.int32, tf.string, etc
§ Default is usually tf.float32
§ Most TF operations support numpy natively
# Tuple of (tf.float32 scalar, tf.int32 array of 100 elements)
(tf.random_uniform([1]), tf.random_uniform([1, 100], dtype=tf.int32))
72. TF.TRAIN.FEATURE
§ Three(3) Feature Types
§ Bytes
§ Float
§ Int64
§ Actually, They Are Lists of 0..* Values of 3 Types Above
§ BytesList
§ FloatList
§ Int64List
73. TF.TRAIN.FEATURES
§ Map of {String -> Feature}
§ Better Name is “FeatureMap”
§ Organize Feature into Categories
§ Access Feature Using
Features[’feature_name’]
75. TF.TRAIN.FEATURELISTS
§ Map of {String -> FeatureList}
§ Better Name is “FeatureListMap”
§ Organize FeatureList into Categories
§ Access FeatureList Using
FeatureLists[’feature_list_name’]
76. TF.TRAIN.EXAMPLE
§ Key-Value Dictionary
§ String -> tf.train.Feature
§ Not a Self-Describing Format (?!)
§ Must Establish Schema Upfront by Writers and Readers
§ Must Obey the Following Conventions
§ Feature K must be of Type T in all Examples
§ Feature K can be omitted, default can be configured
§ If Feature K exists as empty, no default is applied
77. TF.TFRECORD
§ Contains many tf.train.Example’s
=> tf.train.Example contains many tf.train.Feature’s
=> tf.train.Feature contains BytesList, FloatList, Int64List
§ Record-Oriented Format of Binary Strings (ProtoBuffer)
§ Must Convert tf.train.Example to Serialized String
§ Use tf.train.Example.SerializeToString()
§ Used for Large Scale ML/AI Training
§ Not Meant for Random or Non-Sequential Access
§ Compression: GZIP, ZLIB
uint64 length
uint32 masked_crc32_of_length
byte data[length]
uint32 masked_crc32_of_data
78. EMBRACE BINARY FORMATS!
§ Unreadable and Scary, But Much More Efficient
§ Better Use of Memory and Disk Cache
§ Faster Copying and Moving
§ Smaller on the Wire
I
79. CONVERTING MNIST DATA TO TFRECORD
def convert_to_tfrecord(data, name):
images = data.images
labels = data.labels
num_examples = data.num_examples
rows = images.shape[1]
cols = images.shape[2]
depth = images.shape[3]
filename = os.path.join(FLAGS.directory, name + '.tfrecords’)
with tf.python_io.TFRecordWriter(filename) as writer:
for index in range(num_examples):
image_raw = images[index].tostring()
example = tf.train.Example(
features = tf.train.Features(
feature = {'height': tf.train.Feature(int64_list=tf.train.Int64List(value=[rows])),
'width': tf.train.Feature(int64_list=tf.train.Int64List(value=[cols])),
'depth': tf.train.Feature(int64_list=tf.train.Int64List(value=[depth])),
'label': tf.train.Feature(int64_list=tf.train.Int64List(value=[index])),
'image_raw': tf.train.Feature(bytes_list=tf.train.BytesList(value=[image_raw]))
}))
writer.write(example.SerializeToString())
tf.python_io.TFRecordWriter
80. READING TF.TFRECORD’S
§ tf.data.TFRecordDatasetß Preferred (Dataset API)
§ tf.TFRecordReader()ß Not Preferred (Queue API)
§ tf.python_io.tf_record_iterator ß Preferred
§ Used as Python Generator
for serialized_example in tf.python_io.tf_record_iterator(filename):
example = tf.train.Example()
example.ParseFromString(serialized_example)
image_raw example.features.feature['image_raw’].string_list.value
height = example.features.feature[‘height'].int32_list.value[0]
…
81. DE-SERIALIZING TF.TFRECORD’S
feature_map = {'height': tf.train.Feature(int64_list=tf.train.Int64List(value=[rows])),
'width': tf.train.Feature(int64_list=tf.train.Int64List(value=[cols])),
'depth': tf.train.Feature(int64_list=tf.train.Int64List(value=[depth])),
'label': tf.train.Feature(int64_list=tf.train.Int64List(value=[index])),
'image_raw': tf.train.Feature(bytes_list=tf.train.BytesList(value=[image_raw]))
deserialized_features = tf.parse_single_example(serialized_example, features=feature_map)
# Cast height from String to int32
height = tf.cast(deserialized_features[‘height’], tf.int32)
…
# Convert raw image from string to float32
image_raw = tf.decode_raw(deserialized_features[‘image_raw'], tf.float32)
82. MORE TF.TRAIN.FEATURE CONSTRUCTS
§ tf.VarLenFeature
§ tf.FixedLenFeature, tf.FixedLenSequenceFeature
§ tf.SparseFeature
feature_map = {'height': tf.FixedLenFeature((), tf.int32, …)),
…
'image_raw': tf.train.VarLenFeature(tf.string, …))
deserialized_features = tf.parse_single_example(serialized_example, features=feature_map)
# Cast height from String to int32
height = tf.cast(deserialized_features[‘height’], tf.int32)
…
# Convert raw image from string to float32
image_raw = tf.decode_raw(deserialized_features[‘image_raw'], tf.float32)
83. TF.DATA.DATASET
tf.Tensor => tf.data.Dataset
Functional Transformations
Python Generator => tf.data.Dataset
Dataset.from_tensors((features, labels))
Dataset.from_tensor_slices((features, labels))
TextLineDataset(filenames)
dataset.map(lambda x: tf.decode_jpeg(x))
dataset.repeat(NUM_EPOCHS)
dataset.batch(BATCH_SIZE)
def generator():
while True:
yield ...
