for row in range(rows): # for each eta
for col in range(cols): # for each learner
print("Plotting (%d, %d)" % (row, col))
ax = plt.subplot(rows, cols, row * cols + col + 1)
eta = etas[row]
# (n_trials x n_fdfs x n) matrix of errors
# print(type(errors[eta][0]))
# print(errors[eta][0].shape)
error = np.array(
[[errors[eta][itrial][ifdf, col] for ifdf in range(n_fdfs)]
for itrial in range(n_trials)])
error /= Yrms
# error[~np.isfinite(error)] = 2
error[~np.isfinite(error)] = 1e6
error = filt.filt(error, axis=-1)
# error = filt.filtfilt(error, axis=-1)
batch_inds = (n_per_show / 1000.) * np.arange(error.shape[-1])
sns.tsplot(
data=np.transpose(error, (0, 2, 1)),
time=batch_inds,
condition=f_df_labels,
# err_style='unit_traces',
legend=(row == 0 and col == 0))
ax.set_ylim([0.09, 1.2])
ax.set_yscale('log')
if row + 1 == rows:
ax.set_xlabel('thousands of examples')
eta_errors = np.array(eta_errors)
batch_inds = n_per_batch * np.arange(eta_errors.shape[-1])
# --- plot results (traces=learners)
plt.figure(figsize=(6.35, 6.0))
rows, cols = 2, 2
assert len(etas) <= rows * cols
for i, eta in enumerate(etas):
trial_errors = eta_errors[i]
filt = Alpha(30, default_dt=n_per_batch)
trial_errors[~np.isfinite(trial_errors)] = 1e6
trial_errors = trial_errors.clip(None, 1e6)
trial_errors = filt.filt(trial_errors / Yrms, axis=-1)
# trial_errors = filt.filtfilt(trial_errors / Yrms, axis=-1)
ax = plt.subplot(rows, cols, i + 1)
# ax = sns.tsplot(data=np.transpose(trial_errors, (0, 2, 1)),
# time=batch_inds, condition=learner_names)
sns.tsplot(data=np.transpose(trial_errors, (0, 2, 1)),
time=batch_inds,
condition=learner_names,
err_style='unit_traces',
legend=(i == 0))
# ax.set(yscale='log')
plt.ylim([1e-1, 1.2e0])
plt.xlabel('# of examples')
plt.ylabel('normalized RMS error')