Merge branch 'master' of https://github.com/andabi/deep-voice-conversion

docker/Dockerfile

@@ -0,0 +1,14 @@
+FROM tensorflow/tensorflow:1.8.0-gpu
+MAINTAINER Guang Yang <[email protected]>
+
+RUN apt-get update -y && apt-get install -y \
+  sox \
+  libsndfile1-dev \
+  git \
+  ffmpeg \
+ && apt-get clean \
+ && rm -rf /var/lib/apt/lists/*
+
+
+ADD pip-requirements.txt /pip-requirements.txt
+RUN pip install -r /pip-requirements.txt

docker/README.md

@@ -0,0 +1,20 @@
+# Training with Nvidia-Docker
+## Prerequisites
+Because of dependencies of the GPU (cuda etc), there are 3 steps if you're starting from scratch on a 16.04 machine
+- install cuda
+- install docker
+- install nvidia-docker
+
+## Quickstart
+first build the image
+```
+nvidia-docker build -t $USER/deep-voice-transfer-gpu:latest .
+```
+
+then run
+```
+nvidia-docker run -it --rm $USER/deep-voice-transfer-gpu:latest
+```
+
+## Got-chas
+- Notice the pip-requirement does not have tensorflow, but instead tensorflow-gpu. I didn't quite figure out how to force it to use GPU (even with the `-gpu 0` flag) so having tensorflow also installed meant that CPU is used so you "hang" at 0% when you train weights for net2

docker/pip-requirements.txt

@@ -0,0 +1,9 @@
+tensorflow-gpu >= 1.1
+numpy >= 1.11.1
+librosa == 0.5.1
+tensorpack == 0.8.0
+pyyaml
+soundfile
+pydub
+git+https://github.com/wookayin/tensorflow-plot.git@master
+tqdm