Skip to content

Latest commit

 

History

History
72 lines (50 loc) · 2.73 KB

submission-guideline.md

File metadata and controls

72 lines (50 loc) · 2.73 KB

General

The general idea of this tensor-id registry is to provide an identifier for a tensor encoding that is not associated with any specific framework or programming language. The tensor description should be complete enough to allow the interpretation of the tensor without additional reference. This register is meant to facilitate N.N. usage in a framework agnostic way. Internal details and layers of the N.N. are not necessarily in scope, but if you believe their is value in using this register to describe internal layer you are welcome to submit by accompanying your submission with and explanation of the value added. Tensor encoding submission should include at least one reference to the original paper from which this encoding emerge or derive, one link to model using this encoding and one link to a code, publically available, able to interpret the tensor encoding.

Tensor Id

The identifier of the tensor must be unique for a complete definition of a tensor.

Tensor Description File

  • Use markdown
  • At the time of writing this, the filename of the file used to describe the tensor is used as a storage of the tensor-id.
  • Filename is composed of {tensor-id}.md

Selecting a tensor-id

Anatomy of a tensor-id

  1. Only use alpha-numeric characters separated by -
  2. Name of the first architecture where this tensor encoding has emerged
  3. Append "-variant-1" term if this encoding is a derivative of the original tensor encoding. If a "variant-1" already exist, use the next variant index. Ex."variant-2". Append "-out" it the tensor is an output of the inference. Use "-in" it is an input of the inference. Use "-inout" if it's output and an input of the inference.
  4. If the tensor is part of a tensor-group, append a term that identifies the function of this tensor in the tensor-group.

Example:

Original N.N. Architecture: "ssd-mobilenet-v1-variant-1

Keywords

Sections

Classification

Classification of the tensor-encoding and tensor-groups.

tensor-group

Indicate if the content describe a tensor-group or a tensor.

layer-type

Indicate if the content describes input, output or inout tensor.

use-case

In which context this tensor-encoding is used.

part-of-tensor-groups

List tensor-groups using this tensor, if it is not a generic tensor.

Description

Provide a sufficient description to understand the information it contains concerning the analysis result and if it's associated with another tensor.

tensor-shape

Provide the tensor dimensions

tensor-datatype

Provide the datatype used by this tensor

Known Aliases

List known aliases of this tensor encoding

Encoding

Describe any pattern used to encode analysis result in the tensor. Provide a memory layout of the tensor encoding.