def write_data(data_dir, idx, W, L, T, D, N, phi, alpha, beta, chains):
B = generate_lda(T, W, D, N, phi, alpha)
pickle.dump(B, open(os.path.join(data_dir, 'counts.pkl'), 'w'))
write_pb(data_dir, idx, W, T, D, alpha, beta, B)
write_stan(data_dir, idx, W, T, D, alpha, beta, B, chains=chains)
write_prism(data_dir, idx, W, T, D, alpha, beta, B)
write_txt(data_dir, idx, B, T, alpha, beta)
def main():
txt_path = './data.txt'
txt_list = load_txt(txt_path)
data_path = './dataset.txt'
dataset_txt = load_dataset(data_path)
fin_result_list = []
# 对划分的内容逐个处理
print('{} sentence to do.'.format(len(txt_list)))
for txt_line in txt_list:
result_list = []
seg_list = seg_words(txt_line)
print('Sentence index : {}.'.format(txt_list.index(txt_line) + 1))
for seg_word in tqdm.tqdm(list(set(seg_list))):
if len(seg_word) <= 1:
continue
result = judge_txt(seg_word, dataset_txt)
result_list.append((seg_word, result))
# print('{} : {}'.format(seg_word, result))
fin_result_list.extend(result_list)
write_txt(fin_result_list, './result.txt')
result_list = load_txt('./result.txt')
result_list = [
result.replace('(', '').replace(')', '').replace(' ', '').split(',')
for result in result_list
]
top_list = pridict(result_list, 10)
print(top_list)
def __create_concat_txt(files, subTxt):
"""生成 concat.txt
Arguments:
files {[string]} -- 要合并的文件地址
subTxt {[type]} -- concat.txt 的名称
"""
subs = []
sub = "file '{0}'\n"
for f in files:
subs.append(sub.format(f))
utils.write_txt(subTxt, subs)
def save_checkpoint(self, state, stage, index, is_best):
# 保存权重，每一epoch均保存一次，若为最优，则复制到最优权重；index可以区分不同的交叉验证
pth_path = os.path.join(
self.save_path, '%s_%d_%d.pth' % (self.model_type, stage, index))
torch.save(state, pth_path)
if is_best:
print('Saving Best Model.')
write_txt(self.save_path, 'Saving Best Model.')
shutil.copyfile(
os.path.join(self.save_path,
'%s_%d_%d.pth' % (self.model_type, stage, index)),
os.path.join(
self.save_path,
'%s_%d_%d_best.pth' % (self.model_type, stage, index)))
def modify_annotation(self, annotation_file):
# modify the annotation for CAD, as same as VD
new_file = annotation_file.split('.txt')[0] + '_new.txt'
open(new_file, 'w')
f = open(annotation_file)
lines = f.readlines()
cur_frameId = int(lines[0].split('\t')[0])
content_str = ''
action_count = np.zeros([5])
sep = ' '
num = 0
for line in lines:
keywords = line.split('\t')
frame_id = int(keywords[0])
action = int(keywords[5])
if frame_id % 10 == 1 and action != 1:
action = action - 2
x, y, w, h = int(keywords[1]), int(keywords[2]), int(
keywords[3]), int(keywords[4])
x = 0 if x < 0 else x
y = 0 if y < 0 else y
if w <= 0 or h <= 0:
print 'error!'
