def exec_ensemble(models):
# train models
for m in models:
train_model(m,
DATASET_INDEX,
dataset_prefix='lp1_',
epochs=1000,
batch_size=128)
# Predict labels with models
_, _, X_test, y_test, is_timeseries = load_dataset_at(DATASET_INDEX)
labels = []
for m in models:
predicts = np.argmax(m.predict(X_test), axis=1)
labels.append(predicts)
# Ensemble with voting
labels = np.array(labels)
labels = np.transpose(labels, (1, 0))
labels_result = scipy.stats.mode(labels, axis=1)[0]
labels_result = np.squeeze(labels_result)
# cal accuracy
match_count = 0
for i in range(len(labels_result)):
p = labels_result[i]
e = y_test[i][0]
if p == e:
match_count += 1
accuracy = (match_count * 100) / len(labels_result)
print('Accuracy: ', accuracy)
return accuracy
k: width factor
Returns: a keras tensor
'''
filters = input._keras_shape[-1] # channel_axis = -1 for TF
se = GlobalAveragePooling1D()(input)
se = Reshape((1, filters))(se)
se = Dense(filters // 16,
activation='relu',
kernel_initializer='he_normal',
use_bias=False)(se)
se = Dense(filters,
activation='sigmoid',
kernel_initializer='he_normal',
use_bias=False)(se)
se = multiply([input, se])
return se
if __name__ == "__main__":
model = generate_model_2()
train_model(model,
DATASET_INDEX,
dataset_prefix='hhar',
epochs=100,
batch_size=128)
evaluate_model(model, DATASET_INDEX, dataset_prefix='hhar', batch_size=128)
model = Model(inp, output)
model.summary()
return model
if __name__ == "__main__":
# generate_ model
model5 = generate_model_5() #
print("GRU-FCN")
history = train_model(model5,
DATASET_INDEX,
dataset_prefix='adiac',
epochs=4000,
batch_size=128)
accuracy5, loss5 = evaluate_model(model5,
DATASET_INDEX,
dataset_prefix='adiac',
batch_size=128)
print("--- Run Time = %s seconds ---" % ((time.time() - start_time)))
print("--- Run Time = %s minutes ---" %
((time.time() - start_time) / 60.0))
text_file = open("training_time.txt", "w")
text_file.write("--- Run Time =" + str(((time.time() - start_time))) +
" seconds ---" + "\n" + "--- Run Time = " +
str(((time.time() - start_time) / 60.0)) + " minutes ---" +
"\n")
max_seq_lenth = max_seq_len(dataset)
nb_class = nb_classes(dataset)
nb_cnn = int(round(math.log(max_seq_lenth, 2))-3)
print("Number of Pooling Layers: %s" % str(nb_cnn))
#model = lstm_fcn_model(proto_num, max_seq_lenth, nb_class)
#model = alstm_fcn_model(proto_num, max_seq_lenth, nb_class)
#model = cnn_raw_model(nb_cnn, proto_num, max_seq_lenth, nb_class)
#model = cnn_dtwfeatures_model(nb_cnn, proto_num, max_seq_lenth, nb_class)
#model = cnn_earlyfusion_model(nb_cnn, proto_num, max_seq_lenth, nb_class)
#model = cnn_midfusion_model(nb_cnn, proto_num, max_seq_lenth, nb_class)
#model = cnn_latefusion_model(nb_cnn, proto_num, max_seq_lenth, nb_class)
#model = vgg_raw_model(nb_cnn, proto_num, max_seq_lenth, nb_class)
#model = vgg_dtwfeatures_model(nb_cnn, proto_num, max_seq_lenth, nb_class)
#model = vgg_earlyfusion_model(nb_cnn, proto_num, max_seq_lenth, nb_class)
model = vgg_midfusion_model(nb_cnn, proto_num, max_seq_lenth, nb_class)
#model = vgg_latefusion_model(nb_cnn, proto_num, max_seq_lenth, nb_class)
print("Number of Pooling Layers: %s" % str(nb_cnn))
train_model(model, dataset, method, proto_num, dataset_prefix=dataset, nb_iterations=50000, batch_size=50, opt='Nadam', learning_rate=0.0001, early_stop=False, balance_classes=False, run_ver='vgg_')
acc = evaluate_model(model, dataset, method, proto_num, dataset_prefix=dataset, batch_size=50, checkpoint_prefix="vgg_loss")
np.savetxt("output/vgg/vgg-%s-%s-%s-loss-%s" % (dataset, method, str(proto_num), str(acc)), [acc])