dataset.from_generator(generator, tf.int32)
Dataset => One-Shot Iterator
Dataset => Initializable Iter
iter = dataset.make_one_shot_iterator()
next_element = iter.get_next()
while …:
sess.run(next_element)
iter = dataset.make_initializable_iterator()
sess.run(iter.initializer, feed_dict=PARAMS)
next_element = iter.get_next()
while …:
sess.run(next_element)
TIP: Use Dataset.prefetch() and parallel version of Dataset.map()
86. CUSTOM TF.PY_FUNC() TRANSFORMATION
§ Custom Python Function
§ Similar to Spark Python UDF (Eek!)
§ You Will Suffer a Big Performance Penalty
§ Try to Use TensorFlow-Native Operations
§ Remember, you can build your own in C++!
87. TF.DATA.ITERATOR TYPES
§ One Shot: Iterates Once Through the Dataset
§ Currently, best Iterator to use with Estimator API
§ Initializable: Runs iterator.initializer() Once
§ Re-Initializable: Runs iterator.initializer() Many
§ Ie. Random shuffling between iterations (epochs) of training
§ Feedable: Switch Between Different Dataset
§ Uses Feed and Placeholder to explicitly feed the iterator
§ Doesn’t require initialization when switching
88. TF.DATA.ITERATOR SIMPLE EXAMPLE
dataset = tf.data.Dataset.range(5)
iterator = dataset.make_initializable_iterator()
next_element = iterator.get_next()
# Typically `result` will be the output of a model, or an optimizer's
# training operation.
result = tf.add(next_element, next_element)
sess.run(iterator.initializer)
while True:
try:
sess.run(result) # è 0, 2, 4, 6, 8
except tf.errors.OutOfRangeError:
print(‘End of dataset…’)
break
89. TF.DATA.ITERATOR TEXT EXAMPLE
filenames = ["/var/data/file1.txt", "/var/data/file2.txt"]
dataset = tf.data.TextLineDataset(filenames)
filenames = ["/var/data/file1.txt", "/var/data/file2.txt"]
dataset = tf.data.Dataset.from_tensor_slices(filenames)
dataset = dataset.flat_map(
lambda filename: (
tf.data.TextLineDataset(filename)
.skip(1)
.filter(lambda line: tf.not_equal(tf.substr(line, 0, 1), "#"))))
§ Skip 1st Header Line and Comment Lines Starting with `#`
90. TF.DATA.ITERATOR NUMPY EXAMPLE
# Load the training data into two NumPy arrays, for example using `np.load()`.
with np.load("/var/data/training_data.npy") as data:
features = data["features"]
labels = data["labels"]
# Assume that each row of `features` corresponds to the same row as `labels`.
assert features.shape[0] == labels.shape[0]
features_placeholder = tf.placeholder(features.dtype, features.shape)
labels_placeholder = tf.placeholder(labels.dtype, labels.shape)
dataset = tf.data.Dataset.from_tensor_slices((features_placeholder, labels_placeholder))
# …Your Dataset Transformations…
iterator = dataset.make_initializable_iterator()
sess.run(iterator.initializer, feed_dict={features_placeholder: features,
labels_placeholder: labels})
91. TF.DATA.ITERATOR TFRECORD EXAMPLE
filenames = tf.placeholder(tf.string, shape=[None])
dataset = tf.data.TFRecordDataset(filenames)
dataset = dataset.map(...) # Parse the record into tensors.
dataset = dataset.repeat() # Repeat the input indefinitely.
dataset = dataset.batch(32) # Batches of size 32
iterator = dataset.make_initializable_iterator()
# You can feed the initializer with the appropriate filenames for the current
# phase of execution, e.g. training vs. validation.
# Initialize `iterator` with training data.
training_filenames = ["/var/data/file1.tfrecord", "/var/data/file2.tfrecord"]
sess.run(iterator.initializer, feed_dict={filenames: training_filenames})
# Initialize `iterator` with validation data.
validation_filenames = ["/var/data/validation1.tfrecord", ...]
sess.run(iterator.initializer, feed_dict={filenames: validation_filenames})
92. FUTURE OF DATASET API
§ Replaces Queue API
§ More Functional Operators
§ Automatic GPU Data Staging and Pre-Fetching
§ Under-utilized GPUs Assisting with Data Ingestion
§ More Profiling and Recommendations for Ingestion
93. TF.ESTIMATOR.ESTIMATOR (1/2)
§ Supports Keras!