break
anno_str = sep + str(x) + sep + str(y) + sep + str(
w) + sep + str(h) + sep + str(action)
'''if frame_id == cur_frameId:
action_label_count[action_label] += 1
content_str = content_str + anno_str'''
if frame_id != cur_frameId:
activity = np.argmax(action_count)
content_str = str(cur_frameId) + sep + str(
activity) + content_str + '\n'
utils.write_txt(new_file, content_str, 'a')
num += 1
cur_frameId = frame_id
content_str = ''
action_count = np.zeros([5])
action_count[action] += 1
content_str = content_str + anno_str
activity = np.argmax(action_count)
content_str = str(cur_frameId) + sep + str(
activity) + content_str + '\n'
utils.write_txt(new_file, content_str, 'a')
num += 1
return num
def get_src_trace(detailed_config, out_folder):
# process the cmd
trace_cmd = detailed_config['trace_cmd']
poc = detailed_config['poc']
replace_idx = np.where(np.asarray(trace_cmd) == '***')[0]
cmd = dc(trace_cmd)
replace_num = len(replace_idx)
for id in range(replace_num):
cmd[replace_idx[id]] = poc[id]
# write the cmd
cmd_path = os.path.join(out_folder, 'cmd.txt')
utils.write_txt(cmd_path, [' '.join(cmd)])
# get binary path
bin_path = detailed_config['bin_path']
# get the source trace
tmp_folder = './tempDr'
if not os.path.exists(tmp_folder):
os.mkdir(tmp_folder)
my_parser = parse_dwarf.DwarfParser(bin_path)
flineNumberDict, fileBoundRangesList, fileBoundIndexList, src_filepath = my_parser.get_main_addr(
)
ifSrcList = tracer.findIfSrcInOrderDyn(bin_path,
src_filepath,
flineNumberDict,
fileBoundRangesList,
fileBoundIndexList,
cmdFile=cmd_path)
logging.info("Got the source trace!")
# process the source trace
insn2src = {}
src2insn = {}
for item in ifSrcList:
insn = item[0]
src = '-'.join(item[1:3])
if insn not in insn2src:
insn2src[insn] = src
if src in src2insn:
src2insn[src].add(insn)
else:
src2insn[src] = {insn}
info = {'raw': ifSrcList, 'insn2src': insn2src, 'src2insn': src2insn}
# write the source trace
output_path = os.path.join(out_folder, 'poc_source_trace.pkl')
utils.write_pkl(output_path, info)
logging.info("Recorded the source trace -> %s" % output_path)
return insn2src, src2insn
def load_checkpoint(self, load_optimizer=True):
# Load the pretrained Encoder
weight_path = os.path.join(self.save_path, self.resume)
if os.path.isfile(weight_path):
checkpoint = torch.load(weight_path)
# 加载模型的参数，学习率，优化器，开始的epoch，最小误差等
if torch.cuda.is_available:
self.unet.module.load_state_dict(checkpoint['state_dict'])
else:
self.unet.load_state_dict(checkpoint['state_dict'])
self.start_epoch = checkpoint['epoch']
self.max_dice = checkpoint['max_dice']
if load_optimizer:
self.lr = checkpoint['lr']
self.optimizer.load_state_dict(checkpoint['optimizer'])
print('%s is Successfully Loaded from %s' %
(self.model_type, weight_path))
write_txt(
self.save_path, '%s is Successfully Loaded from %s' %
(self.model_type, weight_path))
else:
raise FileNotFoundError(
"Can not find weight file in {}".format(weight_path))
def put_index(movie_id, total, filename):
utils.write_txt('%s,%s' % (movie_id, total), filename)
def put_index(movie_id, total, filename):
utils.write_txt('%s,%s' % (movie_id, total), filename)
async def writelog(message):