acc = evaluate_model(model, dataset, method, proto_num, dataset_prefix=dataset, batch_size=50, checkpoint_prefix="vgg_val_acc")
np.savetxt("output/vgg/vgg-%s-%s-%s-vacc-%s" % (dataset, method, str(proto_num), str(acc)), [acc])
print("Mahis F-Score : ", averagef_mahi)
print("Mahis W F-Score : ", weightedf_mahi)
return averagef
if __name__ == "__main__":
model = generate_model_2()
model.compile('adam',
'categorical_crossentropy',
metrics=['accuracy', f1_score])
train_model(model,
DATASET_INDEX,
dataset_prefix='opportunity_new',
epochs=1000,
batch_size=128,
monitor='val_f1_score',
optimization_mode='max',
compile_model=False)
evaluate_model(model,
DATASET_INDEX,
dataset_prefix='opportunity',
batch_size=128)
predict_model(
model,
DATASET_INDEX,
dataset_prefix='opportunity_weights_attention_9208_512_lstm_128',
batch_size=512)
'''
filters = input._keras_shape[-1] # channel_axis = -1 for TF
se = GlobalAveragePooling1D()(input)
se = Reshape((1, filters))(se)
se = Dense(filters // 16,
activation='relu',
kernel_initializer='he_normal',
use_bias=False)(se)
se = Dense(filters,
activation='sigmoid',
kernel_initializer='he_normal',
use_bias=False)(se)
se = multiply([input, se])
return se
if __name__ == "__main__":
model = generate_model_2()
train_model(model,
DATASET_INDEX,
dataset_prefix='ozone',
epochs=600,
batch_size=128)
evaluate_model(model,
DATASET_INDEX,
dataset_prefix='ozone',
batch_size=128)
'''
filters = input._keras_shape[-1] # channel_axis = -1 for TF
se = GlobalAveragePooling1D()(input)
se = Reshape((1, filters))(se)
se = Dense(filters // 16,
activation='relu',
kernel_initializer='he_normal',
use_bias=False)(se)
se = Dense(filters,
activation='sigmoid',
kernel_initializer='he_normal',
use_bias=False)(se)
se = multiply([input, se])
return se
if __name__ == "__main__":
model = generate_model_2()
train_model(model,
DATASET_INDEX,
dataset_prefix='japanese_vowels',
epochs=600,
batch_size=128)
evaluate_model(model,
DATASET_INDEX,
dataset_prefix='japanese_vowels',
batch_size=128)
'''
filters = input._keras_shape[-1] # channel_axis = -1 for TF
se = GlobalAveragePooling1D()(input)
se = Reshape((1, filters))(se)
se = Dense(filters // 16,
activation='relu',
kernel_initializer='he_normal',
use_bias=False)(se)
se = Dense(filters,
activation='sigmoid',
kernel_initializer='he_normal',
use_bias=False)(se)
se = multiply([input, se])
return se
if __name__ == "__main__":
model = generate_model_2()
train_model(model,
DATASET_INDEX,
dataset_prefix='activity_attention',
epochs=1000,
batch_size=128)
evaluate_model(model,
DATASET_INDEX,
dataset_prefix='activity_attention',
batch_size=128)
'''
filters = input._keras_shape[-1] # channel_axis = -1 for TF
se = GlobalAveragePooling1D()(input)
se = Reshape((1, filters))(se)
se = Dense(filters // 16,
activation='relu',
kernel_initializer='he_normal',
use_bias=False)(se)
se = Dense(filters,
activation='sigmoid',
kernel_initializer='he_normal',
use_bias=False)(se)
se = multiply([input, se])
return se
if __name__ == "__main__":
model = generate_model_2()
train_model(model,
DATASET_INDEX,
dataset_prefix='uwave_attention',
epochs=500,
batch_size=128)
evaluate_model(model,
DATASET_INDEX,
dataset_prefix='uwave_attention',
batch_size=128)
'''
filters = input._keras_shape[-1] # channel_axis = -1 for TF
se = GlobalAveragePooling1D()(input)
se = Reshape((1, filters))(se)
se = Dense(filters // 16,
activation='relu',
kernel_initializer='he_normal',
use_bias=False)(se)
se = Dense(filters,
activation='sigmoid',
kernel_initializer='he_normal',
use_bias=False)(se)
se = multiply([input, se])
return se
if __name__ == "__main__":
model = generate_model_2()
