import matplotlib.pyplot as plt
import numpy as np
from deep_bottleneck.plotter.base import BasePlotter
[docs]def load(run, dataset):
return SignalToNoiseRatioPlotter(run, dataset)
# TODO think about whether plotting snr ratio averaged over multiple runs does make sense
[docs]class SignalToNoiseRatioPlotter(BasePlotter):
plotname = 'snr'
def __init__(self, run, dataset):
self.dataset = dataset
self.run = run
[docs] def plot(self, measures_summary):
activations_summary = measures_summary['activations_summary']
num_layers = len(activations_summary["0"]['weights_norm']) # get number of layers indirectly via number of values
epochs = []
means = []
stds = []
wnorms = []
epochs_in_activation_summary = [int(epoch) for epoch in activations_summary]
epochs_in_activation_summary = np.asarray(sorted(epochs_in_activation_summary))
for epoch in epochs_in_activation_summary:
epochs.append(epoch)
epoch_values = activations_summary[str(epoch)]
wnorms.append(epoch_values['weights_norm'])
means.append(epoch_values['gradmean'])
stds.append(epoch_values['gradstd'])
wnorms, means, stds = map(np.array, [wnorms, means, stds])
plot_layers = range(num_layers)
fig = plt.figure(figsize=(12, 5))
for lndx, layerid in enumerate(plot_layers):
ax = fig.add_subplot(1, len(plot_layers), lndx + 1)
ax.plot(epochs, means[:, layerid], 'b', label='Mean')
ax.plot(epochs, stds[:, layerid], 'orange', label='Std')
ax.plot(epochs, means[:, layerid] / stds[:, layerid], 'red', label='SNR')
ax.plot(epochs, wnorms[:, layerid], 'g', label='||W||')
ax.set_title(f'Layer {layerid}')
ax.set_xlabel('Epoch')
ax.set_xscale('log', nonposx='clip')
ax.set_yscale('log', nonposy='clip')
handles, labels = ax.get_legend_handles_labels()
fig.legend(handles, labels, loc='center left', bbox_to_anchor=(1.0, 0.5))
fig.tight_layout()
return fig