def main():
outputdir = 'result/special_v3/'
if os.path.isdir(outputdir):
print('save in :' + outputdir)
else:
os.makedirs(outputdir)
train_img_path = '/data/MIT-BIH_AD_v3/'
test_img_path = '/data/MIT-BIH_AD_v3/'
train_file = './MIT-BIH_AD_sp_train.txt'
test_file = './MIT-BIH_AD_sp_val.txt'
num_classes = 7
f1 = open(train_file, 'r')
f2 = open(test_file, 'r')
lines = f1.readlines()
f1.close()
train_samples = len(lines)
lines = f2.readlines()
f2.close()
val_samples = len(lines)
batch_size = 32
epochs = 120
input_h = 128
input_w = 128
model = proposed_model(nb_classes=num_classes)
lr = 0.0001
adam = Adam(lr=lr)
model.compile(loss='categorical_crossentropy',
optimizer=adam,
metrics=['accuracy'])
model.summary()
history = model.fit_generator(
generator_train_batch(train_file, batch_size, num_classes,
train_img_path, input_h, input_w),
steps_per_epoch=train_samples // batch_size,
epochs=epochs,
callbacks=[Step()],
validation_data=generator_val_batch(test_file, batch_size, num_classes,
test_img_path, input_h, input_w),
validation_steps=val_samples // batch_size,
verbose=1)
plot_history(history, outputdir)
save_history(history, outputdir)
model.save_weights(outputdir + 'proposed_model')
def main():
proposed = False
if proposed:
outputdir = 'result/NoAugment_{}/'.format(proposed)
if os.path.isdir(outputdir):
print('save in :'+outputdir)
else:
os.makedirs(outputdir)
train_img_path = '/data/MIT-BIH_AD/'
train_file = '/home/ccl/Documents/ECG-Arrhythmia-classification-in-2D-CNN/MIT-BIH_AD_train_paper.txt'
num_classes = 8
k = 10
f1 = open(train_file, 'r')
lines = f1.readlines()
f1.close()
train_samples = len(lines)
val_samples = len(lines)//k
num = len(lines)
new_lines = []
index = [n for n in range(num)]
random.shuffle(index)
for m in range(num):
new_lines.append(lines[index[m]])
lines = new_lines
temp = []
new_lines = []
for i in range(num):
if i % val_samples == 0:
temp = []
new_lines.append(temp)
temp.append(lines[i])
batch_size = 32
epochs = 40
input_h = 96
input_w = 96
augmentation = False
model = proposed_model()
lr = 0.0001
adam = Adam(lr=lr)
model.compile(loss='categorical_crossentropy', optimizer=adam, metrics=['accuracy'])
model.summary()
history = model.fit_generator(generator_train_batch_proposed(new_lines, k, batch_size, num_classes, train_img_path, input_h, input_w, augmentation=augmentation),
steps_per_epoch=train_samples // batch_size,
epochs=epochs,
callbacks=[Step()],
validation_data=generator_val_batch_proposed(new_lines, k, batch_size, num_classes, train_img_path, input_h, input_w, augmentation=augmentation),
validation_steps=val_samples // batch_size,
verbose=1)
plot_history(history, outputdir)
save_history(history, outputdir)
model.save_weights(outputdir+'proposed_model_{}.h5'.format(proposed))
else:
outputdir = 'result/NoAugment_{}/'.format(proposed)
if os.path.isdir(outputdir):
print('save in :' + outputdir)
else:
os.makedirs(outputdir)
train_img_path = '/data/MIT-BIH_AD/'
test_img_path = '/data/MIT-BIH_AD/'
train_file = '/home/ccl/Documents/ECG-Arrhythmia-classification-in-2D-CNN/MIT-BIH_AD_train.txt'
test_file = '/home/ccl/Documents/ECG-Arrhythmia-classification-in-2D-CNN/MIT-BIH_AD_val.txt'
num_classes = 8
f1 = open(train_file, 'r')
f2 = open(test_file, 'r')
lines = f1.readlines()
f1.close()
train_samples = len(lines)
lines = f2.readlines()
f2.close()
val_samples = len(lines)
batch_size = 32
epochs = 40
input_h = 96
input_w = 96
model = proposed_model()
lr = 0.0001
adam = Adam(lr=lr)
model.compile(loss='categorical_crossentropy', optimizer=adam, metrics=['accuracy'])
model.summary()
history = model.fit_generator(
generator_train_batch(train_file, batch_size, num_classes, train_img_path, input_h, input_w),
steps_per_epoch=train_samples // batch_size,
epochs=epochs,
callbacks=[Step()],
validation_data=generator_val_batch(test_file, batch_size, num_classes, test_img_path, input_h, input_w),
validation_steps=val_samples // batch_size,
verbose=1)
plot_history(history, outputdir)
save_history(history, outputdir)
model.save_weights(outputdir+'proposed_model_{}.h5'.format(proposed))
shuffle=True,
use_color=use_color,
use_height=use_height,
use_v1=use_v1,
augment=False)
net = create_votenet(num_points=num_points,
pcd_feature_dims=num_features,
vote_factor=vote_factor,
num_class=Dataset_Config.num_class,
