def objective(params):
print(params)
x_train_scaled = np.concatenate((x_train[:, 0:68]*params['kmer_scale'], x_train[:, 68:85]*params['mean_scale'], x_train[:, 85:102]*params['std_scale'], x_train[:, 102:119]*params['len_scale'], x_train[:, 119:]*params['signal_scale']), axis=1)
x_test_scaled = np.concatenate((x_test[:, 0:68]*params['kmer_scale'], x_test[:, 68:85]*params['mean_scale'], x_test[:, 85:102]*params['std_scale'], x_test[:, 102:119]*params['len_scale'], x_test[:, 119:]*params['signal_scale']), axis=1)
encoder, decoder, vae = load_vae_dna_model_deepsignal(int(params['latent_dim']), params['rc_scale'], params['vae_lr'], params['kmer_loss_scale'])
try:
supervised_size = 10000
es = EarlyStopping(monitor='val_loss', mode='min', patience=20)
vae.fit(x_train_scaled[0:int(len(x_train) * 0.8)], validation_data=(x_train_scaled[int(len(x_train) * 0.8):], None), epochs=args.vae_epochs, batch_size=args.vae_batch_size, verbose=0, callbacks=[es])
predictor = load_vae_predictor(int(params['latent_dim']))
# Prepared predictor training input
x_train_predictor = encoder.predict(x_train_scaled[0:supervised_size])
es = EarlyStopping(monitor='val_loss', mode='min', patience=10)
predictor.fit(x_train_predictor, y_train[0:supervised_size], epochs=args.predictor_epochs, validation_split=0.2,
batch_size=args.predictor_batch_size, callbacks=[es], verbose=0)
# Test model
x_test_predictor = encoder.predict(x_test_scaled)
predictions = predictor.predict(x_test_predictor)
test_results = compute_metrics_standardized(predictions, y_test)
if test_results[0] < 0.51:
encoding_cluster_plt = plot_label_clusters(encoder, x_train, y_train)
save_results(args.output_filename, test_results, encoding_cluster_plt, encoder, predictor)
print_results(test_results)
return test_results[0] * -1
except:
return 0
for idx, row in enumerate(x_val_five_mer):
label = val_five_mer_label[idx]
model_idx = model_names.index(label)
encoding = models[model_idx].predict(np.expand_dims(row, axis=0))
x_val_encoding.append(encoding[0])
x_val_encoding = np.array(x_val_encoding)
x_val_encoding_diff = np.zeros(x_val_encoding.shape)
for i, five_mer in enumerate(y_val_five_mer):
five_mer_idx = np.where(five_mer==1)[0][0]
x_val_encoding_diff[i] = (non_mod_encoding_map[five_mer_idx] - x_val_encoding[i]) ** 2
diff = np.mean((non_mod_encoding_map[five_mer_idx] - x_val_encoding[i]) ** 2)
if diff > threshold[five_mer_idx]:
predictions.append(1)
else:
predictions.append(0)
test_results = compute_metrics_standardized(predictions, y_val)
print_results(test_results)
kmeans = KMeans(n_clusters=2, random_state=0).fit(x_val_encoding_diff)
predictions = kmeans.labels_
print(accuracy_score(y_val, predictions))
print(accuracy_score(y_val, 1-predictions))
'''
predictor = load_vae_predictor(32)
x_test_five_mer, _ = extract_five_mer_data(x_test)
x_test_predictor = model_t.predict(x_test_five_mer[0:10000])
es = EarlyStopping(monitor='val_loss', mode='min', patience=10)
predictor.fit(x_test_predictor, y_test[0:10000], epochs=args.predictor_epochs, validation_split=0.2, batch_size=args.predictor_batch_size, callbacks=[es])
# Evaluate
x_val_five_mer, _ = extract_five_mer_data(x_val)
x_val_predictor = model_t.predict(x_val_five_mer[0:10000])
print(vae.decoder.predict(vae.encoder.predict(x_train[0:5, :]))[:, 24:44])
predictor = vae.predictor
encoder = vae.encoder
plot_label_clusters(args.output_filename, encoder, x_train, y_train)
# Train predictor
predictor_size = int(len(x_train)/10)
x_train_mean, x_train_sd, _ = encoder.predict(x_train[0:predictor_size])
x_train = np.concatenate((x_train_mean, x_train_sd), axis=1)
predictor.fit(x_train, y_train[0:predictor_size], epochs=30, batch_size=128)
# Test model
x_test_mean, x_test_sd, _ = encoder.predict(x_test)
x_test = np.concatenate((x_test_mean, x_test_sd), axis=1)
pred_out = predictor.predict(x_test)
accuracy_val, sensitivity_val, specificity_val, precision_val, au_roc_val, cm_val = compute_metrics_standardized(
pred_out, y_test)
# Save model
#save_vae_model_dna(encoder, predictor, min_values, max_values)
# Print results
print(f"\tAccuracy : {accuracy_val:.3f}")
print(f"\tSensitivity : {sensitivity_val:.3f}")
print(f"\tSpecificity : {specificity_val:.3f}")
print(f"\tPrecision : {precision_val:.3f}")
print(f"\tAUC : {au_roc_val:.3f}")
print(f"{cm_val}")
test_x_10, test_y_10 = load_multiple_reads_data(args)
plot_label_clusters_10(args.output_filename, encoder, test_x_10, test_y_10)