§ Unified API for Local + Distributed
§ Provide Clear Path to Production
§ Enable Rapid Model Experiments
§ Provide Flexible Parameter Tuning
§ Enable Downstream Optimizing & Serving Infra( )
§ Nudge Users to Best Practices Through Opinions
§ Provide Hooks/Callbacks to Override Opinions
94. TF.ESTIMATOR.ESTIMATOR (2/2)
§ “Train-to-Serve” Design
§ Create Custom Estimator or Re-Use Canned Estimator
§ Hides Session, Graph, Layers, Iterative Loops (Train, Eval, Predict)
§ Hooks for All Phases of Model Training and Evaluation
§ Load Input: input_fn()
§ Train: model_fn() and train()
§ Evaluate: eval_fn() and evaluate()
§ Performance Metrics: Loss, Accuracy, …
§ Save and Export: export_savedmodel()
§ Predict: predict() Uses the slow sess.run()
https://github.com/GoogleCloudPlatform/cloudml-samples
/blob/master/census/customestimator/
95. TF.CONTRIB.LEARN.EXPERIMENT
§ Easier-to-Use Distributed TensorFlow
§ Same API for Local and Distributed
§ Combines Estimator with input_fn()
§ Used for Training, Evaluation, & Hyper-Parameter Tuning
§ Distributed Training Defaults to Data-Parallel & Async
§ Cluster Configuration is Fixed at Start of Training Job
§ No Auto-Scaling Allowed, but That’s OK for Training
Note: The Experiment API Will Likely Be Deprecated Soon
96. ESTIMATOR + EXPERIMENT CONFIGS
§ TF_CONFIG
§ Special environment variable for config
§ Defines ClusterSpec in JSON incl. master, workers, PS’s
§ Distributed mode ‘{“environment”:“cloud”}’
§ Local: ‘{environment”:“local”, {“task”:{”type”:”worker”}}’
§ RunConfig: Defines checkpoint interval, output directory,
§ HParams: Hyper-parameter tuning parameters and ranges
§ learn_runner creates RunConfig before calling run() & tune()
§ schedule is set based on {”task”:{”type”:…}}
TF_CONFIG=
'{
"environment": "cloud",
"cluster":
{
"master":["worker0:2222”],
"worker":["worker1:2222"],
"ps": ["ps0:2222"]
},
"task": {"type": "ps",
"index": "0"}
}'
97. ESTIMATOR + KERAS
§ Distributed TensorFlow (Estimator) + Easy to Use (Keras)
§ tf.keras.estimator.model_to_estimator()
# Instantiate a Keras inception v3 model.
keras_inception_v3 = tf.keras.applications.inception_v3.InceptionV3(weights=None)
# Compile model with the optimizer, loss, and metrics you'd like to train with.
keras_inception_v3.compile(optimizer=tf.keras.optimizers.SGD(lr=0.0001, momentum=0.9),
loss='categorical_crossentropy',
metric='accuracy')
# Create an Estimator from the compiled Keras model.
est_inception_v3 = tf.keras.estimator.model_to_estimator(keras_model=keras_inception_v3)
# Treat the derived Estimator as you would any other Estimator. For example,
# the following derived Estimator calls the train method:
est_inception_v3.train(input_fn=my_training_set, steps=2000)
98. “CANNED” ESTIMATORS
§ Commonly-Used Estimators
§ Pre-Tested and Pre-Tuned
§ DNNClassifer, TensorForestEstimator
§ Always Use Canned Estimators If Possible
§ Reduce Lines of Code, Complexity, and Bugs
§ Use FeatureColumn to Define & Create Features
Custom vs. Canned
@ Google, August 2017
99. ESTIMATOR + DATASET API
def input_fn():
def generator():
while True:
yield ...
my_dataset = tf.data.dataset.from_generator(generator, tf.int32)
# A one-shot iterator automatically initializes itself on first use.
iter = my_dataset.make_one_shot_iterator()
# The return value of get_next() matches the dataset element type.
images, labels = iter.get_next()
