Initial commit

Author: djosix

.gitignore

@@ -0,0 +1,6 @@
+.vscode
+*.pyc
+.DS_Store
+*.mid
+*.midi
+*.sess

README.md

@@ -0,0 +1,71 @@
+# Performance RNN - PyTorch
+
+PyTorch implementation of Performance RNN, inspired by
+[https://magenta.tensorflow.org/performance-rnn](https://magenta.tensorflow.org/performance-rnn).
+
+This model is not implemented in the official way!
+
+
+## Directory Structure
+
+```
+.
+├── dataset/
+│   ├── midi/
+│   │   ├── dataset1/
+│   │   │   └── *.mid
+│   │   └── dataset2/
+│   │       └── *.mid
+│   ├── processed/
+│   │   └── dataset1/
+│   │       └── *.data (generated with preprocess.py)
+│   └── scripts/
+│       └── *.sh (dataset download scripts)
+├── generated/
+│   └── *.mid (generated with generate.py)
+└── runs/ (tensorboard logdir)
+```
+
+
+## Getting Started
+
+- Download datasets
+    ```
+    cd dataset/
+    bash scripts/NAME_scraper.sh midi/NAME
+    ```
+
+- Preprocessing
+    ```shell
+    # Will preprocess all MIDI files under dataset/midi/NAME
+    python3 preprocess.py dataset/midi/NAME dataset/processed/NAME
+    ```
+
+- Training
+    ```shell
+    # Train on .data files in dataset/processed/MYDATA,
+    # and save to myModel.sess every 10s.
+    python3 train.py -s myModel.sess -d dataset/processed/MYDATA -i 10
+    ```
+
+- Generating
+    ```shell
+    py generate.py \
+        myModel.sess \  # load trained model from myModel.sess
+        generated/ \  # save to generated/
+        10 \            # generate 10 event sequences
+        2000 \          # generate 2000 event steps
+        0.9 \           # 90% sampling with argmax and 10% multinomial
+        '1,0,1,0,1,1,0,1,0,1,0,1' \ # pitch histogram ([12] or [0])
+        3               # note density (0-5)
+    ```
+
+# Requirements
+
+```
+pretty_midi
+numpy
+pytorch
+tensorboardX
+progress
+```

config.py

@@ -0,0 +1,23 @@
+import torch
+from sequence import EventSeq, ControlSeq
+
+#pylint: disable=E1101
+
+device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
+
+model = {
+    'init_dim': 32,
+    'event_dim': EventSeq.dim(),
+    'control_dim': ControlSeq.dim(),
+    'hidden_dim': 512,
+    'gru_layers': 3,
+    'gru_dropout': 0.3,
+}
+
+train = {
+    'learning_rate': 0.001,
+    'batch_size': 64,
+    'window_size': 200,
+    'stride_size': 10,
+    'use_transposition': False,
+}

data.py

@@ -0,0 +1,48 @@
+import os
+import torch
+import itertools
+import numpy as np
+import random
+
+import config
+import utils
+from sequence import EventSeq, ControlSeq
+
+# pylint: disable=E1101
+# pylint: disable=W0101
+
+class Dataset:
+    def __init__(self, root):
+        paths = utils.find_files_by_extensions(root, ['.data'])
+        self.samples = []
+        self.seqlens = []
+        for path in paths:
+            for eventseq, controlseq in torch.load(path):
+                controlseq = ControlSeq.recover_compressed_array(controlseq)
+                self.samples.append((eventseq, controlseq))
+        for eventseq, controlseq in self.samples:
+            assert len(eventseq) == len(controlseq)
+            self.seqlens.append(len(eventseq))
+    
+    def batches(self, batch_size, window_size, stride_size):
+        indeces = [(i, range(j, j + window_size))
+                   for i, seqlen in enumerate(self.seqlens)
+                   for j in range(0, seqlen - window_size, stride_size)]
+        while True:
+            eventseq_batch = []
+            controlseq_batch = []
+            n = 0
+            for ii in np.random.permutation(len(indeces)):
+                i, r = indeces[ii]
+                eventseq, controlseq = self.samples[i]
+                eventseq = eventseq[r.start:r.stop]
+                controlseq = controlseq[r.start:r.stop]
+                eventseq_batch.append(eventseq)
+                controlseq_batch.append(controlseq)
+                n += 1
+                if n == batch_size:
+                    yield (np.stack(eventseq_batch, axis=1),
+                           np.stack(controlseq_batch, axis=1))
+                    eventseq_batch.clear()
+                    controlseq_batch.clear()
+                    n = 0

dataset/midi/.keep

@@ -0,0 +1,16 @@
+#!/bin/bash
+# Scraper for Classical Piano Midi Page
+[ ! "$1" ] && echo 'Error: please specify output dir' && exit
+dir=$1
+base=http://www.piano-midi.de
+pages=$(curl -s --max-time 5 $base/midi_files.htm \
+    | grep '<tr class="midi"><td class="midi"><a href="' \
+    | egrep '[^"]+\.htm' -o)
+echo Pages: $pages
+mkdir -p $dir
+for page in $pages; do
+    midis=$(curl -s --max-time 5 $base/$page | egrep '[^"]+format0\.mid' -o)
+    for midi in $midis; do
+        echo "http://www.piano-midi.de/$midi"
+    done | tee /dev/stderr | wget -P $dir -i -
+done

dataset/processed/.keep

@@ -0,0 +1,18 @@
+#!/bin/bash
+# Scraper for Yamaha e-Piano Competition dataset
+[ ! "$1" ] && echo 'Error: please specify output dir' && exit
+dir=$1
+pages='http://www.piano-e-competition.com/ecompetition/midi_2002.asp
+http://www.piano-e-competition.com/ecompetition/midi_2004.asp
+http://www.piano-e-competition.com/ecompetition/midi_2006.asp
+http://www.piano-e-competition.com/ecompetition/midi_2008.asp
+http://www.piano-e-competition.com/ecompetition/midi_2009.asp
+http://www.piano-e-competition.com/ecompetition/midi_20011.asp
+'
+mkdir -p $dir
+for page in $pages; do
+    for midi in $(curl -s $page | egrep -i '[^"]+\.mid' -o | sed 's/^\/*/\//g'); do
+        echo "http://www.piano-e-competition.com$midi"
+    done
+done | wget -P $dir -i -
+rm -f $dir/*.{1,2,3,4,5}

dataset/scripts/.keep

@@ -0,0 +1,15 @@
+#!/bin/bash
+[ ! "$1" ] && echo 'Error: please specify output dir' && exit
+dir=$1
+for url in $(curl -s https://thwiki.cc/%E5%88%86%E7%B1%BB:%E5%AE%98%E6%96%B9MIDI \
+    | egrep -o '[^"]+?\.mid' \
+    | egrep '^/' \
+    | sed 's/^/https:\/\/thwiki.cc/g' \
+    | uniq);
+do url=$(curl -s "$url" \
+    | egrep -o '[^"]+?\.mid' \
+    | egrep '^/' \
+    | grep -v '%' \
+    | sed 's/^/https:/g');
+echo $url | tee /dev/stderr
+done | uniq | wget -P $dir -i -