epochs=adv_numberOfEpochs)
#Adversarial training
losses = {"L_f": [], "L_r": [], "L_f - L_r": []}
batch_size = 128
num_epochs=200
for i in range(num_epochs):
print i
l = DRf.evaluate(test_x, [test_y, df_Convert_v2_Test])
losses["L_f - L_r"].append(l[0][None][0])
losses["L_f"].append(l[1][None][0])
losses["L_r"].append(-l[2][None][0])
print(losses["L_r"][-1] / lam)
plot_losses(i, losses, lam, num_epochs)
#Fit "model"
model.trainable = True
advmodel.trainable = False
model.compile(loss=[make_loss_model(c=1.0)], optimizer=opt_model, metrics=['accuracy'])
DRf.compile(loss=[make_loss_model(c=1.0), make_loss_advmodel(c=-lam)], optimizer=opt_DRf)
DfR.compile(loss=[make_loss_advmodel(c=1.0)], optimizer=opt_DfR)
indices = np.random.permutation(len(train_x))[:batch_size]
#model.train_on_batch(train_x.iloc[indices],train_y.iloc[indices])
DRf.train_on_batch(train_x.iloc[indices], [train_y.iloc[indices], df_Convert_v2.iloc[indices]])
#Fit "advmodel"
nJets_binned_bkg, callbacks=[earlystop2], epochs = adv_numberOfEpochs) # Adversarial training num_epochs = 200 for i in range(num_epochs): print 'Adversarial training epoch: ', i+1 l = DRf.evaluate([test_x, test_x.dM_Go_LSP], [test_y, nJets_binned_test]) losses["L_f - L_r"].append(l[0][None][0]) losses["L_f"].append(l[1][None][0]) losses["L_r"].append(-l[2][None][0]) print(losses["L_r"][-1] / lam) plot_losses(i, losses, lam, num_epochs, 'Losses_nJets_PAT_noLO_HT_lambda'+str(lam)) #Fit "model" model.trainable = True advmodel.trainable = False model.compile(loss=[make_loss_model(c=1.0)], optimizer=opt_model, metrics=['accuracy']) DRf.compile(loss=[make_loss_model(c=1.0), make_loss_advmodel(c=-lam)], optimizer=opt_DRf) DfR.compile(loss=[make_loss_advmodel(c=1.0)], optimizer=opt_DfR) indices = np.random.permutation(len(train_x))[:batch_size] DRf.train_on_batch([train_x.iloc[indices], train_x.dM_Go_LSP.iloc[indices]], [train_y.iloc[indices], nJets_binned.iloc[indices]]) #Fit "advmodel" if lam >= 0.0: model.trainable = False
print 'Epoch mean loss: ', round(np.mean(epoch_losses), 5)
for (x_batch, y_batch) in val_data_batched:
classifier_model.compile(loss=make_disco_loss(
x_batch.numpy()[:, decorr_var_col]),
optimizer=opt_model,
metrics=['accuracy'])
val_loss, val_acc = classifier_model.evaluate(x_batch,
y_batch,
batch_size=batchSize)
val_epoch_losses.append(val_loss)
tf.keras.backend.clear_session()
print 'Epoch validation mean loss: ', round(np.mean(val_epoch_losses),
5)
losses["L_t"].append(np.mean(epoch_losses))
losses["L_v"].append(np.mean(val_epoch_losses))
plot_losses(epoch, losses, lam, numberOfEpochs,
'Losses_DisCo_dilepton_lambda_' + str(lam))
tf.keras.backend.clear_session()
epoch_end_time = time.time()
print 'Epoch time elapsed: ', np.round(
(epoch_end_time - epoch_start_time), 3)
print "End of epoch: ", epoch + 1
print '-------------'
print ' - Test set ROC AUC: ', round(
roc_auc_score(test_y, classifier_model.predict(test_x)), 4)
# Save the model
save_path = '/work/kimmokal/susyDNN/models/'
save_name = 'susyDNN_DisCo_dilepton_nJets_lambda' + str(lam) + '_' + str(
mass_point)
classifier_model.save(save_path + save_name + '.h5')
hist_dnn_output_njet_6to8)
js_distances["JS1"].append(js1)
js_distances["JS2"].append(js2)
print 'DNN output Jensen-Shannon distance (nJet = [6,7,8] vs. nJet = [4,5]): ' + str(
js1)
print 'DNN output Jensen-Shannon distance (nJet >= 9 vs. nJet = [6,7,8]): ' + str(
js2)
### Save losses and plot
l = DRf.evaluate(test_x, [test_y, nJets_binned_test], batch_size=512)
losses["L_f - L_r"].append(l[0][None][0])
losses["L_f"].append(l[1][None][0])
losses["L_r"].append(-l[2][None][0])
print("Loss: " + str(losses["L_r"][-1] / lam))
plot_losses(i, losses, lam, num_epochs,
'Losses_adversarial_' + model_name + '_lambda_' + str(lam))
plot_jensenshannon(i, js_distances, lam, num_epochs,