return images, labels
# The input_fn can be used as a regular Estimator input function.
estimator = tf.estimator.Estimator(…)
estimator.train(train_input_fn=input_fn, …)
106. TF.CONTRIB.LEARN.HEAD (OBJECTIVES)
§ Single-Objective Estimator
§ Single classification prediction
§ Multi-Objective Estimator
§ One (1) classification prediction
§ One(1) final layer to feed into next model
§ Multiple Heads Used to Ensemble Models
§ Treats neural network as a feature engineering step
§ Supported by TensorFlow Serving
107. TF.LAYERS
§ Standalone Layer or Entire Sub-Graphs
§ Functions of Tensor Inputs & Outputs
§ Mix and Match with Operations
§ Assumes 1st Dimension is Batch Size
§ Handles One (1) to Many (*) Inputs
§ Metrics are Layers
§ Loss Metric (Per Mini-Batch)
§ Accuracy and MSE (Across Mini-Batches)
108. TF.FEATURE_COLUMN
§ Used by Canned Estimator
§ Declaratively Specify Training Inputs
§ Converts Sparse to Dense Tensors
§ Sparse Features: Query Keyword, ProductID
§ Dense Features: One-Hot, Multi-Hot
§ Wide/Linear: Use Feature-Crossing
§ Deep: Use Embeddings
109. TF.FEATURE_COLUMN EXAMPLE
§ Continuous + One-Hot + Embedding
deep_columns = [
age,
education_num,
capital_gain,
capital_loss,
hours_per_week,
tf.feature_column.indicator_column(workclass),
tf.feature_column.indicator_column(education),
tf.feature_column.indicator_column(marital_status),
tf.feature_column.indicator_column(relationship),
# To show an example of embedding
tf.feature_column.embedding_column(occupation, dimension=8),
]
110. FEATURE CROSSING
§ Create New Features by Combining Existing Features
§ Limitation: Combinations Must Exist in Training Dataset
base_columns = [
education, marital_status, relationship, workclass, occupation, age_buckets
]
crossed_columns = [
tf.feature_column.crossed_column(
['education', 'occupation'], hash_bucket_size=1000),
tf.feature_column.crossed_column(
['age_buckets', 'education', 'occupation'], hash_bucket_size=1000)
]
111. SEPARATE TRAINING + EVALUATION
§ Separate Training and Evaluation Clusters
§ Evaluate Upon Checkpoint
§ Avoid Resource Contention
§ Training Continues in Parallel with Evaluation
Training
Cluster
Evaluation
Cluster
Parameter Server
Cluster
112. BATCH (RE-)NORMALIZATION (2015, 2017)
§ Each Mini-Batch May Have Wildly Different Distributions
§ Normalize per Batch (and Layer)
§ Faster Training, Learns Quicker
§ Final Model is More Accurate
§ TensorFlow is already on 2nd Generation Batch Algorithm
§ First-Class Support for Fusing Batch Norm Layers
§ Final mean + variance Are Folded Into Graph Later
-- (Almost) Always Use Batch (Re-)Normalization! --
z = tf.matmul(a_prev, W)
a = tf.nn.relu(z)
a_mean, a_var = tf.nn.moments(a, [0])
scale = tf.Variable(tf.ones([depth/channels]))
beta = tf.Variable(tf.zeros ([depth/channels]))
bn = tf.nn.batch_normalizaton(a, a_mean, a_var,
beta, scale, 0.001)
113. DROPOUT (2014)
§ Training Technique
§ Prevents Overfitting
§ Helps Avoid Local Minima
§ Inherent Ensembling Technique
§ Creates and Combines Different Neural Architectures
§ Expressed as Probability Percentage (ie. 50%)
§ Boost Other Weights During Validation & Prediction
Perform Dropout
(Training Phase)
Boost for Dropout
(Validation & Prediction Phase)
0%
Dropout
50%
Dropout
114. BATCH NORM, DROPOUT + ESTIMATOR API
§ Must Specify Evaluation or Training Mode
§ These Will Behave Differently Depending on Mode
115. SAVED MODEL FORMAT
§ Different Format than Traditional Exporter
§ Contains Checkpoints, 1..* MetaGraph’s, and Assets
§ Export Manually with SavedModelBuilder
§ Estimator.export_savedmodel()
§ Hooks to Generate SignatureDef
§ Use saved_model_cli to Verify
§ Used by TensorFlow Serving
§ New Standard Export Format? (Catching on Slowly…)
116. TENSORFLOW DEBUGGER
§ Step through Operations
§ Inspect Inputs and Outputs
§ Wrap Session in Debug Session
sess = tf.Session(config=config)
sess =
tf_debug.LocalCLIDebugWrapperSession(sess)
https://www.tensorflow.org/programmers_guide/debugger
117. AGENDA
Part 1: Optimize TensorFlow Training
§ GPUs and TensorFlow
§ Train, Inspect, and Debug TensorFlow Models
§ TensorFlow Distributed Cluster Model Training
§ Optimize Training with JIT XLA Compiler
118. SINGLE NODE, MULTI-GPU TRAINING
§ cpu:0
§ By default, all CPUs
§ Requires extra config to target a CPU
§ gpu:0..n
§ Each GPU has a unique id
§ TF usually prefers a single GPU
§ xla_cpu:0, xla_gpu:0..n
§ “JIT Compiler Device”
§ Hints TensorFlow to attempt JIT Compile
with tf.device(“/cpu:0”):
with tf.device(“/gpu:0”):
with tf.device(“/gpu:1”):
GPU 0 GPU 1
119. DISTRIBUTED, MULTI-NODE TRAINING
§ TensorFlow Automatically Inserts Send and Receive Ops into Graph
§ Parameter Server Synchronously Aggregates Updates to Variables
§ Nodes with Multiple GPUs will Pre-Aggregate Before Sending to PS
Worker0 Worker0
Worker1
Worker0 Worker1 Worker2
gpu0 gpu1
gpu2 gpu3
gpu0 gpu1
gpu2 gpu3
gpu0 gpu1
gpu2 gpu3
gpu0
gpu1
gpu0
gpu0
Single
Node
Multiple
Nodes
120. DATA PARALLEL VS. MODEL PARALLEL
§ Data Parallel (“Between-Graph Replication”)
§ Send exact same model to each device
§ Each device operates on partition of data
§ ie. Spark sends same function to many workers
§ Each worker operates on their partition of data
§ Model Parallel (“In-Graph Replication”)
§ Send different partition of model to each device
§ Each device operates on all data
§ Difficult, but required for larger models with lower-memory GPUs
121. SYNCHRONOUS VS. ASYNCHRONOUS
§ Synchronous
§ Nodes compute gradients
§ Nodes update Parameter Server (PS)
§ Nodes sync on PS for latest gradients
§ Asynchronous
§ Some nodes delay in computing gradients
§ Nodes don’t update PS
§ Nodes get stale gradients from PS
§ May not converge due to stale reads!
