A (non) technical introduction to TensorFlow

What is Tensorflow?

(in general)

General Intro to tf

• Machine Learning Framework
• Developed by Google Brain team
• OpenSource - create an open standard for ML
• Flexible - from research to production

General Intro to tf

• Covers all aspects from testing, training to production and serving of models
• Portable, platform independent
• Scaleable

What is Tensorflow?

(a little bit technical :) )

How is DeepLearning done? (nDim Matrices)

What is a Tensor?

• n-dimensional array or list specified by rank:
  
  • 0: scalar
  • 1: vector
  • 2: matrix
  • 3: 3-tensor (cube of numbers)
• Datatypes: float, double, int ... (specific)

How is DeepLearning done? (Forward Propagation)

What about the Flow?

• Modelling the computations as nodes in a directed graph
• Checking for dimensional fit - pushing tensors through the network

What about the Flow?

How is DeepLearning done? (Computations)

Computations?

• Computations are nodes on the graph
• calculations are hidden from 'description language'
• Operations for different architectures (CPUs, GPUs)
• Large set of kernel methods included

Now, what makes tf so special

• Using grpc (protobuf) for communication between nodes
• Makes parallel computations easy (GPUs, distributed)
• Scalability in mind
• Distributed calculations

What's included?

• Graph View
• TensorBoard includes:
  
  • Graph viewer
  • Log analyzer, see when numbers converge
  • ...
• Serving mode:
  • Train and export model
  • Load model and bring it up as a service

What is tf suited best for?

• All kind of Machine Learning applications (obviously)
• Distributed Calculations e.g. ...
  • Mandelbrot
  • Partial differential equations
• Massive parallelization (CUDA)

Practical Aspects

How and where can it be used?

• Docker*
• Local instance*

Which Programming Languages are involved

• C / C++ for the core
• Python bindings for data scientists and ML (Google)
• Counting on Open-Source community for other languages

Demo

Example Code

							
	import tensorflow as tf

	x = tf.constant(1.0, name='input')
	w = tf.Variable(0.8, name='weight')
	y = tf.mul(w, x, name='output')
	y_ = tf.constant(0.0, name='correct_value')
	loss = tf.pow(y - y_, 2, name='loss')
	train_step = tf.train.GradientDescentOptimizer(0.025).minimize(loss)

	for value in [x, w, y, y_, loss]:
	    tf.scalar_summary(value.op.name, value)

	summaries = tf.merge_all_summaries()
	sess = tf.Session()
	summary_writer = tf.train.SummaryWriter('log_simple_stats', sess.graph)

	sess.run(tf.initialize_all_variables())
	for i in range(100):
	    summary_writer.add_summary(sess.run(summaries), i)
	    sess.run(train_step)
		

Tensorboard