def test_calculate_aic(self):
global table
table = ac.calculate_aic(data = sampler_args['data'],
models = sampler_args['models'],
priors = sampler_args['priors'],
params = accepted.iloc[-10:]
)
assert len(table)
figs=trace_figs,
AX=trace_AX)
print('\ntrace_AX\n', trace_AX)
#modify the figure setting else all the labels overlapped.
for tax in trace_AX.values():
tax.set_xticks(tax.get_xticks()[::2])
tax.set_yticks(tax.get_yticks()[::2])
tax.tick_params(axis='both', labelsize=16)
tax.xaxis.label.set_size(fontsize=16)
tax.yaxis.label.set_size(fontsize=16)
plt.tight_layout(pad=1.0) # set spacing between figures
#Rank models
table = ac.calculate_aic(data=sampler_args['data'],
models=sampler_args['models'],
priors=sampler_args['priors'],
params=accepted.iloc[-10:])
ranked_table = ac.rank_aic(table, inplace=False)
print('\nRanked AIC table:\n', ranked_table.head())
#export the ranked table into csv file inside the output folder
ranked_table.to_csv(output_folder / 'ranked_table.csv')
best = ranked_table.iloc[0]
best_row_index = best['row']
best_model_num = best['model_num']
new_settings = config_data[best_model_num]
new_settings['settings_name'] = 'LogicGate_OR_bestfitted'
new_settings['parameters'] = cf.get_params_for_model(
sampler_args['models'][model_num]['init'] = model_init
sampler_args['models'][model_num]['int_args'][
'modify_params'] = modify_params
traces = {}
result = cf.simulated_annealing(**sampler_args)
accepted = result['a']
posterior = accepted.iloc[-40::4]
traces[1] = accepted
'''
In order to calculate the AIC, we need the data, models, priors and
parameters for evaluation.
'''
table = ac.calculate_aic(data=sampler_args['data'],
models=sampler_args['models'],
priors=sampler_args['priors'],
params=posterior)
'''
rank_aic accepts a DataFrame containing AIC values indexed under
aic_column_name. It then sorts the DataFrame and adds columns for the normalized
AIC and the evidence for that model. The original columns in the input
DataFrame remain untouched.
As can be seen from the result, rank_aic isn't picky about what indices and
columns you use as long as your DataFrame has its AIC values indexed under aic_column_name.
'''
ranked_table = ac.rank_aic(table, inplace=False)
'''
After choosing the best model and its fitted parameters, we can convert it into
a settings data structure so we can reuse it as a template in the future.