id = str(message.author.id)
if not user_info[id]['isAdmin']:
msg = f"({str(message.channel)}) {datetime.now().strftime('%Y-%m-%d %H:%M:%S')} {user_info[id]['username']} : {message.content.strip().lower()}"
write_txt('test_collect.txt', msg)
predict = mlp_model.predict(validation_data)
'''
'''
torch_mlp = model.torch_train(train_data, train_label, validation_data, validation_label)
test_label = model.torch_predict(test_data, torch_mlp)
predict_torch = model.torch_predict(validation_data, torch_mlp)
for i in range(predict.shape[0]):
print(str(predict[i]) + str(predict_torch[i]) + str(validation_label[i]))
'''
# 线性回归
linear_model = model.Linear_model_build(train_data, train_label)
predict_linear = linear_model.predict(validation_data)
print("linear_loss：" + str(mse(predict_linear, validation_label)))
linear_test_label = linear_model.predict(test_data)
utils.write_txt(linear_test_label)
predict = linear_model.predict(train_data)
delete_list = []
for i in range(train_data.shape[0]):
if abs(predict[i] - train_label[i]) > 1:
delete_list.append(i)
data = []
label = []
for i in range(train_data.shape[0]):
if i not in delete_list:
data.append(list(train_data[i]))
label.append(train_label[i])
train_data = np.array(data)
train_label = np.reshape(np.array(label), [-1, 1])
linear_model = model.Linear_model_build(train_data, train_label)
def validation(self, stage=1):
d_h, d_l = 0, 0
# 验证的时候，train(False)是必须的0，设置其中的BN层、dropout等为eval模式
# with torch.no_grad(): 可以有，在这个上下文管理器中，不反向传播，会加快速度，可以使用较大batch size
self.unet.eval()
tbar = tqdm.tqdm(self.valid_loader)
loss_sum, dice_sum = 0, 0
if stage == 1:
criterion = self.criterion
elif stage == 2:
criterion = self.criterion_stage2
elif stage == 3:
criterion = self.criterion_stage3
with torch.no_grad():
for i, (images, masks) in enumerate(tbar):
images = images.to(self.device)
masks = masks.to(self.device)
net_output = self.unet(images)
net_output_flat = net_output.view(net_output.size(0), -1)
masks_flat = masks.view(masks.size(0), -1)
loss_set = criterion(net_output_flat, masks_flat)
try:
loss_num = len(loss_set)
except:
loss_num = 1
# 依据返回的损失个数分情况处理
if loss_num > 1:
loss = loss_set[0]
else:
loss = loss_set
loss_sum += loss.item()
# 计算dice系数，预测出的矩阵要经过sigmoid含义以及阈值，阈值默认为0.5
net_output_flat_sign = (torch.sigmoid(net_output_flat) >
0.5).float()
dice = self.dice_overall(net_output_flat_sign,
masks_flat).mean()
# if dice>0.8:
#
# masks_i = masks.view((masks.shape[0],-1,masks.shape[1],masks.shape[2]))
# # img_show = torch.cat((images,masks,net_output))
# self.writer.add_images('images_Dice>0.8',images, d_h)
# self.writer.add_images('masks_Dice>0.8',masks_i, d_h)
# self.writer.add_images('preds_Dice>0.8',net_output, d_h)
# d_h += 1
#
# if dice<0.15:
#
# masks_i = masks.view((masks.shape[0],-1,masks.shape[1],masks.shape[2]))
# # img_show = torch.cat((images, masks, net_output))
# self.writer.add_images('images_Dice<0.15', images, d_l)
# self.writer.add_images('masks_Dice<0.15', masks_i, d_l)
# self.writer.add_images('preds_Dice<0.15', net_output, d_l)
# d_l += 1
dice_sum += dice.item()
descript = "Val Loss: {:.7f}, dice: {:.7f}".format(
loss.item(), dice.item())
