Initial commit

LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2020 OpenAI
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

README.md

@@ -0,0 +1,102 @@
+**Status:** Archive (code is provided as-is, no updates expected)
+
+# Jukebox
+Code for "Jukebox: A Generative Model for Music"
+
+# Install
+``` 
+# Required: Sampling
+conda create --name jukebox python=3.7.5
+conda activate jukebox
+conda install mpi4py=3.0.3
+conda install pytorch=1.4 torchvision=0.5 cudatoolkit=10.0 -c pytorch
+cd jukebox
+pip install -r requirements.txt
+pip install -e .
+
+# Required: Training
+conda install av=7.0.01 -c conda-forge 
+cd tensorboardX
+python setup.py install
+ 
+# Optional: Apex for faster training with fused_adam
+conda install pytorch=1.1 torchvision=0.3 cudatoolkit=10.0 -c pytorch
+cd apex
+pip install -v --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" ./
+```
+
+# Sampling
+To sample, run the following command. Model can be `5b`, `5b_lyrics`, `1b_lyrics`
+``` 
+python jukebox/sample.py --model=5b_lyrics --name=sample_5b --levels 3 --sample_length 1048576 --sr 44100 --n_samples 3 --hop_fraction 0.5,0.5,0.125
+```
+``` 
+python jukebox/sample.py --model=1b_lyrics --name=sample_1b --levels 3 --sample_length 786432 --sr 44100 --n_samples 16 --hop_fraction 0.5,0.5,0.125
+```
+
+Here sample_length is chosen so that we get a single n_ctx (8192 for 5b model, 6144 for 1b_model, VQ-VAE downsampling is 128X) of tokens at the top-level, however you can choose longer sample_lengths to generate longer samples. For example, for a 2X longer sample, run
+``` 
+python jukebox/sample.py --model=5b_lyrics --name=sample_5b_2x --levels 3 --sample_length 2097152 --sr 44100 --n_samples 12 --hop_fraction 0.5,0.5,0.125
+```
+The length of the generated raw audio is `sample_length/sr` seconds, where sr = 44100 KHz for our VQ-VAE.
+
+On a V100, to generate a single ctx at each level, it should take approximately 
+- 12 min for the 5b model top level (1 ctx = 24 sec of music)
+- 7 min for the 1b model top level (1 ctx = 18 sec of music)
+- 9 min for upsamplers (1 ctx = 6 sec and 1.5 sec respectively for middle and bottom level). 
+
+So, it takes about 3 hrs for 5b model to fully sample 24sec of music. Since this could be a long time, its recommended to use n_samples > 1 so you can generate multiple samples in parallel.   
+
+The samples decoded from each level are stored in `{name}/level_{level}`. You can also view the samples as an html with the aligned lyrics under `{name}/level_{level}/index.html`.
+
+# Training
+## VQVAE
+To train a small vqvae, run
+```
+mpiexec -n {ngpus} python jukebox/train.py --hps=small_vqvae --name=small_vqvae --sample_length 262144 --bs 4 --nworkers 4 --audio_files_dir {audio_files_dir} --labels False --train --aug_shift --aug_blend
+```
+Here, {audio_files_dir} is the directory in which you can put the audio files for your dataset. The above trains a two-level VQ-VAE with downs_t = (5,3), and strides_t = (2, 2) meaning we downsample the audio by 2^5 = 32X to get the first level of codes, and 2^8 = 256X to get the second level codes.  Checkpoints are stored in the `logs` folder. You can monitor the training by running Tensorboard
+```
+tensorboard --logdir logs
+```
+
+## Prior
+### Train prior or upsamplers
+Once the VQ-VAE is trained, we can restore it from its saved checkpoint and train priors on the learnt codes. 
+To train the top-level prior, we can run
+
+```
+mpiexec -n {ngpus} python jukebox/train.py --hps=small_vqvae,small_prior,all_fp16,cpu_ema --name=small_prior --sample_length 2097152 --bs 4 --nworkers 4 --audio_files_dir {audio_files_dir} --labels False --train --aug_shift --aug_blend --restore_vqvae logs/small_vqvae/checkpoint_latest.pth.tar --prior --levels 2 --level 1 --weight_decay 0.01 --save_iters 1000
+```
+
+To train the upsampler, we can run
+```
+mpiexec -n {ngpus} python jukebox/train.py --hps=small_vqvae,small_upsampler,all_fp16,cpu_ema --name=small_upsampler --sample_length 65536 --bs 4 --nworkers 4 --audio_files_dir {audio_files_dir} --labels False --train --aug_shift --aug_blend --restore_vqvae logs/small_vqvae/checkpoint_latest.pth.tar --prior --levels 2 --level 0 --weight_decay 0.01 --save_iters 1000
+```
+
+We chose sample_lengths above so that after the compression factors of the VQ-VAE (32x and 256x at levels 0 and 1), we get an n_ctx which matches that of the prior/upsamplers we're training (8192 for each level). 
+
+### Reuse pre-trained VQ-VAE and retrain top level prior on new dataset.
+Our pre-trained VQ-VAE can produce compressed codes for a wide variety of genres of music, and the pre-trained upsamplers can upsample them back to audio that sound very similar to the original audio.
+To re-use these for a new dataset of your choice, you can retrain just the top-level  
+
+To retrain top-level on a new dataset, run
+```
+mpiexec -n {ngpus} python jukebox/train.py --hps=vqvae,small_prior,all_fp16,cpu_ema --name pretrained_vqvae_small_prior --sample_length=1048576 --bs 4 --nworkers 4 --bs_sample 4 --aug_shift --aug_blend --audio_files_dir {audio_files_dir} --labels False --train --prior --levels 3 --level 2 --weight_decay 0.01 --save_iters 1000
+```
+
+You can then run sample.py with the top-level of our models replaced by your new model. To do so, add an entry 'my_model' in MODELs (in make_models.py) with the (vqvae hps, upsampler hps, top-level prior hps) of your new model, and run sample.py with `--model=my_model` . 
+
+
+# Citation
+
+Please cite using the following bibtex entry:
+
+```
+@article{dhariwal2020jukebox,
+  title={Jukebox: A Generative Model for Music},
+  author={Dhariwal, Prafulla and Jun, Heewoo and Payne, Christine and Kim, Jong Wook and Radford, Alec and Sutskever, Ilya},
+  journal={arXiv preprint arXiv:[TODO]},
+  year={2020}
+}
+```