vae.compile(optimizer=keras.optimizers.Adam())
vae.fit(x_train, epochs=30, batch_size=128)
# Visualize cluster
encoder = vae.encoder
predictor = vae.predictor
plot_label_clusters(args.output_filename, encoder, x_train, y_train)
# Train predictor
x_train_mean, x_train_sd, _ = encoder.predict(x_train[0:5000])
x_train = np.concatenate((x_train_mean, x_train_sd), axis=1)
predictor.fit(x_train, y_train[0:5000], epochs=30, batch_size=128)
# Test model
x_test_mean, x_test_sd, _ = encoder.predict(x_test)
x_test = np.concatenate((x_test_mean, x_test_sd), axis=1)
pred_out = predictor.predict(x_test)
accuracy_val, sensitivity_val, specificity_val, precision_val, au_roc_val, cm_val = compute_metrics_standardized(
pred_out, y_test)
# Save model
save_vae_model_rna(args, encoder, predictor)
# Print results
print(f"\tAccuracy : {accuracy_val:.3f}")
print(f"\tSensitivity : {sensitivity_val:.3f}")
print(f"\tSpecificity : {specificity_val:.3f}")
print(f"\tPrecision : {precision_val:.3f}")
print(f"\tAUC : {au_roc_val:.3f}")
print(f"{cm_val}")
from utils.arguments import parse_args
from utils.data import load_dna_data_gan
from utils.evaluate import compute_metrics_standardized
from utils.gan_model import load_deep_signal_supervised
import numpy as np
import tensorflow as tf
args = parse_args()
np.random.seed(args.seed)
tf.compat.v1.set_random_seed(args.seed)
x_train, x_test, y_test, x_val, y_val = load_dna_data_gan(args)
model = load_deep_signal_supervised(args)
model.fit(x_test,
y_test,
epochs=150,
batch_size=512,
validation_data=(x_val, y_val))
y_predicted = np.squeeze(model.predict_on_batch(x_val))
results = (compute_metrics_standardized(y_predicted, y_val))
print(results)
def train(args, generator, discriminator, GAN, x_train, x_test, y_test, x_val,
y_val):
# Adversarial Training
epochs = args.epochs
batch_size = args.batch_size
v_freq = args.v_freq
latent_dim = args.latent_dim
d_loss, g_loss, best_cm = [], [], []
best_au_roc_val, best_accuracy, best_sensitivity, best_specificity, best_precision, best_au_roc = 0, 0, 0, 0, 0, 0
print('===== Start of Adversarial Training =====')
for epoch in range(epochs):
try:
with trange(x_train.shape[0] // batch_size,
ascii=True,
desc='Epoch {}'.format(epoch + 1)) as t:
for _ in t:
# Train Discriminator
loss_temp = []
set_trainability(discriminator, True)
K.set_value(gamma, [1])
x, y = D_data(int(batch_size / 2), generator, 'normal',
x_train, latent_dim)
loss_temp.append(discriminator.train_on_batch(x, y))
set_trainability(discriminator, True)
K.set_value(gamma, [args.gamma])
x, y = D_data(int(batch_size), generator, 'gen', x_train,
latent_dim)
loss_temp.append(discriminator.train_on_batch(x, y))
d_loss.append(sum(loss_temp) / len(loss_temp))
# Train Generator
set_trainability(discriminator, False)
x = noise_data(batch_size, latent_dim)
y = np.ones(batch_size)
y[:] = args.alpha
g_loss.append(GAN.train_on_batch(x, y))
t.set_postfix(G_loss=g_loss[-1], D_loss=d_loss[-1])
except KeyboardInterrupt:
# hit control-C to exit
break
if (epoch + 1) % v_freq == 0:
# Check for the best validation results
y_predicted = 1 - np.squeeze(discriminator.predict_on_batch(x_val))
x = noise_data(batch_size, latent_dim)
print(generator.predict_on_batch(x)[0, 24:44])
accuracy_val, sensitivity_val, specificity_val, precision_val, au_roc_val, cm_val = compute_metrics_standardized(
y_predicted, y_val)
if au_roc_val > best_au_roc_val:
best_au_roc_val = au_roc_val
# Save the best test results
y_predicted = 1 - np.squeeze(
discriminator.predict_on_batch(x_test))
best_accuracy, best_sensitivity, best_specificity, best_precision, best_au_roc, best_cm = compute_metrics_standardized(
y_predicted, y_test)
save_gan_model(args, discriminator)
print(f"\tAccuracy : {accuracy_val:.3f}")
print(f"\tSensitivity : {sensitivity_val:.3f}")
print(f"\tSpecificity : {specificity_val:.3f}")
print(f"\tPrecision : {precision_val:.3f}")
print(f"\tAUC : {au_roc_val:.3f}")
print(f"{cm_val}")
print(f"\tGen. Loss: {g_loss[-1]:.3f}\n\tDisc. Loss: {d_loss[-1]:.3f}")
print('===== End of Adversarial Training =====')
print(f"\tBest accuracy : {best_accuracy:.3f}")
print(f"\tBest sensitivity : {best_sensitivity:.3f}")
print(f"\tBest specificity : {best_specificity:.3f}")
print(f"\tBest precision : {best_precision:.3f}")
print(f"\tBest AUC : {best_au_roc:.3f}")
print(f"{best_cm}")
results = (best_accuracy, best_sensitivity, best_specificity,
best_precision, best_au_roc)
return results