tbar.set_description(desc=descript)
loss_mean, dice_mean = loss_sum / len(tbar), dice_sum / len(tbar)
print("Val Loss: {:.7f}, dice: {:.7f}".format(loss_mean, dice_mean))
write_txt(
self.save_path,
"Val Loss: {:.7f}, dice: {:.7f}".format(loss_mean, dice_mean))
return loss_mean, dice_mean
def train_stage2(self, index):
# for param in self.unet.module.encoder.parameters():
# param.requires_grad = False
# self.optimizer = optim.Adam(filter(lambda p:p.requires_grad, self.unet.module.parameters()),self.lr_stage2,weight_decay=self.weight_decay)
# # 冻结BN层， see https://zhuanlan.zhihu.com/p/65439075 and https://www.kaggle.com/c/siim-acr-pneumothorax-segmentation/discussion/100736591271 for more information
def set_bn_eval(m):
classname = m.__class__.__name__
if classname.find('BatchNorm') != -1:
m.eval()
# self.optimizer = optim.Adam([{'params': self.unet.decoder.parameters(), 'lr': 1e-5}, {'params': self.unet.encoder.parameters(), 'lr': 1e-7},])
# self.optimizer = optim.Adam(self.unet.module.parameters(), self.lr_stage2, weight_decay=self.weight_decay_stage2
# )
# self.optimizer = NoamLR(self.unet.module.parameters(),10)
self.optimizer = optim.SGD(self.unet.module.parameters(),
lr=self.lr_stage2,
momentum=self.momentum,
weight_decay=self.weight_decay_stage2)
# 加载的resume分为两种情况：之前没有训练第二个阶段，现在要加载第一个阶段的参数；第二个阶段训练了一半要继续训练
if self.resume:
# 若第二个阶段训练一半，要重新加载 TODO
if self.resume.split('_')[-3] == '2':
self.load_checkpoint(
load_optimizer=False) # 当load_optimizer为True会重新加载学习率和优化器
'''
CosineAnnealingLR：若存在['initial_lr']，则从initial_lr开始衰减；
若不存在，则执行CosineAnnealingLR会在optimizer.param_groups中添加initial_lr键值，其值等于lr
重置初始学习率，在load_checkpoint中会加载优化器，但其中的initial_lr还是之前的，所以需要覆盖为self.lr，让其从self.lr衰减
'''
self.optimizer.param_groups[0]['initial_lr'] = self.lr
# 若第一阶段结束后没有直接进行第二个阶段，中间暂停了
elif self.resume.split('_')[-3] == '1':
self.load_checkpoint(load_optimizer=False)
self.start_epoch = 0
self.max_dice = 0
# 第一阶段结束后直接进行第二个阶段，中间并没有暂停
else:
self.start_epoch = 0
self.max_dice = 0
# 防止训练到一半暂停重新训练，日志被覆盖
global_step_before = self.start_epoch * len(self.train_loader)
stage2_epoches = self.epoch_stage2 - self.start_epoch
lr_scheduler = optim.lr_scheduler.CosineAnnealingLR(self.optimizer, 30)
d_h, d_l = 0, 0
for epoch in range(self.start_epoch, self.epoch_stage2):
# if self.epoch >= 20:
# self.unet.apply(set_bn_eval)
epoch += 1
self.unet.train(True)
epoch_loss = 0
self.reset_grad() # 梯度累加的时候需要使用
tbar = tqdm.tqdm(self.train_loader)
for i, (images, masks) in enumerate(tbar):
# GT : Ground Truth
images = images.to(self.device)
masks = masks.to(self.device)
assert images.size(2) == 512
# SR : Segmentation Result
net_output = self.unet(images)
net_output_flat = net_output.view(net_output.size(0), -1)
masks_flat = masks.view(masks.size(0), -1)
loss_set = self.criterion_stage2(net_output_flat, masks_flat)
try:
loss_num = len(loss_set)
except:
loss_num = 1
# 依据返回的损失个数分情况处理
if loss_num > 1:
for loss_index, loss_item in enumerate(loss_set):
if loss_index > 0:
loss_name = 'stage2_loss_%d' % loss_index
self.writer.add_scalar(loss_name, loss_item.item(),
global_step_before + i)
loss = loss_set[0]
else:
loss = loss_set
epoch_loss += loss.item()
