def _save_fig(self, metric, plot_name):
timestamp = datetime.datetime.now().strftime('%Y%m%d_%H%M%S_%f')
filepath = os.path.join(
self.out_dir, '%s__%s__%s.png' % (metric, plot_name, timestamp))
io_utils.makedirs(os.path.dirname(filepath))
with open(filepath, 'wb') as f:
plt.savefig(f)
def write_results(results, outfile_prefix):
"""Write results to JSON."""
io_utils.makedirs(os.path.dirname(outfile_prefix))
results_path = outfile_prefix + 'results.json'
with open(results_path, 'w') as outfile:
json.dump(results, outfile, cls=_NumpyEncoder)
logging.info('Results written to: %s', results_path)
return results_path
def _write_pval_result(self, pval, algo1, algo2, timeframe=None):
"""Write p-value to text file."""
io_utils.makedirs(self.outfile_dir)
outfile_path = ('%s/%s_%s_%s' %
(self.outfile_dir, self.metric, algo1, algo2))
if timeframe is not None:
outfile_path += '_%d' % timeframe
with open(outfile_path, 'w') as outfile:
outfile.write('%g' % pval)
logging.info('P-val result written to: %s', outfile_path)
def write_metric_params(self, outfile_prefix):
"""Write metric parameters to JSON."""
# Load the metric params.
metric_params = get_metric_params(self.metrics)
# Write the metric params.
io_utils.makedirs(os.path.dirname(outfile_prefix))
params_path = outfile_prefix + 'metric_params.json'
with open(params_path, 'w') as outfile:
json.dump(metric_params, outfile, cls=_NumpyEncoder)
logging.info('Metric params written to: %s', params_path)
return params_path
def _write_confidence_interval_result(self,
ci_lower,
ci_upper,
algo,
timeframe=None):
"""Write confidence interval to text file."""
io_utils.makedirs(self.outfile_dir)
outfile_path = '%s/%s_%s' % (self.outfile_dir, self.metric, algo)
if timeframe is not None:
outfile_path += '_%d' % timeframe
with open(outfile_path, 'w') as outfile:
outfile.write('%g,%g' % (ci_lower, ci_upper))
logging.info('Confidence interval written to: %s', outfile_path)
def make_training_curve_plots(algos,
tasks,
n_runs_per_expt,
csv_filepath_template,
figure_outdir,
window_size=None,
subplot_height=5,
subplot_width=8):
"""Make four different plots of the training curves.
(1) Raw training curves, one line per run.
One subplot per (algo, task) combination.
(2) Median + 5th/95th percentiles across runs.
One subplot per (algo, task) combination.
(3) Median + 5th/95th percentiles across runs.
One subplot per task (all algos plotted).
(4) Means + 95% bootstrap CIs across runs.
One subplot per task (all algos plotted).
Args:
algos: List of strings. Which algorithms we are analyzing.
tasks: List of strings. Which tasks we are analyzing.
n_runs_per_expt: Number of runs per (algo, task) combination.
csv_filepath_template: String path to CSV files containing training curves.
e.g. '/my/path/%s_%s_%d.csv'. Should accept (task, algo, run).
figure_outdir: Path to directory for saving figures.
window_size: If the training curves are not aligned (if different curves are
evaluated at different timepoints), we can specify a window_size that
aggregates, to allow computing summaries across curves like means,
medians, percentiles, and confidence intervals.
subplot_height: Height of each subplot.
subplot_width: Width of each subplot.
"""
n_algo = len(algos)
n_task = len(tasks)
plt.figure('raw',
figsize=(subplot_width * n_algo, subplot_height * n_task))
plt.figure('medians_percentiles',
figsize=(subplot_width * n_algo, subplot_height * n_task))
n_subplots_x, n_subplots_y = subplots_square(n_task)
plt.figure('medians_percentiles_pertask',
figsize=(subplot_width * n_subplots_x,
subplot_height * n_subplots_y))
plt.figure('means_CIs_pertask',
figsize=(subplot_width * n_subplots_x,
subplot_height * n_subplots_y))
fig_names = [
'raw', 'medians_percentiles', 'medians_percentiles_pertask',
'means_CIs_pertask'
]
subplot_pos = 0
# Iterate through each task.
for i_task, task in enumerate(tasks):
print('%s...' % task, end='')
# Initialize x- and y-lims.
xlims_extremes = [np.inf, -np.inf]
task_baselines = [
baseline for key, baseline in BASELINES.items() if task in key
]
if task_baselines:
ylims_extremes = [np.inf, max(task_baselines)]
else:
ylims_extremes = [np.inf, -np.inf]
# Iterate through each algorithm.
for i_algo, algo in enumerate(algos):
subplot_pos += 1
algo_color = ALGO_COLORS[i_algo]
plt.figure('raw')
plt.subplot(n_task, n_algo, subplot_pos)
# Load and plot the raw curves.
curves = []
for run in range(n_runs_per_expt):
csv_filepath = csv_filepath_template % (task, algo, run)
with open(csv_filepath, 'r') as csv_file:
csv_reader = csv.reader(csv_file)
curve = []
for _ in range(2):
curve.append(
np.array(csv_reader.next(), dtype=np.float))
curves.append(curve)
plt.plot(curve[0], curve[1])
plot_baseline(algo, task)
# update the xlim/ylim extremes
xlims_extremes, ylims_extremes = update_xylims_extremes(
xlims_extremes, ylims_extremes)
# Compute summaries
curves = np.array(curves)
timesteps, window_means = compute_window_means(curves, window_size)
means = compute_means(window_means)
medians = compute_medians(window_means)
cis = compute_boot_ci(window_means)
percentiles = compute_percentiles(window_means)
# plot the medians + percentiles
plt.figure('medians_percentiles')
plt.subplot(n_task, n_algo, subplot_pos)
fill_plot(timesteps, medians, percentiles[0], percentiles[1], algo,
task, algo_color)
# Plot the medians + percentiles on a single plot per task.
plt.figure('medians_percentiles_pertask')
plt.subplot(n_subplots_y, n_subplots_x, i_task + 1)
fill_plot(timesteps, medians, percentiles[0], percentiles[1], algo,
task, algo_color)
# Plot the mean + CI on a single plot per task.
plt.figure('means_CIs_pertask')
plt.subplot(n_subplots_y, n_subplots_x, i_task + 1)
fill_plot(timesteps, means, cis[0], cis[1], algo, task, algo_color)
# Figure titles.
for fig_name in ['raw', 'medians_percentiles']:
plt.figure(fig_name)
plt.title('%s - %s' % (algo, task))
for fig_name in ['medians_percentiles_pertask', 'means_CIs_pertask']:
plt.figure(fig_name)
plt.title(task)
# equalize axes for the task
for fig_name in ['raw', 'medians_percentiles']:
equalize_axes_for_task(fig_name, xlims_extremes, ylims_extremes,
subplot_pos, n_task, n_algo)
# Add legends
for fig_name in ['medians_percentiles_pertask', 'means_CIs_pertask']:
plt.figure(fig_name)
plt.legend(algos)
# Save the figures.
io_utils.makedirs(figure_outdir)
for fig_name in fig_names:
plt.figure(fig_name)
plt.tight_layout()
output_path = '%s/%s.png' % (figure_outdir, fig_name)
with open(output_path, 'wb') as outfile:
plt.savefig(outfile, dpi=100)