num_head_bin=Dataset_Config.num_heading_bin,
num_size_cluster=Dataset_Config.num_size_cluster,
num_proposal=num_proposals,
mean_size_arr=Dataset_Config.mean_size_arr,
random_proposal=random_proposal,
config=Dataset_Config)
step = Step(lr_decay_step, lr_decay_result, 0)
bn_decay_scheduler = BNDecayScheduler(bn_init=bn_momentum_init,
decay_rate=bn_decay_rate,
interval=bn_decay_interval,
clip=bn_clip)
ckpt = ModelCheckpoint(os.path.join(
log_dir, 'ep{epoch:03d}-loss{loss:.3f}-val_loss{val_loss:.3f}.h5'),
save_best_only=False,
monitor='val_loss',
save_weights_only=True,
period=1)
net.summary()
def loss_components(idx, name):
def choice(y_true, y_pred):
def main():
outputdir = 'result/192_128_class_weight_v3_120eps/'
if os.path.isdir(outputdir):
print('save in :' + outputdir)
else:
os.makedirs(outputdir)
train_img_path = '/data/MIT-BIH_AD_v3/'
test_img_path = '/data/MIT-BIH_AD_v3/'
train_file = './MIT-BIH_AD_train.txt'
test_file = './MIT-BIH_AD_val.txt'
num_classes = 8
f1 = open(train_file, 'r')
f2 = open(test_file, 'r')
lines = f1.readlines()
f1.close()
train_samples = len(lines)
lines = f2.readlines()
f2.close()
val_samples = len(lines)
batch_size = 32
epochs = 120
input_h = 128
input_w = 128
class_weight = {
0: (1 - (75016 / 107620)) * 8,
1: (1 - (8072 / 107620)) * 8,
2: (1 - (7256 / 107620)) * 8,
3: (1 - (2544 / 107620)) * 8,
4: (1 - (7130 / 107620)) * 8,
5: (1 - (7024 / 107620)) * 8,
6: (1 - (106 / 107620)) * 8,
7: (1 - (472 / 107620)) * 8
}
model = proposed_model(nb_classes=num_classes)
lr = 0.0001
adam = Adam(lr=lr)
model.compile(loss='categorical_crossentropy',
optimizer=adam,
metrics=['accuracy'])
model.summary()
history = model.fit_generator(
generator_train_batch(train_file, batch_size, num_classes,
train_img_path, input_h, input_w),
steps_per_epoch=train_samples // batch_size,
epochs=epochs,
callbacks=[Step()],
validation_data=generator_val_batch(test_file, batch_size, num_classes,
test_img_path, input_h, input_w),
validation_steps=val_samples // batch_size,
verbose=1,
class_weight=class_weight)
plot_history(history, outputdir)
save_history(history, outputdir)
model.save_weights(outputdir + 'proposed_model.h5')
def onetenth_4_8_12(lr):
steps = [4, 8,12]
lrs = [lr, lr/10, lr/100,lr/1000]
return Step(steps, lrs)
def wideresnet_step(lr):
steps = [60, 120, 160]
lrs = [lr, lr/5, lr/25, lr/125]
return Step(steps, lrs)
def onetenth_50_75(lr):
steps = [25, 40]
lrs = [lr, lr/10, lr/100]
return Step(steps, lrs)
def dsn_step_200_230(dataset, lr):
steps = [200, 230]
lrs = [lr, lr / 2.5, lr / 25]
return Step(steps, lrs)
def onetenth_200_230(dataset, lr):
steps = [200, 230]
lrs = [lr, lr / 10, lr / 100]
return Step(steps, lrs)
def dsn_step_40_60(dataset, lr):
steps = [40, 60]
lrs = [lr, lr / 2.5, lr / 25]
return Step(steps, lrs)
def nin_nobn_mnist(dataset, lr):
steps = [40, 50]
lrs = [lr, lr / 2, lr / 10]
return Step(steps, lrs)
# severe overfitting (~98% train_acc) at the same time. In conclusion, half data augumentation might be the best choice.
# Perhaps as the network has already used strong dropout (two layers with 0.5 drop rate), strong data
# augumentation harms the performance.
train_set = ModelNet40Dataset(root=data_root,
batch_size=batch_size,
npoints=num_points,
split='train',
shuffle=True,
augment=True)
val_set = ModelNet40Dataset(root=data_root,
batch_size=batch_size,
npoints=num_points,
split='test',
shuffle=False,
augment=False)
lr_scheduler = Step([20, 40, 60], [lr, lr / 10, lr / 100, lr / 1000])
lr_divider = Divide_lr(15, 1.5)
bn_decay_scheduler = BNDecayScheduler(bn_init=bn_momentum_init,
decay_rate=bn_decay_rate,
interval=bn_decay_interval,
clip=bn_clip)
ckpt = ModelCheckpoint(os.path.join(
log_dir, 'ep{epoch:03d}-acc{acc:.3f}-val_acc{val_acc:.3f}.h5'),
save_best_only=True,
monitor='val_acc',
save_weights_only=True,
period=1)
if classifier_name == 'pointnet2':
model = pointnet2_cls_ssg(num_class, num_points, num_dim)
elif classifier_name == 'votenet_backbone':
model = make_classifier(num_points, 0)
def onetenth_10_15_20(lr):
steps = [10, 15, 15]
lrs = [lr, lr / 10, lr / 100, lr / 1000]
return Step(steps, lrs)
def onetenth_35_70(lr):
steps = [35, 70]
lrs = [lr, lr / 10, lr / 100]
return Step(steps, lrs)
def onetenth_20_30_40(lr):
steps = [20, 30, 40]
lrs = [lr, lr/10, lr/100, lr/1000]
return Step(steps, lrs)