def __data_generation(self, list_IDs_temp):
'Generates data containing batch_size samples' # X : (n_samples, *dim, n_channels)
# Initialization
X = np.empty((self.batch_size, *self.dim, self.n_channels), dtype = float)
y = np.empty((self.batch_size), dtype = int)
# Generate data
for i, ID in enumerate(list_IDs_temp):
data = extend_ts(self.h5file[ID]['ecgdata'][:, 0], self.sequence_length)
data = np.reshape(data, (1, len(data)))
if self.augment:
# dropout bursts
data = zero_filter(data, threshold = 2, depth = 10)
# random resampling
data = random_resample(data)
# Generate spectrogram
data_spectrogram = spectrogram(data, nperseg = self.nperseg, noverlap = self.noverlap)[2]
# Normalize
data_transformed = norm_float(data_spectrogram, self.data_mean, self.data_std)
X[i,] = np.expand_dims(data_transformed, axis = 3)
# Assuming that the dataset names are unique (only 1 per label)
y[i] = self.labels[ID]
return X, keras.utils.to_categorical(y, num_classes=self.n_classes)
def __data_generation(self, list_IDs_temp):
'Generates data containing batch_size samples' # X : (n_samples, *dim, n_channels)
# Initialization
X = np.empty((self.batch_size, *self.dim, self.n_channels), dtype=float)
y = np.empty((self.batch_size), dtype=int)
# Generate data
for i, ID in enumerate(list_IDs_temp):
data = extend_ts(self.h5file[ID]['ecgdata'][:, 0],
self.sequence_length)
data = np.reshape(data, (1, len(data)))
if self.augment:
# dropout bursts
data = zero_filter(data, threshold=2, depth=10)
# random resampling
data = random_resample(data)
# Generate spectrogram
data_spectrogram = spectrogram(data,
nperseg=self.nperseg,
noverlap=self.noverlap)[2]
# Normalize spectrogram
#data_transformed = norm_float(data_spectrogram, self.data_mean, self.data_std)
data_norm = (data_spectrogram -
np.mean(data_spectrogram)) / np.std(data_spectrogram)
X[i, ] = np.expand_dims(data_norm, axis=3)
# Assuming that the dataset names are unique (only 1 per label)
y[i] = self.labels[ID]
return X, keras.utils.to_categorical(y, num_classes=self.n_classes)
# Parameters needed for the batch generator
# Maximum sequence length
max_length = sequence_length_max
# Output dimensions
sequence_length = max_length
spectrogram_nperseg = 64 # Spectrogram window
spectrogram_noverlap = 32 # Spectrogram overlap
n_classes = len(label_df.label.unique())
batch_size = 32
# calculate image dimensions
data = fetch_h5data(h5file, [0], sequence_length)
_, _, Sxx = spectrogram(data,
nperseg=spectrogram_nperseg,
noverlap=spectrogram_noverlap)
dim = Sxx[0].shape
print('Maximum sequence length:', max_length)
params = {
'batch_size': batch_size,
'dim': dim,
'nperseg': spectrogram_nperseg,
'noverlap': spectrogram_noverlap,
'n_channels': 1,
'sequence_length': sequence_length,
'n_classes': n_classes,
'shuffle': True
}
from scipy import signal
f1, PSD = signal.periodogram(ts_extended, fs, 'flattop', scaling='density')
fig, ax1 = plt.subplots(figsize=(15, 3))
ax1.plot(f1, PSD, 'b')
ax1.set(xlabel='Frequency [Hz]',
xlim=[0, time[-1]],
xticks=np.arange(0, time[-1] + 5, 10))
ax1.set(ylabel='PSD')
plt.title('Power spectral density (PSD)', fontsize=15)
fig.savefig('physionet_ECG_PSD.png', bbox_inches='tight', dpi=150)
plt.show()
from physionet_processing import spectrogram
# Convert ECG into spectrograms without and with log transform
Sx = spectrogram(np.expand_dims(ts_extended, axis=0), log_spectrogram=False)[2]
Sx_log = spectrogram(np.expand_dims(ts_extended, axis=0),
log_spectrogram=True)[2]
# Get the frequency and time axes
f, t, _ = spectrogram(np.expand_dims(ts_extended, axis=0),
log_spectrogram=False)
# Plot the spectrograms as images
im_list = [Sx[0], Sx_log[0]]
im_title = [
'Spectrogram without log transform', 'Spectrogram with log transform'
]
fig, ax_list = plt.subplots(1, 2, figsize=(15, 3))
for i, ax in enumerate(ax_list):
# Generowanie danych testujacych model - klasa DataGenerator
###########################################################
# Maksymalna dlugosc przebiegu
max_length = sequence_length
# Parametry spektrogramu
sequence_length = sequence_length_max
spectrogram_nperseg = 64
spectrogram_noverlap = 32
n_classes = len(label_df.label.unique())
batch_size = 32
# Wymiary obrazu okreslone za pomoca wymiarow spektrogramu
Sx_log = spectrogram(np.expand_dims(ts, axis = 0),
nperseg = spectrogram_nperseg,
noverlap = spectrogram_noverlap,
log_spectrogram = True)[2]
dim = Sx_log[0].shape
# Inicjalizacja parametrow klasy DataGenerator
params = {'batch_size': batch_size,
'dim': dim,
'nperseg': spectrogram_nperseg,
'noverlap': spectrogram_noverlap,
'n_channels': 1,
'sequence_length': sequence_length,
'n_classes': n_classes,
'shuffle': True}
f1, PSD = signal.periodogram(ts, fs, 'flattop', scaling='density')
fig, ax2 = plt.subplots(figsize=(15, 3))
ax2.plot(f1, PSD, 'b')
#ax2.set(xlabel = 'Frequency [Hz]', xlim = [0, time[-1]], xticks = np.arange(0, time[-1]+5, 10))
#ax2.set(ylabel = 'PSD')
plt.title('Power spectral density (PSD)', fontsize=15)
fig.savefig('physionet_ECG_PSD.png', bbox_inches='tight', dpi=150)
########################################
#### 3. SPECTROGRAM ####################
########################################
from physionet_processing import spectrogram
# Convert ECG into spectrograms without and with log transform
Sx = spectrogram(np.expand_dims(ts, axis=0), log_spectrogram=False)[2]
Sx_log = spectrogram(np.expand_dims(ts, axis=0), log_spectrogram=True)[2]
# Get the frequency and time axes
f, t, _ = spectrogram(np.expand_dims(ts, axis=0), log_spectrogram=False)
# Plot the spectrograms as images
im_list = [Sx[0], Sx_log[0]]
im_title = [
'Spectrogram without log transform', 'Spectrogram with log transform'
]
fig, ax_list = plt.subplots(1, 2, figsize=(15, 3))
for i, ax in enumerate(ax_list):
ax.imshow(np.transpose(im_list[i]), aspect='auto', cmap='jet')