def plot_history_to_image_file(pickle_path):
with open(pickle_path, 'rb') as input:
trainLoss = pickle.load(input)
validLoss = pickle.load(input)
trainAcc = pickle.load(input)
validAcc = pickle.load(input)
x = range(len(validAcc))
y = validAcc
z = np.polyfit(x, y, 3)
p = np.poly1d(z)
fig = plt.figure(1)
plt.subplot(211)
axes = plt.gca()
axes.set_ylim([0, 5])
plt.ylabel("Loss")
plt.plot(trainLoss, 'o-', label="train")
plt.plot(validLoss, 'o-', label="valid")
plt.legend(loc="upper_left")
plt.subplot(212)
axes = plt.gca()
axes.set_ylim([0, 1])
plt.ylabel("Accuracy")
plt.xlabel("Epoch")
plt.plot(trainAcc, 'o-', label="train")
plt.plot(validAcc, 'o-', label="valid")
# plt.plot(p(x), 'o-', label="trend")
plt.legend(loc="upper_left")
basename = utils.get_basename_without_ext(pickle_path)
fig.savefig(basename + ".png")
plt.clf()
plt.close()
def signal_and_noise_spectrogram_from_wave_file(filepath):
(fs, wave) = utils.read_wave_file(filepath)
spectrogram = sp.wave_to_sample_spectrogram(wave, fs)
signal_wave, noise_wave = pp.preprocess_wave(wave, fs)
spectrogram_signal = sp.wave_to_sample_spectrogram(signal_wave, fs)
spectrogram_noise = sp.wave_to_sample_spectrogram(noise_wave, fs)
fig = plt.figure(1)
cmap = plt.cm.get_cmap('jet')
gs = gridspec.GridSpec(2, 2)
# whole spectrogram
ax1 = fig.add_subplot(gs[0, :])
ax1.pcolormesh(spectrogram, cmap=cmap)
ax1.set_title("Sound")
ax2 = fig.add_subplot(gs[1, 0])
ax2.pcolormesh(spectrogram_signal, cmap=cmap)
ax2.set_title("Signal")
ax3 = fig.add_subplot(gs[1, 1])
ax3.pcolormesh(spectrogram_noise, cmap=cmap)
ax3.set_title("Noise")
gs.update(wspace=0.5, hspace=0.5)
basename = utils.get_basename_without_ext(filepath)
fig.savefig(basename + "_noise_signal.png")
fig.clf()
plt.close(fig)
def noise_augmentation_from_dirs(noise_dir, class_dir):
sig_paths = glob.glob(os.path.join(class_dir, "*.wav"))
sig_path = np.random.choice(sig_paths, 1, replace=False)[0]
(fs, sig) = utils.read_wave_file(sig_path)
aug_sig = da.noise_augmentation(sig, noise_dir)
spectrogram_sig = sp.wave_to_sample_spectrogram(sig, fs)
spectrogram_aug_sig = sp.wave_to_sample_spectrogram(aug_sig, fs)
fig = plt.figure(1)
cmap = plt.cm.get_cmap('jet')
gs = gridspec.GridSpec(2, 1)
# whole spectrogram
ax1 = fig.add_subplot(gs[0, 0])
ax1.pcolormesh(spectrogram_sig, cmap=cmap)
ax1.set_title("Original Signal")
ax2 = fig.add_subplot(gs[1, 0])
ax2.pcolormesh(spectrogram_aug_sig, cmap=cmap)
ax2.set_title("Noise Augmented signal")
gs.update(wspace=0.5, hspace=0.5)
basename = utils.get_basename_without_ext(sig_path)
fig.savefig(basename + "_noise_augmentation.png")
fig.clf()
plt.close(fig)
def sprengel_binary_mask_from_wave_file(filepath):
fs, x = utils.read_wave_file(filepath)
Sxx = sp.wave_to_amplitude_spectrogram(x, fs)
Sxx_log = sp.wave_to_log_amplitude_spectrogram(x, fs)
# plot spectrogram
fig = plt.figure(1)
subplot_image(Sxx_log, 411, "Spectrogram")
Sxx = pp.normalize(Sxx)
binary_image = pp.median_clipping(Sxx, 3.0)
subplot_image(binary_image + 0, 412, "Median Clipping")
binary_image = morphology.binary_erosion(binary_image,
selem=np.ones((4, 4)))
subplot_image(binary_image + 0, 413, "Erosion")
binary_image = morphology.binary_dilation(binary_image,
selem=np.ones((4, 4)))
subplot_image(binary_image + 0, 414, "Dilation")
mask = np.array([np.max(col) for col in binary_image.T])
mask = morphology.binary_dilation(mask, np.ones(4))
mask = morphology.binary_dilation(mask, np.ones(4))
# plot_vector(mask, "Mask")
fig.set_size_inches(10, 12)
plt.tight_layout()
fig.savefig(utils.get_basename_without_ext(filepath) + "_binary_mask.png",
dpi=100)
def preprocess_sound_file(filename, class_dir, noise_dir,
segment_size_seconds):
""" Preprocess sound file. Loads sound file from filename, downsampels,
extracts signal/noise parts from sound file, splits the signal/noise parts
into equally length segments of size segment size seconds.
