Use relative links with .md extension (automatically converted to .ht…

…ml for github-pages) (awslabs#979)

docs/_config.yml

@@ -6,3 +6,7 @@ markdown: kramdown
 highlighter: rouge
 url: https://awslabs.github.io
 base_url: /sockeye
+relative-links:
+  enabled: true
+  include:
+    - tutorials

docs/index.md

@@ -1,7 +1,3 @@
----
-layout: default
----
-
 # Sockeye
 
 [![PyPI version](https://badge.fury.io/py/sockeye.svg)](https://badge.fury.io/py/sockeye)
@@ -19,7 +15,7 @@ For a quickstart guide to training a standard NMT model on any size of data, see
 
 If you are interested in collaborating or have any questions, please submit a pull request or [issue](https://github.com/awslabs/sockeye/issues/new).
 You can also send questions to *sockeye-dev-at-amazon-dot-com*.
-Developers may be interested in [our developer guidelines](development.html).
+Developers may be interested in [our developer guidelines](development.md).
 
 ## Citation
 

docs/inference.md

@@ -1,7 +1,3 @@
----
-layout: default
----
-
 # Translation
 
 Decoding (a.k.a. inference or translation) in sockeye is made available through the `sockeye.translate` module.
@@ -41,12 +37,12 @@ The PNG files will be written to files beginning with the prefix given by the `-
 
 ## Source factors
 
-If your [model was trained with source factors](training.html#source-factors), you will need to supply them at test-time, too.
+If your [model was trained with source factors](training.md#source-factors), you will need to supply them at test-time, too.
 Factors can be provided in three formats: (a) separate, token-parallel files (as in training), (b) direct annotations on words, or (c) in a JSON object.
 
 ### Parallel files
 
-You can also provide parallel files, [in the same style as training](training.html#source-factors).
+You can also provide parallel files, [in the same style as training](training.md#source-factors).
 Factor files are token-parallel to the source and are passed in to `sockeye.translate` via the `--input-factors` flag.
 (In this scenario, the source is another file, passed via `--input`).
 

docs/scoring.md

@@ -1,7 +1,3 @@
----
-layout: default
----
-
 # Scoring existing translations
 
 Sockeye provides a fast scoring module that permits the scoring of existing translations.

docs/training.md

@@ -1,7 +1,3 @@
----
-layout: default
----
-
 # Training
 
 ## Data preparation
@@ -110,7 +106,7 @@ In this case a drop in training throughput is expected.
 #### Multi-GPU training
 
 Training can be carried out on multiple GPUs. See the
-[WMT 2014 English-German tutorial](https://awslabs.github.io/sockeye/tutorials/wmt_large.html) for more information.
+[WMT 2014 English-German tutorial](tutorials/wmt_large.md) for more information.
 
 
 ### Checkpoint averaging
@@ -152,7 +148,7 @@ Since these embeddings concatenated to those of the word embeddings, the total s
 You can also sum the embeddings (`--source-factors-combine sum`).
 In this case, you do not need to specify `--source-factors-num-embed`, since they are automatically all set to the size of the word embeddings (`--num-embed`).
 
-You then also have to apply factors for the source side [at inference time](inference.html#source-factors).
+You then also have to apply factors for the source side [at inference time](inference.md#source-factors).
 
 ## Target factors
 

docs/tutorials.md

@@ -2,16 +2,16 @@
 
 ## Setup
 
-For installing Sockeye follow the [installation instructions](setup.html) to manually install Sockeye and all dependencies.
+For installing Sockeye follow the [installation instructions](setup.md) to manually install Sockeye and all dependencies.
 The tutorials below might have additional dependencies that will be mentioned at the beginning of each tutorial.
 
 ## Tutorials
 
 Below is the full list of tutorials we provide. We recommend going through them in order as they will gradually
 introduce different concepts and parameters used for training and translation.
 
-1. [Sequence copy task](tutorials/seqcopy.html)
-1. [WMT German to English news translation](tutorials/wmt.html)
-1. [Domain adaptation of NMT models](tutorials/adapt.html)
-1. [Large data: WMT English-German 2014](tutorials/wmt_large.html)
-1. [Multilingual Zero-shot Translation IWSLT 2017](tutorials/multilingual.html)
+1. [Sequence copy task](tutorials/seqcopy_tutorial.md)
+1. [WMT German to English news translation](tutorials/wmt.md)
+1. [Domain adaptation of NMT models](tutorials/adapt.md)
+1. [Large data: WMT English-German 2014](tutorials/wmt_large.md)
+1. [Multilingual Zero-shot Translation IWSLT 2017](tutorials/multilingual.md)

docs/tutorials/adapt.md

@@ -4,7 +4,7 @@ Although the quality of machine translation systems is nowadays remarkably good,
 These customizations may include preferring some word translation over others or adapting the style of the text, among others.
 In this tutorial, we show two methods on how to perform domain adaptation of a general translation system using Sockeye.
 
-We assume you already have a trained Sockeye model, for example the one trained from the [WMT tutorial tutorial](wmt.html).
+We assume you already have a trained Sockeye model, for example the one trained from the [WMT tutorial](wmt.md).
 We also assume that you have two training sets, one composed of general or out-of-domain (OOD) data, and one composed of in-domain (ID) data on which you want to adapt your system.
 Note that both datasets need to be pre-processed in the same way.
 

docs/tutorials/multilingual.md

@@ -3,9 +3,9 @@
 In this tutorial we will train a multilingual Sockeye model that can translate between several language pairs,
 including ones that we did not have training data for (this is called _zero-shot translation_).
 
-Please note: this tutorial assumes that you are familiar with the introductory tutorials on [copying
-sequences](https://awslabs.github.io/sockeye/tutorials/seqcopy.html)
-and [training a standard WMT model](https://awslabs.github.io/sockeye/tutorials/wmt.html).
+Please note: this tutorial assumes that you are familiar with the introductory tutorials on
+[copying sequences](seqcopy_tutorial.md)
+and [training a standard WMT model](wmt.md).
 
 ## Approach
 
@@ -42,7 +42,7 @@ virtualenv -p python3 sockeye3
 source sockeye3/bin/activate
 ```
 
-Then [install the correct version of Sockeye](https://awslabs.github.io/sockeye/setup.html).
+Then [install the correct version of Sockeye](../setup.md).
 We also install several libraries for preprocessing, monitoring and evaluation:
 
 ```bash

docs/tutorials/seqcopy.md → docs/tutorials/seqcopy_tutorial.md

@@ -6,7 +6,7 @@ The task is then to train a model that copies the sequence from the source to th
 This task is on the one hand difficult enough to be interesting and on the other and allows for quickly training a model.
 
 ## Setup
-For this tutorial we assume that you have successfully [installed](../setup.html) Sockeye.
+For this tutorial we assume that you have successfully [installed](../setup.md) Sockeye.
 We will be using scripts from the Sockeye repository, so you should either clone the repository or manually download the scripts.
 Just as a reminder: Everything is run using Python 3, so depending on your setup you may have to replace `python` with `python3` below.
 All of the commands below assume you are running on a CPU.

docs/tutorials/wmt.md

@@ -26,7 +26,7 @@ pip install tensorboard
 
 All of the commands below assume you're running on a CPU.
 If you have a GPU available you can simply remove `--use-cpu`.
-With multiple GPUs you can use `torchrun` to spawn multiple training processes (see [WMT 2014 English-German tutorial](https://awslabs.github.io/sockeye/tutorials/wmt_large.html)).
+With multiple GPUs you can use `torchrun` to spawn multiple training processes (see [WMT 2014 English-German tutorial](wmt_large.md)).
 
 ## Data