# Backprop + optimize, see https://discuss.pytorch.org/t/why-do-we-need-to-set-the-gradients-manually-to-zero-in-pytorch/4903/20 for Accumulating Gradients
if epoch <= self.epoch_stage2 - self.epoch_stage2_accumulation:
self.reset_grad()
loss.backward()
self.optimizer.step()
else:
# loss = loss / self.accumulation_steps # Normalize our loss (if averaged)
loss.backward() # Backward pass
if (
i + 1
) % self.accumulation_steps == 0: # Wait for several backward steps
self.optimizer.step(
) # Now we can do an optimizer step
self.reset_grad()
params_groups_lr = str()
for group_ind, param_group in enumerate(
self.optimizer.param_groups):
params_groups_lr = params_groups_lr + 'params_group_%d' % (
group_ind) + ': %.12f, ' % (param_group['lr'])
# 保存到tensorboard，每一步存储一个
self.writer.add_scalar('Stage2_train_loss', loss.item(),
global_step_before + i)
descript = "Train Loss: %.7f, lr: %s" % (loss.item(),
params_groups_lr)
tbar.set_description(desc=descript)
# 更新global_step_before为下次迭代做准备
global_step_before += len(tbar)
# Print the log info
print('Finish Stage2 Epoch [%d/%d], Average Loss: %.7f' %
(epoch, self.epoch_stage2, epoch_loss / len(tbar)))
write_txt(
self.save_path,
'Finish Stage2 Epoch [%d/%d], Average Loss: %.7f' %
(epoch, self.epoch_stage2, epoch_loss / len(tbar)))
# 验证模型，保存权重，并保存日志
loss_mean, dice_mean = self.validation(stage=2)
if dice_mean > self.max_dice:
is_best = True
self.max_dice = dice_mean
else:
is_best = False
self.lr = lr_scheduler.get_lr()
state = {
'epoch': epoch,
'state_dict': self.unet.module.state_dict(),
'max_dice': self.max_dice,
'optimizer': self.optimizer.state_dict(),
'lr': self.lr
}
self.save_checkpoint(state, 2, index, is_best)
self.writer.add_scalar('Stage2_val_loss', loss_mean, epoch)
self.writer.add_scalar('Stage2_val_dice', dice_mean, epoch)
self.writer.add_scalar('Stage2_lr', self.lr[0], epoch)
# 学习率衰减
lr_scheduler.step()
def train(self, index):
# for param in self.unet.module.encoder.parameters():
# param.requires_grad = False
# self.optimizer = optim.Adam(filter(lambda p:p.requires_grad, self.unet.module.parameters()),self.lr,weight_decay=self.weight_decay)
self.optimizer = optim.Adam(self.unet.module.parameters(),
self.lr,
weight_decay=self.weight_decay)
# 若训练到一半暂停了，则需要加载之前训练的参数，并加载之前学习率 TODO:resume学习率没有接上，所以resume暂时无法使用
# if self.resume:
# self.load_checkpoint(load_optimizer=True)
# '''
# CosineAnnealingLR：若存在['initial_lr']，则从initial_lr开始衰减；
# 若不存在，则执行CosineAnnealingLR会在optimizer.param_groups中添加initial_lr键值，其值等于lr
# 重置初始学习率，在load_checkpoint中会加载优化器，但其中的initial_lr还是之前的，所以需要覆盖为self.lr，让其从self.lr衰减
# '''
# self.optimizer.param_groups[0]['initial_lr'] = self.lr
stage1_epoches = self.epoch_stage1 - self.start_epoch
lr_scheduler = optim.lr_scheduler.CosineAnnealingLR(self.optimizer, 30)
# lr_scheduler = optim.lr_scheduler.ReduceLROnPlateau(self.optimizer, mode='max', factor=0.2, patience=5, verbose=True,
# threshold=0.0001, threshold_mode='rel', cooldown=0, min_lr=0,
# eps=1e-08)
# 防止训练到一半暂停重新训练，日志被覆盖
global_step_before = self.start_epoch * len(self.train_loader)
for epoch in range(self.start_epoch, self.epoch_stage1):
epoch += 1
self.unet.train(True)
# 学习率重启
if epoch >= 18:
def set_bn_eval(m):
