def visualize_vae(vae, skew_config, report,
resolution=20,
title="VAE Heatmap"):
xlim, ylim = vae.get_plot_ranges()
show_prob_heatmap(vae, xlim=xlim, ylim=ylim, resolution=resolution)
fig = plt.gcf()
prob_heatmap_img = vu.save_image(fig)
report.add_image(prob_heatmap_img, "Prob " + title)
show_weight_heatmap(
vae, skew_config, xlim=xlim, ylim=ylim, resolution=resolution,
)
fig = plt.gcf()
heatmap_img = vu.save_image(fig)
report.add_image(heatmap_img, "Weight " + title)
return prob_heatmap_img
def create_figure(
report: HTMLReport,
target_poses,
*list_of_vector_fields
):
num_vfs = len(list_of_vector_fields)
width = 7
height = 7 * num_vfs
for i, target_pos in enumerate(target_poses):
fig, axes = plt.subplots(
num_vfs, figsize=(width, height)
)
for j, vf in enumerate([vfs[i] for vfs in list_of_vector_fields]):
# `heatmaps` is now a list of heatmaps, such that
# heatmaps[k] = list_of_list_of_heatmaps[k][i]
min_pos = max(target_pos - WaterMaze.TARGET_RADIUS,
-WaterMaze.BOUNDARY_DIST)
max_pos = min(target_pos + WaterMaze.TARGET_RADIUS,
WaterMaze.BOUNDARY_DIST)
"""
Plot Estimated & Optimal QF
"""
ax = axes[j]
vu.plot_vector_field(fig, ax, vf)
ax.vlines([min_pos, max_pos], *ax.get_ylim())
ax.set_xlabel("Position")
ax.set_ylabel("Velocity")
ax.set_title("{0}. t = {1}. Target X Pos = {2}".format(
vf.info['title'],
vf.info['time'],
vf.info['target_pos'],
))
img = vu.save_image(fig)
report.add_image(img, "Target Position = {}".format(target_pos))
def visualize_histogram(histogram, skew_config, report, title=""):
prob = histogram.pvals
weights = prob_to_weight(prob, skew_config)
xrange, yrange = histogram.xy_range
extent = [xrange[0], xrange[1], yrange[0], yrange[1]]
for name, values in [
('Weight Heatmap', weights),
('Prob Heatmap', prob),
]:
plt.figure()
fig = plt.gcf()
ax = plt.gca()
values = values.copy()
values[values == 0] = np.nan
heatmap_img = ax.imshow(
np.swapaxes(values, 0, 1), # imshow uses first axis as y-axis
extent=extent,
cmap=plt.get_cmap('plasma'),
interpolation='nearest',
aspect='auto',
origin='bottom', # <-- Important! By default top left is (0, 0)
# norm=LogNorm(),
)
divider = make_axes_locatable(ax)
legend_axis = divider.append_axes('right', size='5%', pad=0.05)
fig.colorbar(heatmap_img, cax=legend_axis, orientation='vertical')
heatmap_img = vu.save_image(fig)
if histogram.num_bins < 5:
pvals_str = np.array2string(histogram.pvals, precision=3)
report.add_text(pvals_str)
report.add_image(heatmap_img, "{} {}".format(title, name))
return heatmap_img
def visualize_histogram_samples(
histogram,
report,
dynamics,
n_vis=1000,
title="Histogram Samples",
samples=None,
):
plt.figure()
if samples is None:
generated_samples = histogram.sample(n_vis)
samples = dynamics(generated_samples)
plt.plot(
samples[:, 0],
samples[:, 1],
'.',
)
xlim, ylim = histogram.get_plot_ranges()
plt.xlim(*xlim)
plt.ylim(*ylim)
plt.title(title)
fig = plt.gcf()
sample_img = vu.save_image(fig)
report.add_image(sample_img, title)
return sample_img
def visualize_vae_samples(
epoch, training_data, vae,
report, dynamics,
n_vis=1000,
xlim=(-1.5, 1.5),
ylim=(-1.5, 1.5)
):
plt.figure()
plt.suptitle("Epoch {}".format(epoch))
n_samples = len(training_data)
skip_factor = max(n_samples // n_vis, 1)
training_data = training_data[::skip_factor]
reconstructed_samples = vae.reconstruct(training_data)
generated_samples = vae.sample(n_vis)
projected_generated_samples = dynamics(generated_samples)
plt.subplot(2, 2, 1)
plt.plot(generated_samples[:, 0], generated_samples[:, 1], '.')
