def eval_accuracy(net_path, name):
path = 'UCRArchive_2018/' + name + '/' + name + '_TEST.tsv'
net = torch.load(net_path)
net.eval()
test_set, n_class = load_ucr(path)
# x_test = test_set[:, 1:]
# x_test = torch.from_numpy(x_test).unsqueeze(dim=-1).float()
# print(x_test.shape)
# y_true = test_set[:, 0]
# y_pred = net(x_test)
# y_pred = torch.argmax(y_pred, dim=-1)
# print(y_pred)
# y_pred = y_pred.detach().numpy()
dataset = UcrDataset(test_set, channel_last=opt.channel_last)
dataloader = UCR_dataloader(dataset, batch_size=128)
y_pred = []
y_true = []
with torch.no_grad():
for i, (data, label) in enumerate(dataloader):
data = data.float().to(device)
y_ = net(data).cpu()
y_ = torch.argmax(y_, dim=1)
y_pred.append(y_.detach().numpy())
y_true.append(label.view(-1).long())
y_pred = np.concatenate(y_pred)
y_true = np.concatenate(y_true)
res = calculate_metrics(y_true, y_pred)
return res
def train(name):
record = pd.DataFrame(data=np.zeros((1, 4), dtype=np.float),
columns=['precision', 'accuracy', 'recall', 'F1'])
for _ in range(opt.runs):
seed = random.randint(1, 10000)
print("Random Seed: ", seed)
torch.manual_seed(seed)
# mkdirs for checkpoints output
os.makedirs(opt.checkpoints_folder, exist_ok=True)
os.makedirs('%s/%s' % (opt.checkpoints_folder, name), exist_ok=True)
os.makedirs('report_metrics', exist_ok=True)
root_dir = 'report_metrics/%s_aug_%s_IMBA/%s' % (
opt.model, str(opt.n_group), name)
os.makedirs(root_dir, exist_ok=True)
# 加载数据集
path = 'UCRArchive_2018/' + name + '/' + name + '_TRAIN.tsv'
train_set, n_class = load_ucr(path)
print('启用平衡数据增强！')
stratified_train_set = stratify_by_label(train_set)
data_aug_set = data_aug_by_dft(stratified_train_set, opt.n_group)
total_set = np.concatenate((train_set, data_aug_set))
print('Shape of total set', total_set.shape)
dataset = UcrDataset(total_set, channel_last=opt.channel_last)
batch_size = int(min(len(dataset) / 10, 16))
dataloader = UCR_dataloader(dataset, batch_size)
# Common behavior
seq_len = dataset.get_seq_len() # 初始化序列长度
# 创建分类器对象\损失函数\优化器
if opt.model == 'r':
net = ResNet(n_in=seq_len, n_classes=n_class).to(device)
if opt.model == 'f':
net = ConvNet(n_in=seq_len, n_classes=n_class).to(device)
criterion = nn.CrossEntropyLoss().to(device)
optimizer = optim.Adam(net.parameters(), lr=opt.lr)
scheduler = optim.lr_scheduler.ReduceLROnPlateau(optimizer,
mode='min',
factor=0.5,
patience=50,
min_lr=0.0001)
min_loss = 10000
print('############# Start to Train ###############')
net.train()
for epoch in range(opt.epochs):
for i, (data, label) in enumerate(dataloader):
data = data.float()
data = data.to(device)
label = label.long()
label = label.to(device)
optimizer.zero_grad()
output = net(data)
loss = criterion(output, label.view(label.size(0)))
loss.backward()
optimizer.step()
scheduler.step(loss)
# print('[%d/%d][%d/%d] Loss: %.8f ' % (epoch, opt.epochs, i + 1, len(dataloader), loss.item()))
if loss < min_loss:
min_loss = loss
# End of the epoch,save model
print('MinLoss: %.10f Saving the best epoch model.....' %
min_loss)
torch.save(
net, '%s/%s/%s_%s_best_IMBA.pth' %
(opt.checkpoints_folder, name, opt.model, str(
opt.n_group)))
net_path = '%s/%s/%s_%s_best_IMBA.pth' % (opt.checkpoints_folder, name,
opt.model, str(opt.n_group))
one_record = eval_accuracy(net_path, name)
print('The minimum loss is %.8f' % min_loss)
record = record.append(one_record, ignore_index=True)
record = record.drop(index=[0])
record.loc['mean'] = record.mean()
record.loc['std'] = record.std()
record.to_csv(root_dir + '/metrics.csv')
# all_reprot_metrics.loc[name, 'acc_mean'] = record.at['mean', 'accuracy']
# all_reprot_metrics.loc[name, 'acc_std'] = record.at['std', 'accuracy']
# all_reprot_metrics.loc[name, 'F1_mean'] = record.at['mean', 'F1']
# all_reprot_metrics.loc[name, 'F1_std'] = record.at['std', 'F1']
print('\n')
# inverse the frequency domain into time domain with the same time step
X_combined = np.fft.irfft(sampled_X_dft, n_time_steps)
# add the corresponding label
Y = Y[:sampled_X_dft.shape[0], np.newaxis]
data_combined = np.concatenate((Y, X_combined), axis=1)
# print('Shape of th combined data', data_combined.shape)
data_augmented.append(data_combined)
print('\n# Succeed to augment the data! #')
data_augmented = np.concatenate(data_augmented)
print('Shape of the augmented data', data_augmented.shape)
return data_augmented
if __name__ == '__main__':
data_path = 'UCRArchive_2018/FaceAll/FaceAll_TRAIN.tsv'
data, n_class = load_ucr(data_path)
x_ori = data[:, 1:]
y_ori = data[:, 0]
time_step = len(data[0]) - 1
stratified_data = stratify_by_label(data)
data_aug = data_aug_by_dft(stratified_data,
n_group=4,
train_size=len(data))
# x_aug = data_aug[:, 1:]
# y_aug = data_aug[:, 0]
# plot_tsne(x_ori, x_aug, y_ori, y_aug,n_class=n_class)
# plot_ori_tsne(x_ori, y_ori)
x = data_aug[0, 1:]
plt.axis('off')
plt.plot(stratified_data[0][4, 1:], color='b')
plt.plot(x, color='r')
from dft_aug import data_aug_by_dft
from utils import UcrDataset, UCR_dataloader, load_ucr, stratify_by_label
import numpy as np
# Press the green button in the gutter to run the script.
if __name__ == '__main__':
path = 'UCRArchive_2018/Car/Car_TRAIN.tsv'
data, n_class = load_ucr(path)
print(data.shape)
stratified_data = stratify_by_label(data)
data_aug = data_aug_by_dft(stratified_data, ratio=0.4, n_group=4)
data_ori_and_aug = np.concatenate((data, data_aug))
dataset = UcrDataset(data_ori_and_aug)
dataloader = UCR_dataloader(dataset, batch_size=148)
for i, (data, label) in enumerate(dataloader):
print(data.size())
print(label)
# See PyCharm help at https://www.jetbrains.com/help/pycharm/