Building train function for AL

Author: nducthang

activelearning/config.py renamed to AL_config.py

@@ -3,12 +3,10 @@
 unlabeled = 'data/unlabeled'
 labeled = 'data/labeled'
 num_select = 8
-info_predict_path = 'activelearning/predict.txt'
-mode_active = 'sum' # 'sum or mean
 
 # Config general
 project = "/media/thang/New Volume/Active-learning-for-object-detection/"
-weight = 'activelearning/yolov5s.pt'
+weight = 'yolov5s.pt'
 device = '0' # cpu or 0,1,...
 name = 'gun'
 exist_ok = 1

AL_detect.py

@@ -4,7 +4,7 @@
 from utils.datasets import LoadImages
 from utils.general import check_img_size, non_max_suppression, scale_coords, xyxy2xywh, set_logging
 from utils.torch_utils import select_device, time_synchronized
-import activelearning.config as config
+import AL_config as config
 import time
 
 def AL_detect(opt):

AL_run.py

@@ -18,19 +18,53 @@
 
 """
 
+from torch.cuda.memory import reset_accumulated_memory_stats
 from AL_yolov5 import Yolov5
-import activelearning.select_function as select
-import activelearning.config as config
+import AL_config as config
 import glob
 import os
 from shutil import copyfile, move
 import io
 import copy
+def RandomSelect(num_select, result):
+ pass
+
+def UncertaintySamplingBinary(num_select, result, typ):
+ """
+ result = 
+ {"<link ảnh>": 
+ [
+ {"class": cls.item(), "box": [x,y,w,h], "conf": conf.item(),
+ ...
+ ],
+ ...
+ }
+ """
+ probas = {}
+ if typ == 'sum':
+ for item, lst_dic in result.items():
+ conf = 0
+ for dic in lst_dic:
+ conf += (1.0 - dic["conf"])
+ probas[item] = conf
+ elif typ == 'avg':
+ for item, lst_dic in result.items():
+ conf = 0
+ for dic in lst_dic:
+ conf += (1.0 - dic["conf"])
+ probas[item] = conf/len(lst_dic)
+ elif typ == 'max':
+ for item, lst_dic in result.items():
+ conf = 0
+ for dic in lst_dic:
+ conf = max(conf, 1.0 - dic["conf"])
+ probas[item] = conf
+ return sorted(probas, key=probas.get, reverse=True)[:num_select]
+
 
 class ActiveLearning(object):
- def __init__(self, model, select_function):
+ def __init__(self, model):
  self.model = model
- self.select_function = select_function
  self.num_select = config.num_select
  self.type = 'sum' # 'avg' , 'max', 'sum'
 
@@ -41,46 +75,32 @@ def run(self):
  while queried < config.max_queried:
  # Dự đoán các ảnh trong tập unlabeled
  result = self.model.detect()
- # Tổng hợp kết quả
- probas = {str(file.split('/')[-1]): 0.0 for file in glob.glob(config.source + '/*')}
- num_object = probas.copy()
-
- with open(config.info_predict_path, 'r') as f:
- for line in f.readlines():
- *cxywh, prob, file_name = line.split(',')
- file_name = file_name[:-1]
- probas[file_name] += 1.0 - float(prob)
- num_object[file_name] += 1
- 
- if config.mode_active == 'mean':
- for key, value in probas.items():
- probas[key] = value/num_object[key]
- 
- # Chọn ra k samples
- if len(probas) >= self.num_select:
- U_best = self.select_function.select(self.num_select, probas)
- 
- # Gán nhãn cho U_best
- """ GIẢ SỬ ĐOẠN NÀY LÀ NGƯỜI GÁN """
- # Tạo ra các file label cho U_best vào thư mục labeled
+
+ # Chon ra k ảnh có score cao nhất
+ # Sử dụng lấy mẫu không chắc chắn
+ if len(result) >= self.num_select:
+ U_best = UncertaintySamplingBinary(self.num_select, result, 'sum')
+ print(U_best)
+ 
+ # Gán nhãn cho các file trong samples (Người tương tác)
+
+ # Duyệt tất cả các file được chọn
  for f in U_best:
  # Chuyển file ảnh vào thự mục labeled
- move(os.path.join(config.unlabeled, f), os.path.join(config.labeled, f))
+ move(f, f.replace("unlabeled", "labeled"))
  # Tạo file nhãn vào thư mục labeled
- file_name = '.'.join(f.split('.')[:-1])
- # print(f)
- # source = open(os.path.join('data/gun/',f),"r",encoding='utf-8').readlines()
- # dect = open(os.path.join(config.labeled, file_name + '.txt', "w", encoding='utf-8').write(copy.copy(source)))
- copyfile(os.path.join('data/gun/', file_name+'.txt'), os.path.join(config.labeled, file_name + '.txt'))
-
- # Cập nhật file train.txt
- with open('data/train.txt', 'w') as f:
- for fn in U_best:
- f.write(config.project + config.labeled + "/" + fn + "\n")
- 
+ type_file = f.split('.')[-1]
+ copyfile(f.replace("unlabeled","gun").replace(type_file, 'txt'), f.replace("unlabeled", "labeled").replace(type_file, 'txt'))
+
+ # thêm danh sách file đã gán nhãn vào dữ liệu train
+ with open('data/train.txt', "a") as f:
+ for file_name in U_best:
+ f.write(file_name.replace("unlabeled","labeled") + '\n')
+
  # Train model
  self.model.train()
 
+ ####################### LOADING ########################
  # Xoá file weight cũ
  if os.path.exists(config.weight):
  os.remove(config.weight)
@@ -95,7 +115,5 @@ def run(self):
 
 
 if __name__ == '__main__':
- bot = ActiveLearning(model=Yolov5(), select_function=select.RandomSelect())
- bot.run()
- # model = Yolov5()
- # model.train()
+ bot = ActiveLearning(model=Yolov5())
+ bot.run()