Matrix multiplication using Go TF

The previous post, I described using the go TensorFlow bindings to add two integers. Now I describe using go TF to perform somewhat more interesting computation.

As before, create a new scope.

s := op.NewScope()

Create two placeholders. Notice that this time they are of type tf.Float.

ph1 := op.Placeholder(s.SubScope("ph1"), tf.Float)
ph2 := op.Placeholder(s.SubScope("ph2"), tf.Float)

Now we create several operations.

sumOP is the same as before. It take the two placeholders.
matMulOP performs matrix multiplication. It likewise take the two placeholders.
reLuOP is a rectifier. It takes the output of matMulOP.

sumOP := op.Add(s, ph1, ph2)
matMulOP := op.MatMul(s, ph1, ph2)
reLuOP := op.Relu(s, matMulOP)

Finalize the graph.

graph, err := s.Finalize()

Create the two input tensors. Note that they are now arrays of arrays of floats.

tensor1, err := tf.NewTensor([][]float32{[]float32{-4.6,5.91,6.40},[]float32{2.653,-3.21,4.316},[]float32{2.675,3,4}})
tensor2, err := tf.NewTensor([][]float32{[]float32{4,-5.61,6.8},[]float32{2.642,3.653,-4.4},[]float32{2.12,3,4}})

Input is the same as before.

input := map[tf.Output]*tf.Tensor{
  ph1: tensor1,
  ph2: tensor2,
}

However the output list now contains three elements. For their are now three operations whose output we wish to know.

output := []tf.Output{
  sumOP,
  matMulOP,
  reLuOP,
}

And as before, we create and run a session.

sess, err := tf.NewSession(graph, nil)
result, err := sess.Run(input, output, nil)

This time, sess.Run returns an array of three tensors, one for each of the three operations whose output we requested of it.

result[0] contains the sum of the two inputs.
result[1] contains the product of the two inputs.
result[2] contains the rectified product of the two inputs.

Note that matMulOP was only computed once, despite being requested twice. Lazy evaluation allow for performance optimizations.

Conclusion

Here I have described using go TF to perform more complex operations on multi dimensional arrays of floats. I also demonstrate the use of multiple output operations.

Blogroll

Matrix multiplication using Go TF

Conclusion