122. CHIEF WORKER
§ Chief Defaults to Worker Task 0
§ Task 0 is guaranteed to exist
§ Performs Maintenance Tasks
§ Writes log summaries
§ Instructs PS to checkpoint vars
§ Performs PS health checks
§ (Re-)Initialize variables at (re-)start of training
123. NODE AND PROCESS FAILURES
§ Checkpoint to Persistent Storage (HDFS, S3)
§ Use MonitoredTrainingSession and Hooks
§ Use a Good Cluster Orchestrator (ie. Kubernetes, Mesos)
§ Understand Failure Modes and Recovery States
Stateless, Not Bad: Training Continues Stateful, Bad: Training Must Stop Dios Mio! Long Night Ahead…
125. AGENDA
Part 1: Optimize TensorFlow Training
§ GPUs and TensorFlow
§ Train, Inspect, and Debug TensorFlow Models
§ TensorFlow Distributed Cluster Model Training
§ Optimize Training with JIT XLA Compiler
126. XLA FRAMEWORK
§ XLA: “Accelerated Linear Algebra”
§ Reduce Reliance on Custom Operators
§ Intermediate Representation used by Hardware Vendors
§ Improve Portability
§ Increase Execution Speed
§ Decrease Memory Usage
§ Decrease Mobile Footprint
Helps TensorFlow Be Flexible AND Performant!!
127. XLA HIGH LEVEL OPTIMIZER (HLO)
§ HLO: “High Level Optimizer”
§ Compiler Intermediate Representation (IR)
§ Independent of source and target language
§ XLA Step 1 Emits Target-Independent HLO
§ XLA Step 2 Emits Target-Dependent LLVM
§ LLVM Emits Native Code Specific to Target
§ Supports x86-64, ARM64 (CPU), and NVPTX (GPU)
128. XLA IS DESIGNED FOR RE-USE
§ Pluggable Backends
§ HLO “Toolkit”
§ Call BLAS or cuDNN
§ Use LLVM or BYO Low-Level-Optimizer
135. XLA PERFORMANCE OPTIMIZATIONS
§ JIT Training
§ MNIST: 30% Speed Up
§ Inception: 20% Speed Up
§ Basic LSTM: 80% Speed Up
§ Translation Model BNMT: 20% Speed Up
§ AOT Inference (Next Section)
§ LSTM Model Size: 1 MB => 10 KB
136. JIT COMPILER
§ JIT: “Just-In-Time” Compiler
§ Built on XLA Framework
§ Reduce Memory Movement – Especially with GPUs
§ Reduce Overhead of Multiple Function Calls
§ Similar to Spark Operator Fusing in Spark 2.0
§ Unroll Loops, Fuse Operators, Fold Constants, …
§ Scopes: session, device, with jit_scope():
138. VISUALIZING JIT COMPILER IN ACTION
Before JIT After JIT
Google Web Tracing Framework:
http://google.github.io/tracing-framework/
from tensorflow.python.client import timeline
trace =
timeline.Timeline(step_stats=run_metadata.step_stats)
with open('timeline.json', 'w') as trace_file:
trace_file.write(
trace.generate_chrome_trace_format(show_memory=True))
run_options = tf.RunOptions(trace_level=tf.RunOptions.SOFTWARE_TRACE)
run_metadata = tf.RunMetadata()
sess.run(options=run_options,
run_metadata=run_metadata)
140. XLA COMPILATION SUMMARY
§ Generates Code and Libraries for Your Computation
§ Packages Libraries Needed for Your
§ Eliminates Dispatch Overhead of Operations
§ Fuses Operations to Avoid Memory Round Trip
§ Analyzes Buffers to Reuse Memory
§ Updates Memory In-Place
§ Unrolls Loops with Your Data Dimensions (ie.Batch Size)
§ Vectorizes Operations Specific to Your Data Dimensions
141. AGENDA
Part 0: Introductions and Setup
Part 1: Optimize TensorFlow Training
Part 2: Optimize TensorFlow Serving
Part 3: Advanced Model Serving + Traffic Routing
143. AGENDA
Part 2: Optimize TensorFlow Serving
§ AOT XLA Compiler and Graph Transform Tool
§ Key Components of TensorFlow Serving
§ Deploy Optimized TensorFlow Model
§ Optimize TensorFlow Serving Runtime
144. AOT COMPILER
§ Standalone, Ahead-Of-Time (AOT) Compiler
§ Built on XLA framework
§ tfcompile
§ Creates executable with minimal TensorFlow Runtime needed
§ Includes only dependencies needed by subgraph computation
§ Creates functions with feeds (inputs) and fetches (outputs)
