In [1]:

    

import keras
from musicnet.dataset import MusicNet
from keras.models import load_model

%matplotlib inline
import seaborn
from matplotlib import pyplot as plt
from sklearn.metrics import precision_recall_curve, average_precision_score

from complexnn.conv import ComplexConv1D
from complexnn.bn import ComplexBN, ComplexBatchNormalization
from complexnn.init import sqrt_init
from keras import initializers
from keras.models import model_from_yaml
setattr(initializers, "sqrt_init", sqrt_init)


    

    




Using Theano backend.

In [8]:

    

dataset = MusicNet('/Tmp/serdyuk/data/musicnet_11khz.npz', complex_=True, fourier=True)
dataset.load()
Xtest, Ytest = dataset.eval_set('test')


    

In [5]:

    

model = model_from_yaml(open("models/complex_model.yaml"),
                   custom_objects={'ComplexConv1D': ComplexConv1D, 
                                   'ComplexBN': ComplexBN, 
                                   'sqrt_init': sqrt_init,
                                   'ComplexBatchNormalization': ComplexBatchNormalization})


    

In [20]:

    

model2 = model_from_yaml(open("models/real_model.yaml"),
                   custom_objects={'ComplexConv1D': ComplexConv1D, 
                                   'ComplexBN': ComplexBN, 
                                   'sqrt_init': sqrt_init,
                                   'ComplexBatchNormalization': ComplexBatchNormalization})


    

In [6]:

    

model.load_weights("models/complex_model.hdf5")


    

In [22]:

    

model2.load_weights("models/real_model.hdf5")


    

In [11]:

    

pr = model.predict(Xtest)


    

In [23]:

    

pr2 = model2.predict(Xtest)


    

In [52]:

    

seaborn.set_style('white')
seaborn.set_context('poster')
plt.figure()
plt.subplot(211)
plt.imshow(pr[:200, 10:65].T)
plt.yticks(list(range(10, 65, 10)))
plt.ylabel('Note ID')

plt.subplot(212)
plt.imshow(Ytest[:200, 10:65].T)
plt.yticks(list(range(10, 65, 10)))
plt.xlabel('Frame')
plt.ylabel('Note ID')
plt.savefig("pred_gt.png")
plt.show()


    

In [40]:

    

precision1, recall1, _ = precision_recall_curve(Ytest.flatten(), pr.flatten())
precision2, recall2, _ = precision_recall_curve(Ytest.flatten(), pr2.flatten())
seaborn.set_style('whitegrid')
seaborn.set_context('poster')
plt.figure()
mod1, = plt.plot(recall1, precision1)
mod2, = plt.plot(recall2, precision2)
plt.xlabel('Recall')
plt.ylabel('Precision')
plt.legend([mod1, mod2], ["Complex", "Real"])
plt.savefig('precision_recall.png')