import numpy as np
import pickle
import matplotlib.pyplot as plt
import scfb
import scfbutils
################################################################################
if __name__ == "__main__":
chunks, sig_len_n = pickle.load(open("chunks.pkl", "rb"))
freqs = np.logspace(np.log10(50.0), np.log10(4000.0), 100)
lims = np.sqrt(freqs[0:-1] * freqs[1:])
lims = np.concatenate([[0], lims, [20000]])
# limits = []
# for k in range(len(freqs)):
# limits.append((lims[k], lims[k+1]))
limits = [(lims[k], lims[k + 1]) for k in range(len(freqs))]
# freqs = np.linspace(50.0, 4000.0, 100)
bump_heights = [(1 / k)**0.9 for k in range(1, 7)]
bump_heights = np.array(bump_heights)
sigma = 0.03
mu = 1
# turn on limits if want to eliminate "butte" behavior
temp_array = scfb.TemplateArray(chunks, sig_len_n, freqs, sigma,
bump_heights, mu) #, limits=limits)
temp_array.adapt(verbose=True)
pickle.dump((temp_array.templates, freqs), open('template_data.pkl', 'wb'))
f0 = 205.0
fs = 44100
num_h = 6
mt_h = 0.5
len_t = 0.150
len_n = int(fs * len_t)
idcs = np.arange(len_n, dtype=np.int32)
chunks = []
for p in range(1, num_h + 1):
factor = 1.05 if p == mt_h else 1.0
print(f0 * p * factor)
chunks.append((idcs, np.ones(len_n) * f0 * p * factor))
f_vals = np.ones(1) * f0 * 0.95
ta = scfb.TemplateArray(chunks, len_n, f_vals, 8, sigma, mu, scale, beta)
ta.adapt(verbose=True)
print(ta.templates[0].f_vals[-1])
phi = ta.templates[0].f_vals
s = ta.templates[0].strengths
fig = plt.figure()
ax1 = fig.add_subplot(1, 2, 1)
ax2 = fig.add_subplot(1, 2, 2)
ax1.set_xlim(100.0, 1500.0)
ax1.set_ylim(-.5, 1.25)
ax2.set_xlim(100.0, 1500.0)
ax2.set_ylim(-.5, 1.25)
ax1.set_ylabel("Magnitude", fontsize=16)
## process through SCFB
peri = scfb.SCFB(peri_f_lo, peri_f_hi, num_peri_units, f_s)
peri.process_signal(in_sig, verbose=True)
## process through templates
chunks = peri.chunks # probably pickle this
freqs = np.logspace(np.log10(temp_f_lo), np.log10(temp_f_hi), num_templates)
lims = np.sqrt(freqs[0:-1] * freqs[1:])
lims = np.concatenate([[0], lims, [20000]])
limits = []
for k in range(len(freqs)):
limits.append((lims[k], lims[k + 1]))
temp_array = scfb.TemplateArray(chunks,
sig_len_n,
freqs,
temp_num_h,
sigma,
mu,
limits=limits)
temp_array.adapt(verbose=True)
## process through WTAN
templates = temp_array.templates # probably pickle this
t = np.arange(len(templates[0].strengths)) * (1. / 44100)
strengths = [t.strengths for t in templates]
# strengths = np.ascontiguousarray(np.flipud(np.stack(strengths, axis=0)))
strengths = np.ascontiguousarray(np.stack(strengths, axis=0))
k = np.ones(
(strengths.shape[0], strengths.shape[0])) * 5. # inhibition constant
# more elaborate inhibition schemes commented out below
# max_val = strengths.shape[0]*strengths.shape[1]
beta = 0.9
pitch_rel = np.zeros_like(mistunings)
for q in range(num_h):
h_mistuned = q + 1
for k in range(len(mistunings)):
# make the input chunks
chunks = []
for p in range(1, num_h + 1):
if p == h_mistuned:
chunks.append(
(chunk_idcs, np.ones(sig_len_n) * f0 * p * mistunings[k]))
else:
chunks.append((chunk_idcs, np.ones(sig_len_n) * f0 * p))
if k == 0:
temp_array = scfb.TemplateArray(chunks, sig_len_n, t_freqs, num_h,
sigma, mu, scale, beta)
else:
temp_array.new_data(chunks, sig_len_n)
temp_array.adapt()
strengths = [t.strengths[-1] for t in temp_array.templates]
pitch_idx = np.argmax(strengths)
pitch[k] = temp_array.templates[pitch_idx].f_vals[-1]
pitch_rel += pitch / f0
pitch_rel /= num_h
fig = plt.figure()
ax1 = fig.add_subplot(1, 2, 1)
ax2 = fig.add_subplot(1, 2, 2)
ax1.set_ylabel("Perceived pitch shift (percentage)", size=14)