def select_best_generators(self):
# computing metrics
gan_metrics = compute_metrics(self.target, self.cond_input, self.prob_config, self.check_config, self.coord_input, self.neighbours)
min_rmse_val = None
min_mie_val = None
# selecting and sampling gan
for criteria in list(self.check_config["perf_metrics"].keys()): # criteria is either RMSE or MIE
# find best config
criteria_info = self.check_config["pf_metrics_setting"][criteria]
perf_metrics = gan_metrics[criteria_info["metric_level"]]
perf_values = criteria_info["agg_function"](perf_metrics[[criteria]])
# print("perf_values for "+criteria,perf_values) # just out of curiosity
if(criteria == "RMSE"):
min_rmse_val = perf_values[criteria_info["rank_function"](perf_values)]
print("min perf_value for "+criteria, min_rmse_val)
else:
min_mie_val = perf_values[criteria_info["rank_function"](perf_values)]
print("min perf_value for "+criteria, min_mie_val)
best_config = perf_metrics.index[criteria_info["rank_function"](perf_values)] # the training step that has the best generator
print("best "+criteria+" at iteration ",best_config) # just out of curiosity
# generate samples of synthetic data
gan_samples_df = self.generate_samples(best_config)
# pdb.set_trace()
# if curious, print out gan_sample_df and its shape and its header
# export results
self.save_model(gan_samples_df, "grid_" + criteria + ".pkl.gz")
self.save_model(gan_metrics["agg_metrics"], "grid_checkmetrics.pkl.gz")
gen_method = Generator(prob_config["cond_dim"], prob_config["noise_dim"],
prob_config["output_dim"])
gen_method.to(prob_config["device"])
# training SpaceGAN
spacegan = SpaceGAN(prob_config, check_config, disc_method, gen_method)
spacegan.train(x_train=cond_input, y_train=target, coords=coord_input)
# export final model and data
spacegan.checkpoint_model(spacegan.epochs)
spacegan.df_losses.to_pickle(model_save_prefix + "grid_spaceganlosses.pkl.gz")
# pick the best Generator (G) as determined by the MIE and the RMSE criterion.
# computing metrics
gan_metrics = compute_metrics(target, cond_input, prob_config, check_config,
coord_input, neighbours)
# selecting and sampling gan
for criteria in list(check_config["perf_metrics"].keys()):
# find best config
criteria_info = check_config["pf_metrics_setting"][criteria]
perf_metrics = gan_metrics[criteria_info["metric_level"]]
perf_values = criteria_info["agg_function"](perf_metrics[[criteria]])
best_config = perf_metrics.index[criteria_info["rank_function"](
perf_values)]
# get and set best space gan
best_spacegan = get_spacegan_config(int(best_config), prob_config,
check_config, cond_input, target)
# training samples
gan_samples_df = pd.DataFrame(index=range(cond_input.shape[0]),
gen_method = spacegan_config.Generator(
spacegan_config.prob_config["cond_dim"],
spacegan_config.prob_config["noise_dim"],
spacegan_config.prob_config["output_dim"])
gen_method.to(spacegan_config.prob_config["device"])
# training SpaceGAN
spacegan = spacegan_method.SpaceGAN(spacegan_config.prob_config,
spacegan_config.check_config,
disc_method, gen_method)
spacegan.train(x_train=cond_input, y_train=target, coords=coord_input)
spacegan.checkpoint_model(spacegan.epochs) # export final model and data
# computing metrics
gan_metrics = spacegan_selection.compute_metrics(
target, cond_input, spacegan_config.prob_config,
spacegan_config.check_config, coord_input, spacegan_config.neighbours)
# selecting and sampling gan
for criteria in list(spacegan_config.check_config["perf_metrics"].keys()):
# find best config
criteria_info = spacegan_config.check_config["pf_metrics_setting"][
criteria]
perf_metrics = gan_metrics[criteria_info["metric_level"]]
perf_values = criteria_info["agg_function"](perf_metrics[[criteria]])
best_config = perf_metrics.index[criteria_info["rank_function"](
perf_values)]
# get and set best space gan
best_spacegan = spacegan_selection.get_spacegan_config(
int(best_config), spacegan_config.prob_config,