§ Packaged as cc_libary header and object files to link into your app
§ Commonly used for mobile device inference graph
§ Currently, only CPU x86-64 and ARM are supported - no GPU
145. GRAPH TRANSFORM TOOL (GTT)
§ Post-Training Optimization to Prepare for Inference
§ Remove Training-only Ops (checkpoint, drop out, logs)
§ Remove Unreachable Nodes between Given feed -> fetch
§ Fuse Adjacent Operators to Improve Memory Bandwidth
§ Fold Final Batch Norm mean and variance into Variables
§ Round Weights/Variables to improve compression (ie. 70%)
§ Quantize (FP32 -> INT8) to Speed Up Math Operations
148. AFTER STRIPPING UNUSED NODES
§ Optimizations
§ strip_unused_nodes
§ Results
§ Graph much simpler
§ File size much smaller
149. AFTER REMOVING UNUSED NODES
§ Optimizations
§ strip_unused_nodes
§ remove_nodes
§ Results
§ Pesky nodes removed
§ File size a bit smaller
150. AFTER FOLDING CONSTANTS
§ Optimizations
§ strip_unused_nodes
§ remove_nodes
§ fold_constants
§ Results
§ Placeholders (feeds) -> Variables*
(*Why Variables and not Constants?)
151. AFTER FOLDING BATCH NORMS
§ Optimizations
§ strip_unused_nodes
§ remove_nodes
§ fold_constants
§ fold_batch_norms
§ Results
§ Graph remains the same
§ File size approximately the same
152. AFTER QUANTIZING WEIGHTS
§ Optimizations
§ strip_unused_nodes
§ remove_nodes
§ fold_constants
§ fold_batch_norms
§ quantize_weights
§ Results
§ Graph is same, file size is smaller, compute is faster
153. WEIGHT (VARIABLE) QUANTIZATION
§ FP16 or INT8: Smaller & Computationally Faster than FP32
§ Easy to “Linearly Quantize” (Re-Encode) FP32 -> INT8
Easy Breezy!
154. BENEFITS OF 32-BIT TO 8-BIT QUANTIZE
§ First Class Hardware and CUDA Support
§ One 32-Bit GPU Core: 4-Way Dot Product of 8-Bit Ints
§ GPU Compute Capability (CC) >= 6.1 Only
155. ACTIVATION QUANTIZATION
§ Activations Not Known Ahead of Time
§ Depends on input, not easy to quantize
§ Requires Additional Calibration Step
§ Use representative, diverse validation dataset
§ ~1000 samples, ~10 minutes,, cheap hardware
§ Run 32-Bit Inference with Calibration Data
§ Collect histogram of activation values at each layer
§ Generate many quantized distributions at diff saturation thresholds
§ Choose Saturation Threshold That Minimizes Accuracy Loss
156. CHOOSING SATURATION THRESHOLD
§ Trade-off Between Range & Precision
§ INT8 Should Encode Same Information As Original FP32
§ Minimize Loss of Information Across Encoding/Distributions
§ Use KL_Divergence(32bit_dist, 8bit_dist)
§ Compares 2 distributions
§ Similar to Cross-Entropy
157. SATURATE TO MINIMIZE ACCURACY LOSS
§ Helps Preserve Accuracy After Activation Quantization
§ Goal: Find Threshold (T) That Minimizes Accuracy Loss
No Saturation Saturation
158. AUTO-CALIBRATE: PIPELINEAI + TENSOR-RT
Pre-Requisites
§ 32-Bit Trained Model (TensorFlow, Caffe)
§ Small Calibration Dataset (Validation)
PipelineAI + TensorRT Optimizations
§ Run 32-Bit Inference on Calibration Dataset
§ Collect Required Statistics
§ Use KL_Divergence to Determine Saturation Thresholds
§ Perform 32-Bit Float -> 8-Bit Int Quantization
§ Generate Calibration Table and INT8 Execution Engine
159. 32-BIT TO 8-BIT QUANTIZATION RESULTS
Accuracy of INT8 Models Comparable to FP32
163. AGENDA
Part 2: Optimize TensorFlow Serving
§ AOT XLA Compiler and Graph Transform Tool
§ Key Components of TensorFlow Serving
§ Deploy Optimized TensorFlow Model
§ Optimize TensorFlow Serving Runtime
164. MODEL SERVING TERMINOLOGY
§ Inference
§ Only Forward Propagation through Network
§ Predict, Classify, Regress, …
§ Bundle
§ GraphDef, Variables, Metadata, …
§ Assets
§ ie. Map of ClassificationID -> String
§ {9283: “penguin”, 9284: “bridge”}
§ Version