'JS_distance_' + model_name + '_lambda_' + str(lam))
plot_inefficiencies(
i, inefficiencies_compressed, inefficiencies_uncompressed, lam,
num_epochs, 'Inefficiencies_' + model_name + '_lambda' + str(lam))
roc_aoc = 1 - round(roc_auc_score(test_y, model.predict(test_x)), 4)
print 'ROC area over curve: ' + str(roc_aoc)
# Save the metrics for the best models to a .csv file
if (js1 <= 0.08 and js2 <= 0.08 and sig_uncompressed_ineff <= 0.7
and sig_compressed_ineff <= 0.8
and bkg_uncompressed_ineff <= 0.1):
metrics_path = save_path + model_name + '_best_metrics.txt'
metrics_to_csv(metrics_path, i + 1, js1, js2,
print 'Epoch mean loss: ', round(np.mean(epoch_losses), 5)
for (x_batch, y_batch) in val_data_batched:
classifier_model.compile(loss=make_disco_loss(
x_batch.numpy()[:, decorr_var_col]),
optimizer=opt_model,
metrics=['accuracy'])
val_loss, val_acc = classifier_model.evaluate(x_batch,
y_batch,
batch_size=batchSize)
val_epoch_losses.append(val_loss)
tf.keras.backend.clear_session()
print 'Epoch validation mean loss: ', round(np.mean(val_epoch_losses),
5)
losses["L_t"].append(np.mean(epoch_losses))
losses["L_v"].append(np.mean(val_epoch_losses))
plot_losses(epoch, losses, lam, numberOfEpochs,
'Losses_DisCo_all_bkg_lambda_' + str(lam))
tf.keras.backend.clear_session()
epoch_end_time = time.time()
print 'Epoch time elapsed: ', np.round(
(epoch_end_time - epoch_start_time), 3)
print "End of epoch: ", epoch + 1
print '-------------'
print ' - Test set ROC AUC: ', round(
roc_auc_score(test_y, classifier_model.predict(test_x)), 4)
# Save the model
save_path = '/work/kimmokal/susyDNN/models/'
save_name = 'susyDNN_DisCo_all_bkg_nJets_lambda' + str(lam) + '_' + str(
mass_point)
classifier_model.save(save_path + save_name + '.h5')
print "Signal Inefficiency Uncompressed: " + str(Signal_Uncompressed_Ineff)
print "Bkg Inefficiency Uncompressed: " + str(Bkg_Uncompressed_Ineff)
#///////////////////////#
### Save losses and plot
l = DRf.evaluate(test_x, [test_y, nJets_binned_test])
losses["L_f - L_r"].append(l[0][None][0])
losses["L_f"].append(l[1][None][0])
losses["L_r"].append(-l[2][None][0])
print("Loss L_r: " + str(losses["L_r"][-1] / lam))
print("Loss L_f: " + str(losses["L_f"][-1]))
print(" ")
if (i == num_epochs - 1):
plot_losses(i, losses, lam, num_epochs,
'Losses_adversarial_reduced_bkg_lambda' + str(lam))
plot_jensenshannon(i, js_distances, lam, num_epochs,
'JS_distance_reduced_bkg_lambda' + str(lam))
plot_Inefficiencies(i, inefficiencies_Compressed,
inefficiencies_Uncompressed, lam, num_epochs,
'Inefficiencies_reduced_bkg_lambda' + str(lam))
save_name = 'susyDNN_adv_model_reduced_bkg_lambda_' + str(lam) + "_" + str(
i)
model.save(save_path2 + save_name + '.h5')
print '- - - - - - -'
print ' - second test set roc auc: ', round(
roc_auc_score(test_y, model.predict(test_x)), 4)
# Check decorrelation for test set
nJets_binned_bkg, callbacks=[earlystop2], epochs = adv_numberOfEpochs) # Adversarial training num_epochs = 200 for i in range(num_epochs): print 'Adversarial training epoch: ', i+1 l = DRf.evaluate(test_x, [test_y, nJets_binned_test]) losses["L_f - L_r"].append(l[0][None][0]) losses["L_f"].append(l[1][None][0]) losses["L_r"].append(-l[2][None][0]) print(losses["L_r"][-1] / lam) plot_losses(i, losses, lam, num_epochs, 'Losses_nJets_dilepton_lambda'+str(lam)) #Fit "model" model.trainable = True advmodel.trainable = False model.compile(loss=[make_loss_model(c=1.0)], optimizer=opt_model, metrics=['accuracy']) DRf.compile(loss=[make_loss_model(c=1.0), make_loss_advmodel(c=-lam)], optimizer=opt_DRf) DfR.compile(loss=[make_loss_advmodel(c=1.0)], optimizer=opt_DfR) indices = np.random.permutation(len(train_x))[:batch_size] DRf.train_on_batch(train_x.iloc[indices], [train_y.iloc[indices], nJets_binned.iloc[indices]]) #Fit "advmodel" if lam >= 0.0: model.trainable = False