classname = m.__class__.__name__
if classname.find('BatchNorm') != -1:
m.eval()
self.unet.apply(set_bn_eval)
# if epoch == 25:
# self.optimizer.param_groups[0]['initial_lr'] = 0.00005
# lr_scheduler = optim.lr_scheduler.CosineAnnealingLR(self.optimizer, 25)
epoch_loss = 0
tbar = tqdm.tqdm(self.train_loader)
for i, (images, masks) in enumerate(tbar):
# GT : Ground Truth
images = images.to(self.device)
masks = masks.to(self.device)
# SR : Segmentation Result
net_output = self.unet(images)
net_output_flat = net_output.view(net_output.size(0), -1)
masks_flat = masks.view(masks.size(0), -1)
loss_set = self.criterion(net_output_flat, masks_flat)
try:
loss_num = len(loss_set)
except:
loss_num = 1
# 依据返回的损失个数分情况处理
if loss_num > 1:
for loss_index, loss_item in enumerate(loss_set):
if loss_index > 0:
loss_name = 'stage1_loss_%d' % loss_index
self.writer.add_scalar(loss_name, loss_item.item(),
global_step_before + i)
loss = loss_set[0]
else:
loss = loss_set
epoch_loss += loss.item()
# Backprop + optimize
self.reset_grad()
loss.backward()
self.optimizer.step()
params_groups_lr = str()
for group_ind, param_group in enumerate(
self.optimizer.param_groups):
params_groups_lr = params_groups_lr + 'params_group_%d' % (
group_ind) + ': %.12f, ' % (param_group['lr'])
# 保存到tensorboard，每一步存储一个
self.writer.add_scalar('Stage1_train_loss', loss.item(),
global_step_before + i)
descript = "Train Loss: %.7f, lr: %s" % (loss.item(),
params_groups_lr)
tbar.set_description(desc=descript)
# 更新global_step_before为下次迭代做准备
global_step_before += len(tbar)
# Print the log info
print('Finish Stage1 Epoch [%d/%d], Average Loss: %.7f' %
(epoch, self.epoch_stage1, epoch_loss / len(tbar)))
write_txt(
self.save_path,
'Finish Stage1 Epoch [%d/%d], Average Loss: %.7f' %
(epoch, self.epoch_stage1, epoch_loss / len(tbar)))
# 验证模型，保存权重，并保存日志
loss_mean, dice_mean = self.validation(stage=1)
if dice_mean > self.max_dice:
is_best = True
self.max_dice = dice_mean
else:
is_best = False
self.lr = lr_scheduler.get_lr()
state = {
'epoch': epoch,
'state_dict': self.unet.module.state_dict(),
'max_dice': self.max_dice,
'optimizer': self.optimizer.state_dict(),
'lr': self.lr
}
self.save_checkpoint(state, 1, index, is_best)
self.writer.add_scalars('Stage1_val_loss_dice', {
'val_loss': loss_mean,
'val_dice': dice_mean
}, epoch)
self.writer.add_scalar('Stage1_lr', self.lr[0], epoch)
# 学习率衰减
lr_scheduler.step()
def write_pb_txt(data_dir, idx, W, L, T, D, N, phi, alpha, beta):
#B = generate_lda(T, W, D, N, phi, alpha)
B,phi = gen_lda2(T, W, D, N, alpha)
write_pb(data_dir, idx, W, T, D, alpha, beta, B)
write_txt(data_dir, idx, B)
def write_data(data_dir, idx, W, L, T, D, N, phi, alpha, beta, chains):
B = generate_lda(T, W, D, N, phi, alpha)
write_pb(data_dir, idx, W, T, D, alpha, beta, B, write_params=False)
write_stan(data_dir, idx, W, T, D, alpha, beta, B, chains=chains, write_params=False)
write_prism(data_dir, idx, W, T, D, alpha, beta, B, write_params=False)
write_txt(data_dir, idx, B, T, alpha, beta, write_params=False)
def save_setting(dc):
"""保存设置到json文件
bool 类型用字符串 1 或 0来表示
"""
s = json.dumps(dc, ensure_ascii=False, indent=2)
utils.write_txt(setting_file, s)
