Tensorflow + Go

Vadim Markovtsev, source{d}.

Tensorflow + Go

Vadim Markovtsev, source{d}

Intent

Overview of Tensorflow internals
Tensorflow Hub
Show how to run Tensorflow models in Go
MobileNetv2 101
Bonus: better Tensorflow builds

Tensorflow terminology

Tensor
Variable
Constant
Operation
Graph
Session
Device

Tensorflow binaries

CPU
GPU=CUDA
OpenCL
TPU^*
Custom

Default CPU builds lack SSE and AVX.

Building Tensorflow

Bazel - Python
Bazel - dynamic, C API
autotools - static, no GPU, no C API

~40 minutes on i7-7500U (2x2 threads) laptop.

Tensorflow binaries

```
_pywrap_tensorflow_internal.so
```
Python API, ops
```
libtensorflow
```
C API, ops, [core]
```
libtensorflow_framework
```
core
```
libtensorflow-core
```
core + CPU ops

Tensorflow binaries for Go

cgo! Installation.

```
libtensorflow
```
C API, ops, [core]
```
libtensorflow_framework
```
core

Static linkage is not supported in 2018.

Life of a deep learning model

Research, debug
Hyperparameter optimization
Export
Import
Execution

Tensorflow model format zoo

GraphDef = graph + [variable values*]
Checkpoint = variable values
Summary (aka Tensorboard) = GraphDef + key-value
MetaGraph = GraphDef + tags + i/o + Checkpoint
SavedModel = one or more MetaGraph-s
ModuleDef (aka Hub) = special SavedModel

* 2GB limit.

Formats supported by Go

GraphDef = graph + [variable values]
SavedModel = one or more MetaGraph-s incompatible with ModuleDef

Imagenet

14.2mm labelled photos from Flickr
1000 classes
Est. 2009, last rev. 2011
State-of-the-art: 86% accuracy
Deep learning hype ignition
Link

Action

Plan

Load MobileNetv2
Read an image
Classify it

Action

Plan

Load BiDiSentiment
Classify text

Action

Thank you

My contacts: