def test_epoch_end(self, outputs):
# avg_loss = torch.stack([x['test_loss'] for x in outputs]).mean()
self.combine_movies_with_z_values()
avg_loss = np.array([x['test_loss'] for x in outputs]).mean()
mse_test = np.array([x['MSE-Test'] for x in outputs])
kld_test =np.array([x['KLD-Test'] for x in outputs])
# ls_mce = {x['mce'] for x in outputs}
utils.save_dict_as_json(self.mce_batch_test, 'mce_results.json', self.experiment_path_test)
# avg_mce = dict(utils.calculate_mean_of_ls_dict(ls_mce))
# avg_rmse = np.array([x['rmse'] for x in outputs]).mean()
# avg_rmse_w_zeros = np.array([x['rmse_w_zeros'] for x in outputs]).mean()
# avg_mse = np.array([x['mse'] for x in outputs]).mean()
# avg_mse_w_zeros = np.array([x['mse_w_zeros'] for x in outputs]).mean()
tensorboard_logs = {'test_loss': avg_loss, 'MSE-Test':mse_test,'KLD-Test': kld_test }
assert len(mse_test)==len(kld_test)
for i in range(0, len(mse_test)):
wandb_logger.log_metrics({'MSE-Test': mse_test[i],'KLD-Test': kld_test[i]} )
# wandb_logger.log_metrics({'rmse': avg_rmse,
# 'rmse_w_zeros':avg_rmse_w_zeros,
# 'mse': avg_mse,
# 'mse_w_zeros': avg_mse_w_zeros})#, 'kld_matrix':self.kld_matrix
return {'test_loss': avg_loss,
'log': tensorboard_logs,
# 'rmse': avg_rmse,
'MSE-Test':mse_test,
'KLD-test': kld_test }#, , 'mce':avg_mce
def training_epoch_end(self, outputs):
#print("Saving MCE before KLD is applied...")
avg_logvar = np.array([x['logvar'] for x in outputs]).mean(axis=0)
# avg_var = np.array([x['var'] for x in outputs]).mean(axis=0)
avg_mu = np.array([x['mu'] for x in outputs]).mean(axis=0)
self.np_logvar_train = np.append(self.np_logvar_train, [avg_logvar], axis=0)
self.np_mu_train = np.append(self.np_mu_train, [avg_mu], axis=0)
if(self.current_epoch == self.sigmoid_annealing_threshold ):
utils.save_dict_as_json(self.mce_batch_train, 'mce_results_wo_kld.json', self.experiment_path_train)
return {}
def test_epoch_end(self, outputs):
# avg_loss = torch.stack([x['test_loss'] for x in outputs]).mean()
self.combine_movies_with_z_values()
avg_loss = np.array([x['test_loss'] for x in outputs]).mean()
mse_test = np.array([x['MSE-Test'] for x in outputs])
kld_test = np.array([x['KLD-Test'] for x in outputs])
avg_var = np.array([x['var'] for x in outputs]).mean(axis=0)
avg_mean = np.array([x['mean'] for x in outputs]).mean(axis=0)
self.np_var_test = np.append(self.np_var_test, [avg_var], axis=0)
# ls_mce = {x['mce'] for x in outputs}
# utils.save_dict_as_json(outputs[0]['mce'], 'mce_results.json', self.experiment_path) #TODO This is wrong as only one batch is processed, should be all
utils.save_dict_as_json(self.mce_batch_test, 'mce_results.json',
self.experiment_path_test)
# avg_mce = dict(calculate_mean_of_ls_dict(ls_mce))
avg_rmse = np.array([x['rmse'] for x in outputs]).mean()
avg_rmse_w_zeros = np.array([x['rmse_w_zeros']
for x in outputs]).mean()
avg_mse = np.array([x['mse'] for x in outputs]).mean()
avg_mse_w_zeros = np.array([x['mse_w_zeros'] for x in outputs]).mean()
tensorboard_logs = {
'test_loss': avg_loss,
'MSE-Test': mse_test,
'KLD-Test': kld_test
}
assert len(mse_test) == len(kld_test)
for i in range(0, len(mse_test)):
wandb_logger.log_metrics({
'MSE-Test': mse_test[i],
'KLD-Test': kld_test[i]
})
wandb_logger.log_metrics({
'rmse': avg_rmse,
'rmse_w_zeros': avg_rmse_w_zeros,
'mse': avg_mse,
'mse_w_zeros': avg_mse_w_zeros
}) #, 'kld_matrix':self.kld_matrix
return {
'test_loss': avg_loss,
'log': tensorboard_logs,
'rmse': avg_rmse,
'MSE-Test': mse_test,
'KLD-test': kld_test
} #, , 'mce':avg_mce
def generate_distribution_df():
dct_attribute_distribution = utils.compute_relative_frequency(
pd.read_csv('../data/generated/syn.csv'))
utils.save_dict_as_json(dct_attribute_distribution, 'syn_attribute_distribution.json')
