test_rnn.py

import tensorflow as tf
from src.basic_rnn import RNNMusic
from src.bidirectional_rnn import BidirectionalRNNMusic
from src.stacked_rnn import StackedRNNMusic
import os
import glob
import numpy as np
from src.midi import midi_to_matrix, matrix_to_midi
from collections import defaultdict
import pdb

input_len = 5
num_units = 32
lr = 0.003
num_epochs = 500
FOLDER = './data/overfit/'
saved_models_folder = './models/shitty_birnn/'

# Some stats about the folder
# stats = defaultdict(int)
# files = glob.glob(FOLDER + '*.mid')
# for file in files:
    # matrix = midi_to_matrix(file)
    # # matrix_to_midi(matrix, os.path.join(saved_models_folder, 'input.mid'))
    # notes = np.argmax(matrix, axis=1)
    # for note in notes:
        # stats[note] += 1
# keys = [key for key in stats]
# values = [stats[key] for key in keys]
# most_occurring_note = keys[values.index(max(values))]
# with open(os.path.join(saved_models_folder, 'train_stats.txt'),'w') as f:
    # f.write('Number of notes used to predict next note: ' + str(input_len) + '\n')
    # f.write('Hidden vector dimension: ' + str(num_units) + '\n')
    # f.write('Learning rate: ' + str(lr) + '\n')
    # f.write('Number of epochs: ' + str(num_epochs) + '\n')
    # f.write('Stats about the frequency of notes\n')
    # f.write('Most occurring note: ' + str(most_occurring_note) + '\n')
    # for key in stats:
        # f.write(str(key) + ': ' + str(stats[key]) + '\n')

# Train
# with tf.Graph().as_default():
    # rnn_music = StackedRNNMusic(input_len=input_len, num_units=num_units, lr=lr, num_epochs=num_epochs)
    # init = tf.global_variables_initializer()
    # saver = tf.train.Saver()
    # with tf.Session() as sess:
        # sess.run(init)
        # rnn_music.fit(sess, saver, FOLDER, saved_models_folder)

# Test
num = max([int(val.split('_')[2]) for val in glob.glob(saved_models_folder + 'epoch_*')])
saved_models = saved_models_folder + 'epoch_' + str(num) + '/'
with tf.Graph().as_default():
    rnn_music = BidirectionalRNNMusic(input_len=input_len, num_units=num_units, lr=lr, num_epochs=num)
    saver = tf.train.Saver()
    with tf.Session() as sess:
        saver.restore(sess, os.path.join(saved_models, 'epoch_' + str(num) + '.ckpt'))
        rnn_music.generate(sess, 100, saved_models_folder + 'test_output.mid', os.path.join(saved_models_folder, 'test_stats.txt'))