##################################################
for run in range(RUNS_IN_FOLD):
print('\nFold {0} runs {1}'.format(FLAGS.fold, run))
# use model check point callbacks
bst_model_path = '../data/output/model/nn_fold{0}_run{1}_{2}.h5'.format(FLAGS.fold, run, FE_TYPE)
model_checkpoint = ModelCheckpoint(
bst_model_path,
monitor='val_acc',
mode='max',
save_best_only=True,
save_weights_only=True
)
model = get_model()
# st(context=21)
hist = model.fit_generator(
generator=train_generator.generator(FE_TYPE),
steps_per_epoch=train_generator.steps_per_epoch,
epochs=epochs,
validation_data=(X_valid, y_valid),
callbacks=[lr_plateau, model_checkpoint, early_stopping, csv_logger],
shuffle=True,
)
gc.collect()
bst_acc = max(hist.history['val_acc'])
print('\nBest model val_acc : %.5f\n' % bst_acc)
model.load_weights(bst_model_path)
# os.remove(bst_model_path)
# st(context=21)
preds_test += model.predict(X_test, batch_size=batch_size) / RUNS_IN_FOLD
preds_valid += model.predict(X_valid, batch_size=batch_size) / RUNS_IN_FOLD
preds = np.zeros((len(fname_test), n_classes))
for run in range(RUNS_IN_FOLD):
print('fold {0} runs {1}'.format(FLAGS.fold, run))
# use model check point callbacks
bst_model_path = './tmp/nn_fold{0}_run{1}_{2}.h5'.format(FLAGS.fold, run, FE_TYPE)
model_checkpoint = ModelCheckpoint(
bst_model_path,
monitor='val_acc',
mode='max',
save_best_only=True,
save_weights_only=True
)
model = get_model()
# st(context=21)
hist = model.fit_generator(
generator=train_generator.generator(FE_TYPE),
steps_per_epoch=train_generator.steps_per_epoch,
epochs=epochs,
validation_data=valid_generator.generator(FE_TYPE),
validation_steps=valid_generator.steps_per_epoch,
callbacks=[lr_plateau, model_checkpoint],
shuffle=True,
)
gc.collect()
bst_acc = max(hist.history['val_acc'])
print('best model val_acc : {0}'.format(bst_acc))
model.load_weights(bst_model_path)
# os.remove(bst_model_path)
preds += model.predict(X_test, batch_size=256) / RUNS_IN_FOLD
del model
gc.collect()
augmentation_prob=50,
)
train_generator.steps_per_epoch = train_generator.steps_per_epoch * 2
valid_generator = AudioGenerator(
root_dir= '../data/input/train/audio/',
k=FLAGS.fold,
file_temp=VALID_SPLIT_FILE_TEMP,
ori_batch_size=batch_size,
train_or_valid='valid',
)
preds = np.zeros((len(fname_test), n_classes))
for run in range(RUNS_IN_FOLD):
print('fold {0} runs {1}'.format(FLAGS.fold, run))
model = get_model()
model.fit_generator(
generator=train_generator.generator(),
steps_per_epoch=train_generator.steps_per_epoch,
epochs=epochs,
validation_data=valid_generator.generator(),
validation_steps=valid_generator.steps_per_epoch,
# callbacks=[lr_scheduler],
shuffle=True,
)
gc.collect()
preds += model.predict(X_test, batch_size=256) / RUNS_IN_FOLD
del model
gc.collect()
labels_index =np.argmax(preds, axis=1)
del X_test
gc.collect()