§ Every Model Has a Version Number (Integer)
§ Version Policy
§ ie. Serve Only Latest (Highest), Serve Both Latest and Previous, …
165. TENSORFLOW SERVING FEATURES
§ Supports Auto-Scaling
§ Custom Loaders beyond File-based
§ Tune for Low-latency or High-throughput
§ Serve Diff Models/Versions in Same Process
§ Customize Models Types beyond HashMap and TensorFlow
§ Customize Version Policies for A/B and Bandit Tests
§ Support Request Draining for Graceful Model Updates
§ Enable Request Batching for Diff Use Cases and HW
§ Supports Optimized Transport with GRPC and Protocol Buffers
167. PREDICTION SERVICE
§ Predict (Original, Generic)
§ Input: List of Tensor
§ Output: List of Tensor
§ Classify
§ Input: List of tf.Example (key, value) pairs
§ Output: List of (class_label: String, score: float)
§ Regress
§ Input: List of tf.Example (key, value) pairs
§ Output: List of (label: String, score: float)
169. MULTI-HEADED INFERENCE
§ Inputs Pass Through Model One Time
§ Model Returns Multiple Predictions:
1. Human-readable prediction (ie. “penguin”, “church”,…)
2. Final layer of scores (float vector)
§ Final Layer of floats Pass to the Next Model in Ensemble
§ Optimizes Bandwidth, CPU/GPU, Latency, Memory
§ Enables Complex Model Composing and Ensembling
170. BUILD YOUR OWN MODEL SERVER
§ Adapt GRPC(Google) <-> HTTP (REST of the World)
§ Perform Batch Inference vs. Request/Response
§ Handle Requests Asynchronously
§ Support Mobile, Embedded Inference
§ Customize Request Batching
§ Add Circuit Breakers, Fallbacks
§ Control Latency Requirements
§ Reduce Number of Moving Parts
#include
“tensorflow_serving/model_servers/server_core.h”
class MyTensorFlowModelServer {
ServerCore::Options options;
// set options (model name, path, etc)
std::unique_ptr<ServerCore> core;
TF_CHECK_OK(
ServerCore::Create(std::move(options), &core)
);
}
Compile and Link with
libtensorflow.so
171. RUNTIME OPTION: NVIDIA TENSOR-RT
§ Post-Training Model Optimizations
§ Specific to Nvidia GPU
§ Similar to TF Graph Transform Tool
§ GPU-Optimized Prediction Runtime
§ Alternative to TensorFlow Serving
§ PipelineAI Supports TensorRT!
172. AGENDA
Part 2: Optimize TensorFlow Serving
§ AOT XLA Compiler and Graph Transform Tool
§ Key Components of TensorFlow Serving
§ Deploy Optimized TensorFlow Model
§ Optimize TensorFlow Serving Runtime
173. AGENDA
Part 2: Optimize TensorFlow Serving
§ AOT XLA Compiler and Graph Transform Tool
§ Key Components of TensorFlow Serving
§ Deploy Optimized TensorFlow Model
§ Optimize TensorFlow Serving Runtime
174. REQUEST BATCH TUNING
§ max_batch_size
§ Enables throughput/latency tradeoff
§ Bounded by RAM
§ batch_timeout_micros
§ Defines batch time window, latency upper-bound
§ Bounded by RAM
§ num_batch_threads
§ Defines parallelism
§ Bounded by CPU cores
§ max_enqueued_batches
§ Defines queue upper bound, throttling
§ Bounded by RAM
Reaching either threshold
will trigger a batch
Separate, Non-Batched Requests
Combined, Batched Requests
175. ADVANCED BATCHING & SERVING TIPS
§ Batch Just the GPU/TPU Portions of the Computation Graph
§ Batch Arbitrary Sub-Graphs using Batch / Unbatch Graph Ops
§ Distribute Large Models Into Shards Across TensorFlow Model Servers
§ Batch RNNs Used for Sequential and Time-Series Data
§ Find Best Batching Strategy For Your Data Through Experimentation
§ BasicBatchScheduler: Homogeneous requests (ie Regress or Classify)
§ SharedBatchScheduler: Mixed requests, multi-step, ensemble predict
§ StreamingBatchScheduler: Mixed CPU/GPU/IO-bound Workloads
§ Serve Only One (1) Model Inside One (1) TensorFlow Serving Process
§ Much Easier to Debug, Tune, Scale, and Manage Models in Production.
