facebookarchive/caffe2

MKL-DNN integration does not support backward pass

Open

#2,376 opened on Mar 22, 2018

View on GitHub
 (1 comment) (0 reactions) (0 assignees)Shell (8,424 stars) (1,954 forks)batch import
help wanted

Description

MKL-DNN integration (in caffe2/mkl) implements only forward pass operators, although MKL-DNN itself provides primitives for both forward and backward pass. Implementation of ConvGradient operator would let Caffe2 use MKL-DNN for training on CPU, and support some models which need backprop (e.g. deep dream).

Contributor guide

Tech stack
cpppythonshell
Domain
machine learningai
Issue type
feature
DifficultyEstimated implementation difficulty for a new contributor, from 1 for very small changes to 5 for expert-level work.
4
Estimated timeA rough time range for an experienced contributor to investigate, implement, test, and prepare a pull request.
3-5 days
Activity statusHow available the issue appears right now: fresh, active, stale, blocked, or waiting on maintainer input.
stale
ClarityHow clearly the issue explains the expected change, acceptance criteria, and next step.
clear
Prerequisites
Knowledge of MKL DNN primitiveC++ development experienceFamiliarity with Caffe2 operator registration
Newbie friendlinessA 1-100 score estimating how approachable this issue is for first-time contributors.
15
Research direction
Review existing forward pass implementation in caffe2/mkl directory to understand operator structure. Study MKL DNN documentation for backward pass primitives, particularly for convolution. Examine Caffe2's gradient operator patterns elsewhere (e.g., CPU or CUDA operators) to define ConvGradient. Propose a plan with specific files to modify and tests to add.