if xlim is not None:
plt.xlim(*xlim)
if ylim is not None:
plt.ylim(*ylim)
plt.title("Generated Samples")
plt.subplot(2, 2, 2)
plt.plot(projected_generated_samples[:, 0],
projected_generated_samples[:, 1], '.')
if xlim is not None:
plt.xlim(*xlim)
if ylim is not None:
plt.ylim(*ylim)
plt.title("Projected Generated Samples")
plt.subplot(2, 2, 3)
plt.plot(training_data[:, 0], training_data[:, 1], '.')
if xlim is not None:
plt.xlim(*xlim)
if ylim is not None:
plt.ylim(*ylim)
plt.title("Training Data")
plt.subplot(2, 2, 4)
plt.plot(reconstructed_samples[:, 0], reconstructed_samples[:, 1], '.')
if xlim is not None:
plt.xlim(*xlim)
if ylim is not None:
plt.ylim(*ylim)
plt.title("Reconstruction")
fig = plt.gcf()
sample_img = vu.save_image(fig)
report.add_image(sample_img, "Epoch {} Samples".format(epoch))
return sample_img
def visualize_samples(epoch,
vis_samples_np,
histogram,
report,
dynamics,
n_vis=1000,
title=None,
xlim=(-1.5, 1.5),
ylim=(-1.5, 1.5)):
if title is None:
title = "Epoch {}".format(epoch)
plt.figure()
plt.suptitle(title)
generated_samples = histogram.sample(n_vis)
projected_generated_samples = dynamics(generated_samples)
plt.subplot(3, 1, 1)
plt.plot(generated_samples[:, 0], generated_samples[:, 1], '.')
if xlim is not None:
plt.xlim(*xlim)
if ylim is not None:
plt.ylim(*ylim)
plt.title("Generated Samples")
plt.subplot(3, 1, 2)
plt.plot(projected_generated_samples[:, 0],
projected_generated_samples[:, 1], '.')
if xlim is not None:
plt.xlim(*xlim)
if ylim is not None:
plt.ylim(*ylim)
plt.title("Projected Generated Samples")
plt.subplot(3, 1, 3)
plt.plot(vis_samples_np[:, 0], vis_samples_np[:, 1], '.')
if xlim is not None:
plt.xlim(*xlim)
if ylim is not None:
plt.ylim(*ylim)
plt.title("Original Samples")
fig = plt.gcf()
sample_img = vu.save_image(fig)
report.add_image(sample_img, "Epoch {} Samples".format(epoch))
return sample_img
def main():
parser = argparse.ArgumentParser()
parser.add_argument("folder_path", type=str)
# parser.add_argument("--num_iters", type=int)
args = parser.parse_args()
base = Path(os.getcwd())
base = base / args.folder_path
path_and_iter = get_path_and_iters(base)
resolution = 20
x_bounds = (-1, 1)
y_bounds = (-1, 1)
report = HTMLReport(
str(base / 'report.html'), images_per_row=1
)
# for path, itr in takespread(path_and_iter, args):
for path, itr in [
path_and_iter[-1]
]:
report.add_text("Path: %s" % path)
print("Loading: %s" % path)
data = joblib.load(str(path))
qf = data['qf']
env = data['env']
qf.train(False)
start_state = env.reset()
report.add_text("Start State = {}".format(start_state))
report.add_text(
"Start XY = {}".format(
position_from_angles(np.expand_dims(start_state, 0))
)
)
goal_states = [start_state]
goal_states += [
env.sample_goal_for_rollout()
for _ in range(5)
]
for goal_state in goal_states:
qf_eval = create_qf_eval_fnct(qf, start_state, goal_state)
qf_heatmap = vu.make_heat_map(
qf_eval,
x_bounds=x_bounds,
y_bounds=y_bounds,
resolution=resolution,
)
fig = create_figure(
['Estimated'],
[qf_heatmap],
)
img = vu.save_image(fig)
report.add_image(img, "Goal State = {}\nGoal XY = {}".format(
goal_state,
position_from_angles(np.expand_dims(goal_state, 0))
))
abs_path = osp.abspath(report.path)