177. AGENDA
Part 0: Introductions and Setup
Part 1: Optimize TensorFlow Training
Part 2: Optimize TensorFlow Serving
Part 3: Advanced Model Serving + Traffic Routing
178. AGENDA
Part 3: Advanced Model Serving + Traffic
Routing
§ Kubernetes Ingress, Egress, Networking
§ Istio and Envoy Architecture
§ Intelligent Traffic Routing and Scaling
§ Metrics, Chaos Monkey, Production Readiness
179. KUBERNETES PRIORITY SCHEDULING
Workloads can …
§ access the entire cluster up
to the autoscaler max size
§ trigger autoscaling until
higher-priority workload
§ “fill the cracks” of resource
usage of higher-priority work
(i.e., wait to run until resources are feed
180. KUBERNETES INGRESS
§ Single Service
§ Can also use Service (LoadBalancer or NodePort)
§ Fan Out & Name-Based Virtual Hosting
§ Route Traffic Using Path or Host Header
§ Reduces # of load balancers needed
§ 404 Implemented as default backend
§ Federation / Hybrid-Cloud
§ Creates Ingress objects in every cluster
§ Monitors health and capacity of pods within each cluster
§ Routes clients to appropriate backend anywhere in federation
apiVersion: extensions/v1beta1
kind: Ingress
metadata:
name: gateway-fanout
annotations:
kubernetes.io/ingress.class: istio
spec:
rules:
- host: foo.bar.com
http:
paths:
- path: /foo
backend:
serviceName: s1
servicePort: 80
- path: /bar
backend:
serviceName: s2
servicePort: 80
Fan Out (Path)
apiVersion: extensions/v1beta1
kind: Ingress
metadata:
name: gateway-virtualhost
annotations:
kubernetes.io/ingress.class: istio
spec:
rules:
- host: foo.bar.com
http:
paths:
backend:
serviceName: s1
servicePort: 80
- host: bar.foo.com
http:
paths:
backend:
serviceName: s2
servicePort: 80
Virtual Hosting
181. KUBERNETES INGRESS CONTROLLER
§ Ingress Controller Types
§ Google Cloud: kubernetes.io/ingress.class: gce
§ Nginx: kubernetes.io/ingress.class: nginx
§ Istio: kubernetes.io/ingress.class: istio
§ Must Start Ingress Controller Manually
§ Just deploying Ingress is not enough
§ Not started by kube-controller-manager
§ Start Istio Ingress Controller
kubectl apply -f
$ISTIO_INSTALL_PATH/install/kubernetes/istio.yaml
193. ISTIO AUTO-SCALING
§ Traffic Routing and Auto-Scaling Occur Independently
§ Istio Continues to Obey Traffic Splits After Auto-Scaling
§ Auto-Scaling May Occur In Response to New Traffic Route
194. A/B & BANDIT MODEL TESTING
§ Perform Live Experiments in Production
§ Compare Existing Model A with Model B, Model C
§ Safe Split-Canary Deployment
§ Pro Tip: Keep Ingress Simple – Use Route Rules Instead!
apiVersion: config.istio.io/v1alpha2
kind: RouteRule
metadata:
name: predict-mnist-20-5-75
spec:
destination:
name: predict-mnist
precedence: 2 # Greater than global deny-all
route:
- labels:
version: A
weight: 20 # 20% still routes to model A
- labels:
version: B # 5% routes to new model B
weight: 5
- labels:
version: C # 75% routes to new model C
weight: 75
apiVersion: config.istio.io/v1alpha2
kind: RouteRule
metadata:
name: predict-mnist-1-2-97
spec:
destination:
name: predict-mnist
precedence: 2 # Greater than global deny-all
route:
- labels:
version: A
weight: 1 # 1% routes to model A
- labels:
version: B # 2% routes to new model B
weight: 2
- labels:
version: C # 97% routes to new model C
weight: 97
apiVersion: config.istio.io/v1alpha2
kind: RouteRule
metadata:
name: predict-mnist-97-2-1
spec:
destination:
name: predict-mnist
precedence: 2 # Greater than global deny-all
route:
- labels:
version: A
weight: 97 # 97% still routes to model A
- labels:
version: B # 2% routes to new model B
weight: 2
- labels:
version: C # 1% routes to new model C
weight: 1
195. AGENDA
Part 3: Advanced Model Serving + Traffic
Routing
§ Kubernetes Ingress, Egress, Networking
§ Istio and Envoy Architecture
§ Intelligent Traffic Routing and Scaling
§ Metrics, Chaos Monkey, Production Readiness
198. SPECIAL THANKS TO CHRISTIAN POSTA
§ http://blog.christianposta.com/istio-workshop
199. AGENDA
Part 0: Introductions and Setup
Part 1: Optimize TensorFlow Training
Part 2: Optimize TensorFlow Serving
Part 3: Advanced Model Serving + Traffic Routing
202. THANK YOU!!
§ Please Star this GitHub Repo!
§ All slides, code, notebooks, and Docker images here:
https://github.com/PipelineAI/pipeline
Contact Me
chris@pipeline.ai
@cfregly