def test_dtw_symmetric(self):
x = mfcc(self.tidigits[17]['samples'])
y = mfcc(self.tidigits[3]['samples'])
dist = lambda x, y: np.linalg.norm(x - y, ord=2)
d1 = dtw(x, y, dist=dist, debug=False)
d2 = dtw(y, x, dist=dist, debug=False)
self.assertEqual(d2, d1)
def test_dtw(self):
x = mfcc(self.tidigits[12]['samples'])
y = mfcc(self.tidigits[22]['samples'])
dist = lambda x, y: np.linalg.norm(x - y, ord=2)
d1, cost1, acc_cost1, path1 = dtw_mod.dtw(x, y, dist=dist)
d2 = dtw(x, y, dist=dist, debug=False)
assert_almost_equal(d1, d2)
def test_dtw_zero(self):
x = mfcc(self.tidigits[7]['samples'])
dist = lambda x, y: np.linalg.norm(x - y, ord=2)
d1, cost1, acc_cost1, path1 = dtw_mod.dtw(x, x, dist=dist)
d2 = dtw(x, x, dist=dist, debug=False)
assert_almost_equal(d1, d2)
# assert_allclose(d2, 0.0)
self.assertEqual(d2, 0.0)
def q7_global_distance():
global_distances = np.zeros((44, 44))
all_mfcc = []
for i in range(44):
all_mfcc.append(mfcc(data_dict[i]['samples']))
for i in range(44):
for j in range(44):
if i == j:
continue
elif global_distances[j, i] != 0:
global_distances[i, j] = global_distances[j, i]
else:
global_d, _, acc_d, _ = dtw(all_mfcc[i], all_mfcc[j],
euclidean)
global_distances[i, j] = global_d
plot_p_color_mesh(global_distances, 'global distance matrix')
np.save('global distance.npy', global_distances)
def test_dtw_ideal(self):
x = np.array([2, 0, 1, 1, 2, 4, 2, 1, 2, 0]).reshape(-1, 1)
y = np.array([1, 1, 2, 4, 2, 1, 2, 0]).reshape(-1, 1)
dist = lambda x, y: np.linalg.norm(x - y, ord=2)
d1, cost1, acc_cost1, path1 = dtw_mod.dtw(x, y, dist=dist)
d2, cost2, acc_cost2, path2 = dtw(x, y, dist=dist, debug=True)
assert_almost_equal(d1, d2)
assert_allclose(cost1, cost2)
assert_allclose(acc_cost1, acc_cost2)
path_r_idx, path_c_idx = path1
self.assertEqual(len(path2), len(path_r_idx))
self.assertEqual(len(path2), len(path_c_idx))
for idx, (i, j) in enumerate(path2):
self.assertEqual(i, path_r_idx[idx])
self.assertEqual(j, path_c_idx[idx])
Calculate a global distance matrix and save the gdist file
"""
import numpy as np
import config
from lab1_proto import dtw
from lab1_proto import mfcc
from scipy.cluster.hierarchy import dendrogram, linkage
from matplotlib import pyplot as plt
if __name__ == "__main__":
data = np.load('data/lab1_data.npz')['data']
# calcualte the global distance matrix
ndata = len(data)
# global_dist matrix is symmetric
# the diagonal terms are zeros
global_dist = np.zeros((ndata, ndata))
cnt = 0
for i, j in zip(*np.triu_indices(ndata, k=1)):
feature_i = mfcc(data[i]['samples'])
feature_j = mfcc(data[j]['samples'])
d = dtw(feature_i, feature_j)
global_dist[i, j] = d
global_dist[j, i] = d
cnt += 1
if cnt % 100 == 0:
print("Calculated %d global distances" % cnt, flush=True)
np.save(config.gdist_npy_file, global_dist)
from sklearn.mixture import GaussianMixture
gmm = GaussianMixture(n_components=4).fit(data_mfcc)
labels = gmm.predict(data_mfcc)
from sklearn.manifold import TSNE
X_embedded = TSNE(n_components=2).fit_transform(data_mfcc)
plt.scatter(X_embedded[:, 0], X_embedded[:, 1], c=labels, s=40, cmap='viridis')
for i in [16, 17, 38, 39]:
samples = data[i]['samples']
s = mfcc(samples)
if (i == 0):
data_gmm = s
else:
data_gmm = np.append(data_mfcc, s, axis=0)
gmm = GaussianMixture(n_components=32).fit(data_gmm)
labels = gmm.predict(data_gmm)
sample_ = data.shape[0]
result = np.zeros([sample_, sample_])
for i in range(sample_):
mfcc1 = mfcc(data[i]['samples'])
for j in range(sample_):
mfcc2 = mfcc(data[j]['samples'])
result[i, j] = dtw(mfcc1, mfcc2, scipy.spatial.distance.euclidean)
plt.plot(result)
from scipy.cluster.hierarchy import dendrogram, linkage
Z = linkage(result, 'complete')
fig = plt.figure(figsize=(25, 10))
dn = dendrogram(Z)