apex/.gitignore

@@ -0,0 +1,5 @@
+apex.egg-info
+dist
+build
+docs/build
+*~

apex/LICENSE

@@ -0,0 +1,11 @@
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without modification, are permitted provided that the following conditions are met:
+
+1. Redistributions of source code must retain the above copyright notice, this list of conditions and the following disclaimer.
+
+2. Redistributions in binary form must reproduce the above copyright notice, this list of conditions and the following disclaimer in the documentation and/or other materials provided with the distribution.
+
+3. Neither the name of the copyright holder nor the names of its contributors may be used to endorse or promote products derived from this software without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.

apex/README.md

@@ -0,0 +1,99 @@
+# Introduction
+
+This repository holds NVIDIA-maintained utilities to streamline 
+mixed precision and distributed training in Pytorch. 
+Some of the code here will be included in upstream Pytorch eventually.
+The intention of Apex is to make up-to-date utilities available to 
+users as quickly as possible.
+
+## Full API Documentation: [https://nvidia.github.io/apex](https://nvidia.github.io/apex)
+
+# Contents
+
+## 1. Amp:  Automatic Mixed Precision
+
+`apex.amp` is a tool to enable mixed precision training by changing only 3 lines of your script.
+Users can easily experiment with different pure and mixed precision training modes by supplying
+different flags to `amp.initialize`.
+
+[Webinar introducing Amp](https://info.nvidia.com/webinar-mixed-precision-with-pytorch-reg-page.html)
+(The flag `cast_batchnorm` has been renamed to `keep_batchnorm_fp32`).
+
+[API Documentation](https://nvidia.github.io/apex/amp.html)
+
+[Comprehensive Imagenet example](https://github.com/NVIDIA/apex/tree/master/examples/imagenet)
+
+[DCGAN example coming soon...](https://github.com/NVIDIA/apex/tree/master/examples/dcgan)
+
+[Moving to the new Amp API](https://nvidia.github.io/apex/amp.html#transition-guide-for-old-api-users) (for users of the deprecated "Amp" and "FP16_Optimizer" APIs)
+
+## 2. Distributed Training
+
+`apex.parallel.DistributedDataParallel` is a module wrapper, similar to 
+`torch.nn.parallel.DistributedDataParallel`.  It enables convenient multiprocess distributed training,
+optimized for NVIDIA's NCCL communication library.
+
+[API Documentation](https://nvidia.github.io/apex/parallel.html)
+
+[Python Source](https://github.com/NVIDIA/apex/tree/master/apex/parallel)
+
+[Example/Walkthrough](https://github.com/NVIDIA/apex/tree/master/examples/simple/distributed)
+
+The [Imagenet example](https://github.com/NVIDIA/apex/tree/master/examples/imagenet)
+shows use of `apex.parallel.DistributedDataParallel` along with `apex.amp`.
+
+### Synchronized Batch Normalization
+
+`apex.parallel.SyncBatchNorm` extends `torch.nn.modules.batchnorm._BatchNorm` to
+support synchronized BN.
+It allreduces stats across processes during multiprocess (DistributedDataParallel) training.
+Synchronous BN has been used in cases where only a small
+local minibatch can fit on each GPU.
+Allreduced stats increase the effective batch size for the BN layer to the
+global batch size across all processes (which, technically, is the correct
+formulation).
+Synchronous BN has been observed to improve converged accuracy in some of our research models.
+
+# Requirements
+
+Python 3
+
+CUDA 9 or newer
+
+PyTorch 0.4 or newer.  The CUDA and C++ extensions require pytorch 1.0 or newer.
+
+We recommend the latest stable release, obtainable from
+[https://pytorch.org/](https://pytorch.org/).  We also test against the latest master branch, obtainable from [https://github.com/pytorch/pytorch](https://github.com/pytorch/pytorch).
+
+It's often convenient to use Apex in Docker containers.  Compatible options include:
+* [NVIDIA Pytorch containers from NGC](https://ngc.nvidia.com/catalog/containers/nvidia%2Fpytorch), which come with Apex preinstalled.  To use the latest Amp API, you may need to `pip uninstall apex` then reinstall Apex using the **Quick Start** commands below.
+* [official Pytorch -devel Dockerfiles](https://hub.docker.com/r/pytorch/pytorch/tags), e.g. `docker pull pytorch/pytorch:nightly-devel-cuda10.0-cudnn7`, in which you can install Apex using the **Quick Start** commands.
+
+See the [Docker example folder](https://github.com/NVIDIA/apex/tree/master/examples/docker) for details.
+
+# Quick Start
+
+### Linux
+
+For performance and full functionality, we recommend installing Apex with
+CUDA and C++ extensions via
+```
+$ git clone https://github.com/NVIDIA/apex
+$ cd apex
+$ pip install -v --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" .
+```
+
+Apex also supports a Python-only build (required with Pytorch 0.4) via
+```
+$ pip install -v --no-cache-dir .
+```
+A Python-only build omits:
+- Fused kernels required to use `apex.optimizers.FusedAdam`.
+- Fused kernels required to use `apex.normalization.FusedLayerNorm`.
+- Fused kernels that improve the performance and numerical stability of `apex.parallel.SyncBatchNorm`.
+- Fused kernels that improve the performance of `apex.parallel.DistributedDataParallel` and `apex.amp`.
+`DistributedDataParallel`, `amp`, and `SyncBatchNorm` will still be usable, but they may be slower.
+
+### Windows support
+Windows support is experimental, and Linux is recommended.  `pip install -v --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" .` may work if you were able to build Pytorch from source
+on your system.  `pip install -v --no-cache-dir .` (without CUDA/C++ extensions) is more likely to work.  If you installed Pytorch in a Conda environment, make sure to install Apex in that same environment.

