Welcome to mirror list, hosted at ThFree Co, Russian Federation.

github.com/marian-nmt/marian-regression-tests.git - Unnamed repository; edit this file 'description' to name the repository.
summaryrefslogtreecommitdiff
diff options
context:
space:
mode:
authorRoman Grundkiewicz <rgrundki@exseed.ed.ac.uk>2018-10-01 13:57:56 +0300
committerRoman Grundkiewicz <rgrundki@exseed.ed.ac.uk>2018-10-01 13:57:56 +0300
commitcce3a5daebf854a02ae9889b38455721d4d90f27 (patch)
treecac0322663fec40d396cba4d8fbff2a6a50cd69d /README.md
parentd5ea3bd476f4f850377447a5b275a62fe93d2233 (diff)
Update README
Diffstat (limited to 'README.md')
-rw-r--r--README.md29
1 files changed, 23 insertions, 6 deletions
diff --git a/README.md b/README.md
index 76a7da9..314cfb7 100644
--- a/README.md
+++ b/README.md
@@ -1,12 +1,11 @@
Marian regression tests
=======================
-**Marian** is a C++ GPU-specific parallel automatic differentiation library
-with operator overloading. It is the training framework used in the Marian
-toolkit.
+<b>Marian</b> is an efficient Neural Machine Translation framework written in
+pure C++ with minimal dependencies.
-This repository contains the regression test framework for the repo of
-`https://github.com/marian-nmt/marian-dev`.
+This repository contains the regression test framework for the main development
+repository: `https://github.com/marian-nmt/marian-dev`.
## Structure
@@ -20,7 +19,7 @@ Directories:
Each test consists of:
-* `test_*.sh` files
+* `test_*.sh` file
* `setup.sh` (optional)
* `teardown.sh` (optional)
@@ -57,6 +56,24 @@ Clean test artifacts:
make clean
+## Adding new tests
+
+Use templates provided in `tests/_template`.
+
+Please follow these recommendations:
+
+* For comparing outputs with numbers, please use float-friendly
+ `tools/diff-floats.py` instead of GNU `diff`
+* Make your tests deterministic using `--no-shuffle --seed 1111` or similar
+* Make training execution as short as possible, for instance, by reducing the
+ size of the network and the number of iterations
+* Do not run decoding or scoring on files longer than ca. 10-100 lines
+* If your tests require downloading and running a custom model, please keep it
+ as small as possible, and contact one of the main contributors to upload it
+ into our storage
+* Test one thing at a time
+
+
## Acknowledgements
The development of Marian received funding from the European Union's