CVPR 2018 的一篇少样本学习论文
Learning to Compare: Relation Network for Few-Shot Learning
源码地址：https://github.com/floodsung/LearningToCompare_FSL
在自己的破笔记本上跑了下这个源码，windows 系统，pycharm + Anaconda3 + pytorch-cpu 1.0.1
报了一堆bug, 总结如下：
procs_images.py里 ‘cp’报错
用procs_images.py处理 miniImangenet 数据集的时候：
报错信息：
/LearningToCompare_FSL-master/datas/miniImagenet/proc_images.py
'cp' �����ڲ����ⲿ���Ҳ���ǿ����еij������������ļ���
具体位置是
/datas/miniImagenet/procs_images.py Line 48:
os.system('cp images/' + image_name + ' ' + cur_dir)
这个‘cp’是linux环境运行的。
用windows系统的话要改成：
os.rename('images/' + image_name, cur_dir + image_name)
除此之外，所有的 os.system('mkdir ' + filename)
也要改成 os.mkdir(filename)，虽然不一定会报错。
cpu RuntimeError: Attempting to deserialize object on a CUDA device but torch.cuda.is_available() is False.
我的torch版本是是cpu, 所以把所有 .cuda(GPU)删了，另外
使用torch.load时添加 ，map_location ='cpu'
以miniImagenet_train_few_shots.py 为例
Line 150:
feature_encoder.load_state_dict(torch.load(str("./models/omniglot_feature_encoder_" + str(CLASS_NUM) +"way_" + str(SAMPLE_NUM_PER_CLASS) +"shot.pkl")))
改成
feature_encoder.load_state_dict(torch.load(str("./models/omniglot_feature_encoder_" + str(CLASS_NUM) +"way_" + str(SAMPLE_NUM_PER_CLASS) +"shot.pkl"),map_location = 'cpu'
))
Line:153:
relation_network.load_state_dict(torch.load(str("./models/miniimagenet_relation_network_"+ str(CLASS_NUM) +"way_" + str(SAMPLE_NUM_PER_CLASS) +"shot.pkl")))
改成
relation_network.load_state_dict(torch.load(str("./models/miniimagenet_relation_network_"+ str(CLASS_NUM) +"way_" + str(SAMPLE_NUM_PER_CLASS) +"shot.pkl"),map_location = 'cpu'))
KeyError: '..\\datas\\omniglot_resized'
报错信息：
File "LearningToCompare_FSL-master/omniglot/omniglot_train_few_shot.py", line 163, in main
task = tg.OmniglotTask(metatrain_character_folders,CLASS_NUM,SAMPLE_NUM_PER_CLASS,BATCH_NUM_PER_CLASS)
File "LearningToCompare_FSL-master\omniglot\task_generator.py", line 72, in
self.train_labels = [labels[self.get_class(x)] for x in self.train_roots]
KeyError: '..\\datas\\omniglot_resized'
关键的地方其实是在：
task_generator.py, line 74:
def get_class(self, sample):
return os.path.join(*sample.split('/')[:-1])
print (os.path.join(*sample.split('/')[:-1])) 结果是
..\datas\omniglot_resized
而labels是
{'../datas/omniglot_resized/Malay_(Jawi_-_Arabic)\\character25': 0, '../datas/omniglot_resized/Japanese_(hiragana)\\character15': 1, '…}
而 print(os.path.join(*sample.split('\\')[:-1])) 结果正是
../datas/omniglot_resized/Malay_(Jawi_-_Arabic)\character25
解决方法：把'/'改成'\\'即可
def get_class(self, sample):
return os.path.join(*sample.split('\\')[:-1])
RuntimeError: Expected object of scalar type Long but got scalar type Int for argument #3 'index'
报错信息：
File "/LearningToCompare_FSL-master/miniimagenet/miniimagenet_train_few_shot.py", line 193, in main
torch.zeros(BATCH_NUM_PER_CLASS * CLASS_NUM, CLASS_NUM).scatter_(1, batch_labels.view(-1, 1), 1))
RuntimeError: Expected object of scalar type Long but got scalar type Int for argument #3 'index'
解决方法：在前面加一句
batch_labels = batch_labels.long()
RuntimeError: Expected object of scalar type Long but got scalar type Int for argument #2 'other'
报错信息：
File "LearningToCompare_FSL-master/miniimagenet/miniimagenet_test_few_shot.py", line 247, in
rewards = [1 if predict_labels[j]==test_labels[j] else 0 for j in range(batch_size)]
RuntimeError: Expected object of scalar type Long but got scalar type Int for argument #2 'other'
解决方法：在前面加上
predict_labels = predict_labels.long()
test_labels = test_labels.long()
这两个好像是使用torch的数据格式问题
IndexError: invalid index of a 0-dim tensor. Use tensor.item() to convert a 0-dim tensor to a Python number
报错信息：
File "LearningToCompare_FSL-master/miniimagenet/miniimagenet_train_few_shot.py", line 212, in main
print("episode:",episode+1,"loss",loss.data[0])
IndexError: invalid index of a 0-dim tensor. Use tensor.item() to convert a 0-dim tensor to a Python number
按要求改成
print("episode:", episode + 1, "loss", loss.item())
就可以了
RuntimeError: output with shape [1, 28, 28] doesn't match the broadcast shape [3, 28, 28]
报错信息：
File "LearningToCompare_FSL-master\omniglot\task_generator.py", line 107, in __getitem__
image = self.transform(image)
File "...\Anaconda3\envs\python36\lib\site-packages\torchvision\transforms\transforms.py", line 60, in __call__
img = t(img)
File "...\Anaconda3\envs\python36\lib\site-packages\torchvision\transforms\transforms.py", line 163, in __call__
return F.normalize(tensor, self.mean, self.std, self.inplace)
File "...\Anaconda3\envs\python36\lib\site-packages\torchvision\transforms\functional.py", line 208, in normalize
tensor.sub_(mean[:, None, None]).div_(std[:, None, None])
RuntimeError: output with shape [1, 28, 28] doesn't match the broadcast shape [3, 28, 28]
这个是使用Omniglot数据集时的报错，主要原因在于
"\omniglot\task_generator.py", line 139:
def get_data_loader(task, num_per_class=1, split='train',shuffle=True,rotation=0):
normalize = transforms.Normalize(mean=[0.92206, 0.92206, 0.92206], std=[0.08426, 0.08426, 0.08426])
dataset = Omniglot(task,split=split,transform=transforms.Compose([Rotate(rotation),transforms.ToTensor(),normalize]))
使用 torch.transforms 中 normalize 用了 3 通道，而实际使用的数据集Omniglot 图片大小是 [1, 28, 28]
解决方法：
把
normalize = transforms.Normalize(mean=[0.92206, 0.92206, 0.92206], std=[0.08426, 0.08426, 0.08426])
改成
normalize = transforms.Normalize(mean=[0.92206], std=[0.08426])
UserWarning: nn.functional.sigmoid is deprecated.
类似的warning 还有
UserWarning : torch.nn.utils.clip_grad_norm is now deprecated in favor of torch.nn.utils.clip_grad_norm_.
按要求改就行
torch.nn.utils.clip_grad_norm(feature_encoder.parameters(), 0.5)
改成
torch.nn.utils.clip_grad_norm_(feature_encoder.parameters(), 0.5)
def forward里的
out = F.sigmoid(self.fc2(out))
改成
out = F.torch.sigmoid(self.fc2(out))