def plot_random_fields():
width = 32
height = 32
x = np.arange(width)
y = np.arange(height)
x_grid, y_grid = np.meshgrid(x, y)
test_distances = np.arange(1, 30)
distances = distance_matrix(x_grid, y_grid)
length_scales = np.array([2, 16])
rand_gen = random_field_generator(x_grid, y_grid, length_scales, spatial_pattern="blended")
rand_fields = [next(rand_gen) for x in range(25)]
for rand_field in rand_fields:
print(rand_field.std(), rand_field.mean())
covariances = np.array([spatial_covariance(distances,
rand_field,
test_distances, 0.5) for rand_field in rand_fields])
plt.figure(figsize=(6, 4))
# plt.fill_between(test_distances, covariances.max(axis=0), covariances.min(axis=0), color='red', alpha=0.2)
for cov_inst in covariances:
plt.plot(test_distances, cov_inst, color='pink', marker='o', ls='-')
plt.plot(test_distances, covariances.mean(axis=0), 'ro-')
plt.plot(test_distances, exp_kernel(test_distances, length_scales[0]), 'bo-')
fig, axes = plt.subplots(5, 5, figsize=(9, 9))
plt.subplots_adjust(wspace=0,hspace=0)
axef = axes.ravel()
for a, ax in enumerate(axef):
ax.contourf(rand_fields[a], np.linspace(-4, 4, 20), cmap="RdBu_r", extend="both")
ax.tick_params(axis='both', which='both', bottom='off', top='off', right='off', left='off',
labelleft='off', labelbottom='off')
plt.show()
return
def generate_random_fields(set_size, data_width, length_scale_str):
length_scale_list = length_scale_str.split(";")
spatial_pattern = length_scale_list[0]
length_scales = [float(v) for v in length_scale_list[1:]]
x = np.arange(data_width)
y = np.arange(data_width)
x_grid, y_grid = np.meshgrid(x, y)
rand_gen = random_field_generator(x_grid, y_grid, length_scales, spatial_pattern=spatial_pattern)
data = np.stack([next(rand_gen) for i in range(set_size)], axis=0)
return data
def eval_full_gan_config(gan_index,
length_scale_str,
epochs,
gan_path,
data_width,
eval_distances,
tolerance,
batch_size=256):
"""
Evaluate GAN configurations with the full configuration
Args:
gan_index:
epochs:
gan_path:
data_width:
eval_distances:
tolerance:
Returns:
"""
gan_mean_cols = ["cov_mean_{0:02d}".format(x) for x in eval_distances]
gan_sd_cols = ["cov_sd_{0:02d}".format(x) for x in eval_distances]
gan_t_cols = ["cov_t_{0:02d}".format(x) for x in eval_distances]
gan_stat_cols = ["Index", "Epoch"
] + gan_mean_cols + gan_sd_cols + gan_t_cols + [
"mean_tscore", "max_tscore"
]
gan_stats = pd.DataFrame(np.ones(
(len(epochs), len(gan_stat_cols))) * np.nan,
columns=gan_stat_cols,
index=np.arange(len(epochs)))
x = np.arange(data_width)
y = np.arange(data_width)
x_g, y_g = np.meshgrid(x, y)
distances = distance_matrix(x_g, y_g)
length_scale_list = length_scale_str.split(";")
spatial_pattern = length_scale_list[0]
length_scales = [float(v) for v in length_scale_list[1:]]
corr = exp_kernel(distances, length_scales[0])
cho = cholesky(corr, lower=True)
cho_inv = np.linalg.inv(cho)
rand_covs = np.zeros((batch_size, eval_distances.size))
random_gen = random_field_generator(x_g, y_g, length_scales[0])
random_fields = np.stack([next(random_gen) for x in range(batch_size)],
axis=0)[:, :, :, 0]
random_noise = np.stack([
np.matmul(cho_inv, random_field.reshape(random_field.size, 1)).reshape(
random_fields.shape[1:]) for random_field in random_fields
],
axis=0)
noise_cov = np.zeros((batch_size, eval_distances.size))
for p, patch in enumerate(random_noise):
noise_cov[p] = spatial_covariance(distances, patch, eval_distances,
tolerance)
noise_mean = noise_cov.mean(axis=0)
noise_sd = noise_cov.std(axis=0)
for e, epoch in enumerate(epochs):
gan_patch_file = join(
gan_path, "gan_gen_patches_{0:04d}_epoch_{1:04d}.nc".format(
gan_index, epoch))
if exists(gan_patch_file):
gan_patch_ds = Dataset(gan_patch_file)
gan_patches = np.array(gan_patch_ds.variables["gen_patch"][:, :, :,
0])
gan_patch_ds.close()
for p, patch in enumerate(gan_patches):
rand_patch = np.matmul(cho_inv,
patch.reshape(patch.size,
1)).reshape(patch.shape)
rand_covs[p] = spatial_covariance(distances, rand_patch,
eval_distances, tolerance)
gan_stats.loc[e, "Index"] = gan_index
gan_stats.loc[e, "Epoch"] = epoch
gan_stats.loc[e, gan_mean_cols] = rand_covs.mean(axis=0)
gan_stats.loc[e, gan_sd_cols] = rand_covs.std(axis=0)
gan_stats.loc[e, gan_t_cols] = ttest_ind_from_stats(
gan_stats.loc[e, gan_mean_cols].values,
gan_stats.loc[e, gan_sd_cols].values,
np.ones(len(eval_distances) * batch_size),
noise_mean,
noise_sd,
np.ones(len(eval_distances) * batch_size),
equal_var=False)[0]
gan_stats.loc[e, "mean_tscore"] = np.abs(
gan_stats.loc[e, gan_t_cols]).mean()
gan_stats.loc[e, "max_tscore"] = np.abs(
gan_stats.loc[e, gan_t_cols]).max()
return gan_stats