train_model(model,
DATASET_INDEX,
dataset_prefix='movement_aal',
epochs=1000,
batch_size=128)
evaluate_model(model,
DATASET_INDEX,
dataset_prefix='movement_aal',
batch_size=128)
'''
filters = input._keras_shape[-1] # channel_axis = -1 for TF
se = GlobalAveragePooling1D()(input)
se = Reshape((1, filters))(se)
se = Dense(filters // 16,
activation='relu',
kernel_initializer='he_normal',
use_bias=False)(se)
se = Dense(filters,
activation='sigmoid',
kernel_initializer='he_normal',
use_bias=False)(se)
se = multiply([input, se])
return se
if __name__ == "__main__":
model = generate_model()
train_model(model,
DATASET_INDEX,
dataset_prefix='gesture_phase',
epochs=1000,
batch_size=128)
evaluate_model(model,
DATASET_INDEX,
dataset_prefix='gesture_phase',
batch_size=128)
from utils.model_utils import lstm_fcn_model, alstm_fcn_model
from utils.model_utils import cnn_raw_model, cnn_dtwfeatures_model, cnn_earlyfusion_model, cnn_midfusion_model, cnn_latefusion_model
import sys
if __name__ == "__main__":
dataset = sys.argv[1]
method = sys.argv[2]
proto_num = int(sys.argv[3])
nb_cnn = int(sys.argv[4])
max_seq_lenth = max_seq_len(dataset)
nb_class = nb_classes(dataset)
model = lstm_fcn_model(proto_num, max_seq_lenth, nb_class)
#model = alstm_fcn_model(proto_num, max_seq_lenth, nb_class)
#model = cnn_raw_model(nb_cnn, proto_num, max_seq_lenth, nb_class)
#model = cnn_dtwfeatures_model(nb_cnn, proto_num, max_seq_lenth, nb_class)
#model = cnn_earlyfusion_model(nb_cnn, proto_num, max_seq_lenth, nb_class)
#model = cnn_midfusion_model(nb_cnn, proto_num, max_seq_lenth, nb_class)
#model = cnn_latefusion_model(nb_cnn, proto_num, max_seq_lenth, nb_class)
train_model(model, dataset, method, proto_num, dataset_prefix=dataset, nb_iterations=20000, batch_size=50, learning_rate=0.0001, early_stop=True)
evaluate_model(model, dataset, method, proto_num, dataset_prefix=dataset, batch_size=50, checkpoint_prefix="loss")
evaluate_model(model, dataset, method, proto_num, dataset_prefix=dataset, batch_size=50, checkpoint_prefix="val_acc")
'''
filters = input._keras_shape[-1] # channel_axis = -1 for TF
se = GlobalAveragePooling1D()(input)
se = Reshape((1, filters))(se)
se = Dense(filters // 16,
activation='relu',
kernel_initializer='he_normal',
use_bias=False)(se)
se = Dense(filters,
activation='sigmoid',
kernel_initializer='he_normal',
use_bias=False)(se)
se = multiply([input, se])
return se
if __name__ == "__main__":
model = generate_model_2()
train_model(model,
DATASET_INDEX,
dataset_prefix='walk_vs_run_',
epochs=1000,
batch_size=128)
evaluate_model(model,
DATASET_INDEX,
dataset_prefix='walk_vs_run_',
batch_size=128)
se = GlobalAveragePooling1D()(input)
se = Reshape((1, filters))(se)
se = Dense(filters // 16,
activation='relu',
kernel_initializer='he_normal',
use_bias=False)(se)
se = Dense(filters,
activation='sigmoid',
kernel_initializer='he_normal',
use_bias=False)(se)
se = multiply([input, se])
return se
if __name__ == "__main__":
model = generate_model()
train_model(model,
DATASET_INDEX,
dataset_prefix='vehicle_data',
epochs=500,
batch_size=16)
evaluate_model(model,
DATASET_INDEX,
dataset_prefix='vehicle_data',
batch_size=16)
model.save('/Users/jhk1m/J_ZOOM/test_model.h5')
'''
filters = input._keras_shape[-1] # channel_axis = -1 for TF
se = GlobalAveragePooling1D()(input)
se = Reshape((1, filters))(se)
se = Dense(filters // 16,
activation='relu',
kernel_initializer='he_normal',
use_bias=False)(se)
se = Dense(filters,
activation='sigmoid',
kernel_initializer='he_normal',
use_bias=False)(se)
se = multiply([input, se])
return se
if __name__ == "__main__":
model = generate_model_2()
