The paper download address: https://arxiv.org/pdf/1904.04232.pdf
Source link: https://github.com/wyharveychen/CloserLookFewShot
Operating environment: Windows10
only reproduces the default source code. That is, the baseline trained on the CUB data set.
The main parameters:
parser = argparse.ArgumentParser(description= 'few-shot script %s' %(script))
parser.add_argument('--dataset' , default='CUB', help='CUB/miniImagenet/cross/omniglot/cross_char')
parser.add_argument('--model' , default='Conv4', help='model: Conv{4|6} / ResNet{10|18|34|50|101}') # 50 and 101 are not used in the paper
parser.add_argument('--method' , default='baseline', help='baseline/baseline++/protonet/matchingnet/relationnet{_softmax}/maml{_approx}') #relationnet_softmax replace L2 norm with softmax to expedite training, maml_approx use first-order approximation in the gradient for efficiency
parser.add_argument('--train_n_way' , default=5, type=int, help='class num to classify for training') #baseline and baseline++ would ignore this parameter
parser.add_argument('--test_n_way' , default=5, type=int, help='class num to classify for testing (validation) ') #baseline and baseline++ only use this parameter in finetuning
parser.add_argument('--n_shot' , default=5, type=int, help='number of labeled data in each class, same as n_support') #baseline and baseline++ only use this parameter in finetuning
parser.add_argument('--train_aug' , action='store_true', help='perform data augmentation or not during training ') #still required for save_features.py and test.py to find the model path correctly
parser.add_argument('--num_classes' , default=200, type=int, help='total number of classes in softmax, only used in baseline') #make it larger than the maximum label value in base class
parser.add_argument('--save_freq' , default=50, type=int, help='Save frequency')
parser.add_argument('--start_epoch' , default=0, type=int,help ='Starting epoch')
parser.add_argument('--stop_epoch' , default=-1, type=int, help ='Stopping epoch') #for meta-learning methods, each epoch contains 100 episodes. The default epoch number is dataset dependent. See train.py
parser.add_argument('--resume' , action='store_true', help='continue from previous trained model with largest epoch')
parser.add_argument('--warmup' , action='store_true', help='continue from baseline, neglected if resume is true') #never used in the paper
image_size = 224
optimization = 'Adam'
Problems encountered and solutions:
1. CUB data set
There is no data set, you need to download the data set.
The download link given by the source code: http://www.vision.caltech.edu/visipedia-data/CUB-200-2011/CUB_200_2011.tgz You
need to access the external network. At that time, I didn’t find resources, so I bought a vpn for 15 yuan. , You can visit the website later, but it says that there are too many visitors or the file is too large, so download is not allowed. Fifteen yuan has been lost.
Then find resources,
Baidu Netdisk download link: https://pan.baidu.com/s/1OL3s7XmzoaYmbBocYJjYyQ
Extraction code: 7jeb
2. Code modification
"""把反斜杠改为斜杠"""
data_path = join(cwd,'CUB_200_2011\images')
"""读json文件出现的问题，加入斜杠转义字符，让原来的斜杠存在真实意义"""
with open(data_file, encoding='utf-8') as f:
s = ""
for i in f:
s += i
s = s.replace('\\', '\\\\')
self.meta = json.loads(s)
"""
本地没有cuda，需要修改model.cuda(), 凡 **.cuda() 的都需要改成 **.to(device)
当然修改方式多种多样，本人秉持能就行的态度，这个没太大必要深究
修改如下：
"""
device = ('cuda' if torch.cuda.is_available() else 'cpu')
model = model.to(device)
"""
baselinetrain.py 中的 for i, (x, y) in enumerate(train_loader) 不能迭代
修改方法：把datamgr.py中的SimpleDataManager类中get_data_loader函数的data_loader_params参数的内核工作数改为0
即：num_workers = 0
"""
The final test set result (there is an error with the paper result, which may be caused by the division of training and test data):
