higher performance on random,entropy and coreset methods compared to paper
ChenggangLu opened this issue · 15 comments
I try to reproduce random,entropy and coreset methods based on mmdetection. However, results are much higher compared to paper. Do you have any idea about this?
Thanks.
And I have tried 3 times with different init dataset, the result seems not change.
The results of all methods except MI-AOD using SSD is copied from Figure 6 in LL4AL and Figure 4 in CDAL. You can ask them for more reasons.
Okay. Thank you for your reply.
@ChenggangLu Would u mind sharing the scripts used to evaluate the entropy and coreset method? I am also working on similar issues, thanks!
The dataloader is initialized in the same way as author here
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
from mmdet.utils import get_root_logger
class SelectionMethod:
"""
Abstract base class for selection methods,
which allow to select a subset of indices from the pool set as the next batch to label for Batch Active Learning.
"""
def __init__(self):
super().__init__()
self.logger=get_root_logger()
def select(self, selection_size):
"""
Select selection_size elements from the pool set
(which is assumed to be given in the constructor of the corresponding subclass).
This method needs to be implemented by subclasses.
It is assumed that this method is only called once per object, since it can modify the state of the object.
Args:
selection_size (int): how much images selected in one cycle
Returns:
idxs_selected (np.ndarray): index of chosen images
"""
raise NotImplementedError()
class RandomSelectionMethod(SelectionMethod):
def __init__(self,idxs_unlabeled):
super().__init__()
self.idxs_unlabeled=idxs_unlabeled
def select(self, selection_size):
np.random.shuffle(self.idxs_unlabeled)
return self.idxs_unlabeled[:selection_size]
class EntropySelectionMethod(SelectionMethod):
def __init__(self,data_loader,idxs_unlabeled,model):
super().__init__()
self.data_loader=data_loader
self.idxs_unlabeled=idxs_unlabeled
self.model=model
def select(self, selection_size):
self.logger.info(f'------ computing image entropy ------')
entropy=torch.zeros(len(self.idxs_unlabeled)).cuda()
for i,data in enumerate(self.data_loader):
with torch.no_grad():
if i in self.idxs_unlabeled:
idx=np.where(i==self.idxs_unlabeled)[0][0]
img=data['img'][0].cuda()
# calculate entropy
feat = self.model.extract_feat(img)
cls_scores, _ = self.model.bbox_head(feat)
all_cls_scores = torch.cat([
s.permute(0, 2, 3, 1).reshape(
s.shape[0], -1, self.model.bbox_head.cls_out_channels) for s in cls_scores
], 1)
all_cls_scores=F.softmax(all_cls_scores,dim=2)
entropy[idx]=torch.sum(all_cls_scores*torch.log(all_cls_scores),dim=2).mean()
if i%500==0:
self.logger.info(f'------ {i}/{len(self.data_loader.dataset)} ------')
arg=entropy.argsort().cpu().numpy()
return self.idxs_unlabeled[arg[:selection_size]]
class CoresetSelectionMethod(SelectionMethod):
def __init__(self,data_loader,idxs_labeled,idxs_unlabeled,model):
super().__init__()
self.data_loader=data_loader
self.idxs_labeled=idxs_labeled
self.idxs_unlabeled=idxs_unlabeled
self.model=model
def select(self, selection_size):
self.logger.info(f'------ computing image feat ------')
feats=compute_feat(self.data_loader,self.model)
feats_labeled=feats[self.idxs_labeled]
feats_unlabeled=feats[self.idxs_unlabeled]
self.logger.info(f'------ computing distances ------')
dist=torch.cdist(feats_unlabeled,feats_labeled,p=2).min(dim=1)[0]
self.logger.info(f'------ choose images ------')
idxs_chosen=[]
for i in range (selection_size):
idx=dist.argmax()
idxs_chosen.append(self.idxs_unlabeled[idx])
new_dist=torch.cdist(feats_unlabeled,feats_unlabeled[idx].unsqueeze(0),p=2).squeeze()
dist=torch.where(new_dist<dist,new_dist,dist)
if i%20==0:
self.logger.info(f'------ {i}/{selection_size} ------')
return np.array(idxs_chosen)
def compute_feat(data_loader,model):
logger=get_root_logger()
feats=torch.zeros([len(data_loader.dataset),1024]).cuda()
for i,data in enumerate(data_loader):
with torch.no_grad():
img=data['img'][0].cuda()
feat = model.extract_feat(img)
# feat shape [1024]
feats[i]=nn.AdaptiveAvgPool2d((1, 1))(feat[1]).view(-1)
if i%500==0:
logger.info(f'------ {i}/{len(data_loader.dataset)} ------')
return feats
@ChenggangLu Would u mind sharing the scripts used to evaluate the entropy and coreset method? I am also working on similar issues, thanks!
I use the same settings in Learning Loss.
I find out the reason. The data augmentation used in mmdetection is a little different from SSD in LL4AL. It seems that SSD in LL4AL miss sampling a patch the minimum Jaccard overlap with the objects is 0.5.
When i change the settings in mmdetection to be the same as SSD in LL4AL, the results are almost the same.
OK, thanks for your research.
Hi, did you rerun the SSD with Pascal VOC? How about your results? My results are lower than the paper,but my log file is same to the author'.
Hi, did you rerun the SSD with Pascal VOC? How about your results? My results are lower than the paper,but my log file is same to the author'.
Do you mean use MIAOD to select samples to train SSD with PASCAL VOC? I haven't tried yet.
Hi, did you rerun the SSD with Pascal VOC? How about your results? My results are lower than the paper,but my log file is same to the author'.
Do you mean use MIAOD to select samples to train SSD with PASCAL VOC? I haven't tried yet.
yes,i train SSD with VOC,but mAP is lower than the paper
@jiahao12121 Ok. I will try to use MIAOD and reply you later after getting results.
@jiahao12121 Ok. I will try to use MIAOD and reply you later after getting results.
Thank you! Recently i'm studying the code, i have some valuable questions, may i discuss with you? would you mind leaving a mail?
@jiahao12121 Ok. I will try to use MIAOD and reply you later after getting results.
Thank you! Recently i'm studying the code, i have some valuable questions, may i discuss with you? would you mind leaving a mail?
My email address is luchenggang@zju.edu.cn
@ChenggangLu Thanks for sharing the code and sorry for the late reply. For your original question, personally I think it might be due to the use of pretrained models.