yepesasecas / tensorflex

TensorflEx

Contents

• How to run
• Documentation
• Pre-application Milestones
• Summer and post-application milestones

How to run

• You need to have the Tensorflow C API installed. Look here for details.
• Clone this repository and cd into it
• Run mix deps.get to install the dependencies
• Run mix compile to compile the code
• Open up iex using iex -S mix

Documentation

Tensorflow requires the creation of graphs to represent the operations in a network. It also requires the creation of tensors, which are just the basic multi-dimensional datatype that contain data arrays and matrices for computation purposes. Once the graph is created an operation also needs to be created, which contains information about the kind of graph we are going to be running. The graph is finally run in a session, in which input and output tensors are defined.

For the purposes of describing the current capabilities of TensorflEx, the same things hold true. We will be creating a graph that just does a constant operation of running a constant string tensor as input and passing this as an output to another tensor. Instead of using the tensor to hold large amounts of data, for this example it just holds a string containing Hello World!. After getting iex started, we first define a new graph:

iex(1)> graph = TensorflEx.new_graph

Next, we define the graph operation. Here "Const" is the type of operation and "test" is the name we have given it so that we can reference it when we want:

iex(2)> op_desc = TensorflEx.new_op(graph, "Const", "test")

Then we define our string constant tensor. Make sure you do not use double quotes ("") in describing the tensor input as it has to be a char list:

iex(3)> tensor = TensorflEx.string_constant('Hello World!')

Lastly, we create and simultaneously run the session for our graph and then print the output:

iex(4)> IO.puts TensorflEx.create_and_run_sess(graph, op_desc, tensor)

This will give us the value of the string constant tensor as output, that is Hello World! as output.

Pre-application Milestones

• Simple TF_Version Hello World test

  • The first commit refers to the rather simplistic hello world program for the C API and not in general for Tensorflow. This is present here. This Hello world test works!

  Code in TensorflEx:

    
    iex(1)> IO.puts "Hello World! Tensorflow #{TensorflEx.version}"
    Hello World! Tensorflow 1.4.1
    :ok
    

• Graph, Tensor & Session based proper TF Hello World test

  • The better hello world test is to be able to program a real "Hello World!" TF program. This would involve declaring a string as a Tensorflow constant and then run a session to print the Tensor in a graph. As Python is still the best client for TF, this would look something like this in Python:

  import tensorflow as tf
  
  with tf.Session() as sess:
      print(sess.run(tf.constant("Hello World!")))

  OUTPUT: "Hello World!"
  

  Code in TensorflEx:

      
      iex(1)> graph = TensorflEx.new_graph
      2018-02-03 21:06:07.923328: I tensorflow/core/platform/cpu_feature_guard.cc:137] Your CPU supports instructions that this TensorFlow binary was not compiled to use: SSE4.1 SSE4.2 AVX AVX2 FMA
      #Reference<0.85340593.4211212290.58261>
      
      iex(2)> op_desc = TensorflEx.new_op(graph, "Const", "test")
      #Reference<0.85340593.4211212290.58539>
      
      iex(3)> tensor = TensorflEx.string_constant('Hello World!')
      'Hello World!'
      
      iex(4)> IO.puts TensorflEx.create_and_run_sess(graph, op_desc, tensor)
      => [INFO] Loaded Graph correctly
      => [INFO] Loaded Operation Description correctly
      => [INFO] Session Run Complete
      Hello World!
      :ok

Summer and post-application milestones

• Will be added in time