def kroghInterpolation():
sample_rate, sample = wavfile.read('songs/hakuna_matata.wav')
sample = sample[5000000:5000100]
sample_rateBAD, sampleBAD = wavfile.read(
'songs/bad_songs/not_good_song.wav')
sampleBAD = sampleBAD[5000000:5000100]
BadSample = sample.copy()
dz.theEvilMethod(BadSample, 0.5)
matches = recognize.cheat(sample, BadSample)
x, y = utils.tovalidxy(BadSample, matches)
f = KroghInterpolator(x, y)
utils.repair(BadSample, matches, f)
IwannaSee(sample, BadSample, sampleBAD)
#! /usr/bin/env python
from scipy.io import wavfile
from interpolate import NewtonInterpolator
import damage, recognize, utils
sample_rate, samples = wavfile.read('songs/hakuna_matata.wav')
samples = samples[5000000:5000100]
newsamples = samples.copy()
damage.zerofill(newsamples, 0.3)
matches = recognize.cheat(samples, newsamples)
x, y = utils.tovalidxy(newsamples, matches)
f = NewtonInterpolator(x, y)
utils.repair(newsamples, matches, f)
import matplotlib.pyplot as plt
plt.title('Newton interpolation')
plt.xlabel('Frame')
plt.ylabel('Amplitude')
plt.plot(samples, label='real')
plt.plot(newsamples, label='interpolated')
plt.legend(loc='best')
plt.show()
gt_test_set = ColoredDataset(test_set, classes=10, colors=[1, 1], std=0)
gt_test_loader = DataLoader(gt_test_set, batch_size=args.batch_size, shuffle=False, num_workers=args.num_workers, pin_memory=True)
# measure bias & generalization before resampling
color_fn = lambda x: x.view(x.size(0), x.size(1), -1).max(2)[0] # color of digit
color_dim = 3 # rgb
bias_0 = measure_bias(train_loader, test_loader, color_fn, color_dim)[0]
model = create_mnist_model(args.model)
gt_acc_0 = measure_generalization(train_loader, [test_loader, gt_test_loader], model)[1]
print('Color bias before resampling = {:.3f}'.format(bias_0 + 0))
print('Generalization accuracy before resampling = {:.2%}'.format(gt_acc_0))
# learn resampling weights
repair_dataset = IndexedDataset(ConcatDataset([colored_train_set, colored_test_set]))
train_loader = DataLoader(repair_dataset, batch_size=args.batch_size, shuffle=True, num_workers=args.num_workers, pin_memory=True)
w, q, cls_idx, cls_w, bias = repair(train_loader, color_fn, color_dim, args.epochs, args.lr, args.lr_w)
# plot histogram of reampling weights
sns.distplot(w.cpu(), bins=100, kde=False)
plt.xlabel('Resampling weights')
plt.ylabel('# Examples')
plt.show()
# perform resampling
print('Resampling strategy:', args.sampling)
keep_idx = get_keep_idx(w, cls_idx, mode=args.sampling)
keep_idx_train = keep_idx[keep_idx < len(colored_train_set)]
keep_idx_test = keep_idx[keep_idx >= len(colored_train_set)] - len(colored_train_set)
print('Keep examples: {}/{} ({:.2%})'.format(len(keep_idx), len(w), len(keep_idx) / len(w)))
train_loader = DataLoader(Subset(colored_train_set, keep_idx_train), batch_size=args.batch_size, shuffle=True, num_workers=args.num_workers, pin_memory=True)
test_loader = DataLoader(Subset(colored_test_set, keep_idx_test), batch_size=args.batch_size, shuffle=False, num_workers=args.num_workers, pin_memory=True)
samples = samples[5000000:5000100]
newsamples = samples.copy()
damage.noiseadd(newsamples, 0.7, 0.3)
matches = recognize.cheat(samples,
newsamples,
false_positives=0.04,
false_negatives=0.1)
x, y = utils.tovalidxy(newsamples, matches)
flinear = interp1d(x, y, fill_value='extrapolate')
fcubic = interp1d(x, y, kind='cubic', fill_value='extrapolate')
plt.subplot(211)
utils.repair(newsamples, matches, flinear)
plt.title('Linear')
plt.ylabel('Amplitude')
plt.plot(samples, label='real')
plt.plot(newsamples, label='interpolated')
plt.legend(loc='best')
plt.subplot(212)
utils.repair(newsamples, matches, fcubic)
plt.title('Cubic')
plt.xlabel('Frame')
plt.ylabel('Amplitude')
plt.plot(samples, label='real')