print("Report saved to: {}".format(abs_path))
report.save()
open_report = query_yes_no("Open report?", default="yes")
if open_report:
cmd = "xdg-open {}".format(abs_path)
print(cmd)
subprocess.call(cmd, shell=True)
def generate_report(fanova_info: FanovaInfo, base_dir, param_name_to_log=None):
if param_name_to_log is None:
param_name_to_log = {}
f, config_space, X, Y, categorical_remapping, variants_list = fanova_info
report = HTMLReport(
osp.join(base_dir, 'report.html'),
images_per_row=3,
)
vis = visualizer.Visualizer(f, config_space)
cs_params = config_space.get_hyperparameters()
importances = [get_param_importance(f, param) for param in cs_params]
are_logs = [
is_data_log_uniformly_distributed(X[:, i])
for i in range(len(cs_params))
]
data = sorted(
zip(cs_params, importances, are_logs),
key=lambda x: -x[1],
)
"""
List out how the categorical hyperparameters were mapped.
"""
for name, remapping in categorical_remapping.items():
report.add_text("Remapping for {}:".format(name))
for key, value in remapping.inverse_dict.items():
report.add_text("\t{} = {}\n".format(key, value))
"""
Plot individual marginals.
"""
print("Creating marginal plots")
for param, importance, is_log in data:
param_name = param.name
if param_name in param_name_to_log:
is_log = param_name_to_log[param_name]
if isinstance(param, CategoricalHyperparameter):
vis.plot_categorical_marginal(param_name, show=False)
else:
vis.plot_marginal(param_name, show=False, log_scale=is_log)
img = vu.save_image()
report.add_image(
img,
"Marginal for {}.\nImportance = {}".format(param_name, importance),
)
"""
Plot pairwise marginals.
"""
print("Creating pairwise-marginal plots")
num_params = len(cs_params)
num_pairs = num_params * (num_params + 1) // 2
pair_and_importance = (f.get_most_important_pairwise_marginals(num_pairs))
for combi, importance in pair_and_importance.items():
param_names = []
for p in combi:
param_names.append(cs_params[p].name)
info_text = "Pairwise Marginal for {}.\nImportance = {}".format(
param_names,
importance,
),
if any(is_categorical(f, name) for name in param_names):
report.add_text(info_text)
continue
plot_pairwise_marginal(vis, combi, show=False)
img = vu.save_image()
report.add_image(
img,
txt=info_text,
)
"""
List the top 10 parameters.
"""
N = min(10, len(Y))
Y[np.isnan(Y)] = np.nanmin(Y) - 1
best_idxs = Y.argsort()[-N:][::-1]
all_param_names = [p.name for p in config_space.get_hyperparameters()]
for rank, i in enumerate(best_idxs):
variant = variants_list[i]
report.add_text("Rank {} params, with score = {}:".format(
rank + 1, Y[i]))
for name, value in zip(all_param_names, X[i, :]):
report.add_text("\t{} = {}\n".format(name, value))
report.add_text("\texp_name = {}\n".format(variant['exp_name']))
report.add_text("\tunique_id = {}\n".format(variant['unique_id']))
print("Guesses for is_log")
print("{")
for param, _, is_log in data:
name = param.name
print(" '{}': {},".format(name, is_log))
print("}")
"""
Ask user if they want to see the report
"""
abs_path = osp.abspath(report.path)
print("Report saved to: {}".format(abs_path))
report.save()
open_report = query_yes_no("Open report?", default="yes")
if open_report:
cmd = "xdg-open {}".format(abs_path)
print(cmd)
subprocess.call(cmd, shell=True)