# Arguments
filename : the sound file to preprocess
class_dir : the directory to save the extracted signal segments in
noise_dir : the directory to save the extracted noise segments in
segment_size_seconds : the size of each segment in seconds
# Returns
nothing, simply saves the preprocessed sound segments
"""
samplerate, wave = utils.read_wave_file(filename)
signal_wave, noise_wave = preprocess_wave(wave, samplerate)
basename = utils.get_basename_without_ext(filename)
if signal_wave.shape[0] > 0:
signal_segments = split_into_segments(signal_wave, samplerate,
segment_size_seconds)
save_segments_to_file(class_dir, signal_segments, basename, samplerate)
if noise_wave.shape[0] > 0:
noise_segments = split_into_segments(noise_wave, samplerate,
segment_size_seconds)
save_segments_to_file(noise_dir, noise_segments, basename, samplerate)
def preprocess_sound_file(filename, class_dir, noise_dir,
segment_size_seconds):
""" Preprocess sound file. Loads sound file from filename, downsampels,
extracts signal/noise parts from sound file, splits the signal/noise parts
into equally length segments of size segment size seconds.
# Arguments
filename : the sound file to preprocess
class_dir : the directory to save the extracted signal segments in
noise_dir : the directory to save the extracted noise segments in
segment_size_seconds : the size of each segment in seconds
# Returns
nothing, simply saves the preprocessed sound segments
"""
samplerate, wave = utils.read_wave_file_not_normalized(filename)
print(wave)
if len(wave) == 0:
print("An empty sound file..")
wave = np.zeros(samplerate * segment_size_seconds, dtype=np.int16)
wave = wave.astype(np.float)
signal_wave, noise_wave = preprocess_wave(wave, samplerate)
if signal_wave.shape[0] == 0:
signal_wave = np.zeros(samplerate * segment_size_seconds,
dtype=np.int16)
basename = utils.get_basename_without_ext(filename)
# print(filename)
# print(class_dir)
# print(noise_dir)
if signal_wave.shape[0] > 0:
signal_segments = split_into_segments(signal_wave, samplerate,
segment_size_seconds)
save_segments_to_file(class_dir, signal_segments, basename, samplerate)
if noise_wave.shape[0] > 0:
noise_segments = split_into_segments(noise_wave, samplerate,
segment_size_seconds)
save_segments_to_file(noise_dir, noise_segments, basename, samplerate)
def same_class_augmentation_from_dir(class_dir):
sig_paths = glob.glob(os.path.join(class_dir, "*.wav"))
sig_path = np.random.choice(sig_paths, 1, replace=False)[0]
(fs, sig) = utils.read_wave_file(sig_path)
aug_sig_path = np.random.choice(sig_paths, 1, replace=False)[0]
(fs, aug_sig) = utils.read_wave_file(aug_sig_path)
alpha = np.random.rand()
combined_sig = (1.0 - alpha) * sig + alpha * aug_sig
spectrogram_sig = sp.wave_to_sample_spectrogram(sig, fs)
spectrogram_aug_sig = sp.wave_to_sample_spectrogram(aug_sig, fs)
spectrogram_combined_sig = sp.wave_to_sample_spectrogram(combined_sig, fs)
fig = plt.figure(1)
cmap = plt.cm.get_cmap('jet')
gs = gridspec.GridSpec(3, 1)
# whole spectrogram
ax1 = fig.add_subplot(gs[0, 0])
ax1.pcolormesh(spectrogram_sig, cmap=cmap)
ax1.set_title("Signal 1")
ax2 = fig.add_subplot(gs[1, 0])
ax2.pcolormesh(spectrogram_aug_sig, cmap=cmap)
ax2.set_title("Signal 2")
ax3 = fig.add_subplot(gs[2, 0])
ax3.pcolormesh(spectrogram_combined_sig, cmap=cmap)
ax3.set_title("Augmented Signal (alpha=" + str(alpha) + ")")
gs.update(wspace=0.5, hspace=0.5)
basename = utils.get_basename_without_ext(sig_path)
fig.savefig(basename + "_same_class_augmentation.png")
fig.clf()
plt.close(fig)
def img_spectrogram_from_wave_file(filepath):
fs, x = utils.read_gzip_wave_file(filepath)
Sxx = sp.wave_to_log_amplitude_spectrogram(x, fs, 512, 128)[:256]
baseName = utils.get_basename_without_ext(filepath)
save_matrix_to_file(Sxx, "Log Amplitude Spectrogram",
baseName + "_AMP.png")