apex/apex.patch

@@ -0,0 +1,42 @@
+diff --git a/csrc/fused_adam_cuda_kernel.cu b/csrc/fused_adam_cuda_kernel.cu
+index 34f7aa2..95581d1 100644
+--- a/csrc/fused_adam_cuda_kernel.cu
++++ b/csrc/fused_adam_cuda_kernel.cu
+@@ -19,8 +19,8 @@ typedef enum{
+
+ template <typename T, typename GRAD_T>
+ __global__ void adam_cuda_kernel(
+-        T* __restrict__ p,
+-        GRAD_T* __restrict__ p_copy, // For mixed precision training, pass NULL if not needed
++        GRAD_T* __restrict__ p,
++        T* __restrict__ p_copy, // For mixed precision training, pass NULL if not needed
+         T* __restrict__ m,
+         T* __restrict__ v,
+         const GRAD_T * __restrict__ g,
+@@ -50,7 +50,7 @@ __global__ void adam_cuda_kernel(
+                 else // Mode 1
+                     denom = sqrtf(v[j]) + eps;
+                 float update = (m[j]/denom) + (decay*p[j]);
+-                p[j] = p[j] - (step_size*update);
++                p[j] = (GRAD_T) (p[j] - (step_size*update));
+                 if (p_copy != NULL) p_copy[j] = (GRAD_T) p[j];
+         }
+ }
+@@ -93,14 +93,14 @@ void fused_adam_cuda(
+
+         if (g.scalar_type() == at::ScalarType::Half) {
+ //all other values should be fp32 for half gradients
+-            AT_ASSERTM(p.scalar_type() == at::ScalarType::Float, "expected parameter to be of float type");
++//            AT_ASSERTM(p.scalar_type() == at::ScalarType::Float, "expected parameter to be of float type");
+ //dispatch is done on the gradient type
+             using namespace at; // prevents "toString is undefined" errors
+             DISPATCH_FLOAT_AND_HALF(g.scalar_type(), 0, "adam_cuda_kernel", 
+                 using accscalar_t = at::acc_type<scalar_t_0, true>;
+                 adam_cuda_kernel<accscalar_t, scalar_t_0><<<blocks,threadsPerBlock, 0, stream>>>(
+-                        p.data<accscalar_t>(),
+-                        p_copy.numel() ? p_copy.data<scalar_t_0>() : NULL,
++                        p.data<scalar_t_0>(),
++                        NULL, //don't output p_copy for fp32, it's wasted write
+                         m.data<accscalar_t>(),
+                         v.data<accscalar_t>(),
+                         g.data<scalar_t_0>(),

apex/apex/RNN/README.md

@@ -0,0 +1 @@
+Under construction...