Adding methods to create constant tensor.

[arafatkatze]

Before we review this a few questions

• Are you cool with the dependency on narray for Complex data types, I used because ruby array does not have pack for complex data type ?
• I hope the tests for constant are working fine.

If this works fine for you,then I will add support for a few data types that I had left before in set_type method of tensor(and also for constant).

Another important thing to be noted is that tf.variable will also be made using advantage of constant op.

[chrhansen]

@Arafatk shouldn't we try and follow as much as possible the same interface (order of arguments) as the Python wrapper? I.e. tf.constant(value, dtype=None, shape=None, name='Const')

https://www.tensorflow.org/versions/r0.9/api_docs/python/constant_op.html#constant

Nice, short and sweet implementation!

[zverok]

Are you cool with the dependency on narray for Complex data types, I used because ruby array does not have pack for complex data type ?

As far as I can tell, SciRuby currently prefers NMatrix over NArray big time (and NArray is in the strange development state, it seems). Does NMatrix have functionality you need? Or maybe it is not so complex to implement by hands, sticking with just Ruby Array?

[zverok]

First, you don't need result variable at all: all branches just finish with calculation of result, so it will be returned in this case:

   case attribute_type
    when "type"
      Tensorflow::AttrValue.new(type: value) # just put the value number here
    when "tensor"
      tensor_element_type = value.type_num
      case value.type_num
      when Tensorflow::TF_DOUBLE
        Tensorflow::AttrValue.new(tensor: Tensorflow::TensorProto.new(dtype: value.type_num, tensor_shape: value.tensor_shape_proto, tensor_content: value.flatten.pack("d*")))
      when Tensorflow::TF_INT32
        Tensorflow::AttrValue.new(tensor: Tensorflow::TensorProto.new(dtype: value.type_num, tensor_shape: value.tensor_shape_proto, tensor_content: value.flatten.pack("l*")))
      when Tensorflow::TF_INT64
        Tensorflow::AttrValue.new(tensor: Tensorflow::TensorProto.new(dtype: value.type_num, tensor_shape: value.tensor_shape_proto, tensor_content: value.flatten.pack("q*")))
      when Tensorflow::TF_COMPLEX128
        tensor_narray = NArray.complex(value.flatten.length)
        (0..value.flatten.length - 1).each do |i|
          tensor_narray[i] = value.flatten[i]
        end
        Tensorflow::AttrValue.new(tensor: Tensorflow::TensorProto.new(dtype: value.type_num, tensor_shape: value.tensor_shape_proto, tensor_content: tensor_narray.to_s))
      end
    end

But it seems code could be simplified even more: there is exactly the same lines in all "tensor" branch:

Tensorflow::AttrValue.new(tensor: Tensorflow::TensorProto.new(dtype: value.type_num, tensor_shape: value.tensor_shape_proto, tensor_content: "the only difference is here"))

So, what do you think about this:

    case attribute_type
    when "type"
      Tensorflow::AttrValue.new(type: value) # just put the value number here
    when "tensor"
      tensor_element_type = value.type_num
      content =
        case value.type_num
        when Tensorflow::TF_DOUBLE
          value.flatten.pack("d*")
        when Tensorflow::TF_INT32
          value.flatten.pack("l*")
        when Tensorflow::TF_INT64
          value.flatten.pack("q*")
        when Tensorflow::TF_COMPLEX128
          tensor_narray = NArray.complex(value.flatten.length)
          (0..value.flatten.length - 1).each do |i|
            tensor_narray[i] = value.flatten[i]
          end
          tensor_narray.to_s
        end
      Tensorflow::AttrValue.new(
        tensor: Tensorflow::TensorProto.new(
          dtype: value.type_num,
          tensor_shape: value.tensor_shape_proto,
          tensor_content: content
        )
      )
    end

Now we can see that content = calculation could be easily extracted into separate method, making code to look somewhat like this:

   case attribute_type
    when "type"
      Tensorflow::AttrValue.new(type: value) # just put the value number here
    when "tensor"
      Tensorflow::AttrValue.new(
        tensor: Tensorflow::TensorProto.new(
          dtype: value.type_num,
          tensor_shape: value.tensor_shape_proto,
          tensor_content: value_to_tensor_content(value)
        )
      )
    end

And, finally, typically it is a good practice to have else branch in case (even if you are sure it is impossible to reach -- impossibility handler, raising an error, could save you a hours of debugging).

[chrhansen]

@zverok we had NMatrix in shortly, but removed it again (https://github.com/Arafatk/tensorflow.rb/pull/16) since it didn't install without manual intervention on macs, SciRuby/nmatrix#505 (comment). But actually, it looks like they've fixed this by now: SciRuby/nmatrix#530. @Arafatk do you wanna swap in NMatrix instead of NArray (provided specs still pass)?

[arafatkatze]

@chrhansen Thanks for the suggestion. I have used narray for the simple reason that it has to_s method for Complex arrays. We don't have that feature in Ruby Array and also its not available in NMAtrix. I have tried searching but I could not find any other popular gem that does this.