train_model(model,
DATASET_INDEX,
dataset_prefix='floodPrediction_',
epochs=200,
batch_size=128)
evaluate_model(model,
DATASET_INDEX,
dataset_prefix='floodPrediction_',
batch_size=128)
#model = cnn_raw_model(nb_cnn, proto_num, max_seq_lenth, nb_class)
#model = cnn_dtwfeatures_model(nb_cnn, proto_num, max_seq_lenth, nb_class)
#model = cnn_earlyfusion_model(nb_cnn, proto_num, max_seq_lenth, nb_class)
model = cnn_midfusion_model_v2(nb_cnn, dim_num, proto_num, max_seq_lenth,
nb_class)
#model = cnn_latefusion_model(nb_cnn, proto_num, max_seq_lenth, nb_class)
print("Number of Pooling Layers: %s" % str(nb_cnn))
train_model(model,
dataset,
method,
proto_num,
dataset_prefix=dataset,
nb_iterations=50000,
batch_size=32,
normalize_timeseries=True,
learning_rate=0.0001,
early_stop=False,
balance_classes=False)
acc = evaluate_model(model,
dataset,
method,
proto_num,
dataset_prefix=dataset,
batch_size=50,
normalize_timeseries=True,
checkpoint_prefix="loss")
np.savetxt(
'''
filters = input._keras_shape[-1] # channel_axis = -1 for TF
se = GlobalAveragePooling1D()(input)
se = Reshape((1, filters))(se)
se = Dense(filters // 16,
activation='relu',
kernel_initializer='he_normal',
use_bias=False)(se)
se = Dense(filters,
activation='sigmoid',
kernel_initializer='he_normal',
use_bias=False)(se)
se = multiply([input, se])
return se
if __name__ == "__main__":
model = generate_model()
train_model(model,
DATASET_INDEX,
dataset_prefix='daily_sport_no_attention',
epochs=500,
batch_size=128)
evaluate_model(model,
DATASET_INDEX,
dataset_prefix='daily_sport_no_attention',
batch_size=128)
y = Conv1D(128, 3, padding='same', kernel_initializer='he_uniform')(y)
y = BatchNormalization()(y)
y = Activation('relu')(y)
y = GlobalAveragePooling1D()(y)
x = concatenate([x, y])
out = Dense(NB_CLASS, activation='softmax')(x)
model = Model(ip, out)
model.summary()
# add load model code here to fine-tune
return model
if __name__ == "__main__":
model = generate_model_2()
train_model(model, DATASET_INDEX, dataset_prefix='synthetic_control', epochs=4000, batch_size=16)
evaluate_model(model, DATASET_INDEX, dataset_prefix='synthetic_control', batch_size=128)
# visualize_context_vector(model, DATASET_INDEX, dataset_prefix='synthetic_control', visualize_sequence=True,
# visualize_classwise=True, limit=1)
# visualize_cam(model, DATASET_INDEX, dataset_prefix='synthetic_control', class_id=0)
if not os.path.exists('weights/' + weights_dir):
os.makedirs('weights/' + weights_dir)
dataset_name_ = weights_dir + dname
# try:
model = model_fn(MAX_SEQUENCE_LENGTH, NB_CLASS, cell)
print('*' * 20, "Training model for dataset %s" % (dname),
'*' * 20)
# comment out the training code to only evaluate !
train_model(model,
did,
dataset_name_,
epochs=2000,
batch_size=128,
normalize_timeseries=normalize_dataset)
acc = evaluate_model(model,
did,
dataset_name_,
batch_size=128,
normalize_timeseries=normalize_dataset)
s = "%d,%s,%s,%0.6f\n" % (did, dname, dataset_name_, acc)
file.write(s)
file.flush()
successes.append(s)
se = GlobalAveragePooling1D()(input)
se = Reshape((1, filters))(se)
se = Dense(filters // 16,
activation='relu',
kernel_initializer='he_normal',
use_bias=False)(se)
se = Dense(filters,
activation='sigmoid',
kernel_initializer='he_normal',
use_bias=False)(se)
se = multiply([input, se])
return se
if __name__ == "__main__":
#model = generate_model_2()
model = generate_model()
train_model(model,
DATASET_INDEX,
dataset_prefix='geo',
epochs=500,
batch_size=16,
normalize_timeseries=False)
evaluate_model(model,
DATASET_INDEX,
dataset_prefix='geo',
batch_size=16,
normalize_timeseries=False)
