ownstyledu / translate-to-recognize-networks Goto Github PK

How to know the division of sun rgbd for scene recognition

我将use_fake设置为True以后，会报如下错误：
TypeError: define_TrecgNet() got an unexpected keyword argument 'use_noise'
于是将trecg_model.py中的代码：
sample_model = networks.define_TrecgNet(cfg_sample, use_noise=not self.use_noise, upsample=True, device=self.device)
改为：
sample_model = networks.define_TrecgNet(cfg_sample, upsample=True, device=self.device)
所以，use_noise这个参数不能使用吗？谢谢。

您好，您的工作非常棒，感谢分享代码。
在使用train_fusion.py 训练fusion model 的时候，发现fusion.py中的total_epoch没有定义，在将其设置为100后，模型可以开始训练啦，但是结果很差，而且learning rate一直都是负的：

########## Class Acc Report ##########
acc bathroom: 0.000
acc bedroom: 0.000
acc classroom: 100.000
acc computer_room: 0.000
acc conference_room: 0.000
acc corridor: 0.000
acc dining_area: 0.000
acc dining_room: 0.000
acc discussion_area: 0.000
acc furniture_store: 0.000
acc home_office: 0.000
acc kitchen: 0.000
acc lab: 0.000
acc lecture_theatre: 0.000
acc library: 0.000
acc living_room: 0.000
acc office: 0.000
acc rest_space: 0.000
acc study_space: 0.000
##############################
mean_acc: [5.263157894736842]
Mean Acc Epoch <18> * Prec@1 <5.263>
End of iter 2196 / 100 Time Taken: 54.42968773841858 sec

default lr 0.0002
/////////learning rate = -0.0052375

在fusion model训练之前，已经用train.py分别训练了两个tregnets：trecg_AtoB_best.pth、trecg_BtoA_best.pth，并在resner_sunrgbd_config.py中设置了resume_path_A和resume_path_B。Mean Acc 分别是50.05和46.02，结果应该还算正常。

以下是user config：

GPU_IDS : 2,3
nTHREADS : 8
WORKERS : 8
MODEL : fusion
ARCH : resnet18
PRETRAINED : place
CONTENT_PRETRAINED : place
NO_UPSAMPLE : False
FIX_GRAD : False
IN_CONC : False
DATA_DIR_TRAIN : /home/cfang/Downloads/dataset/RGB-D/OFFICIAL_SUNRGBD/data_in_class/conc_data/train
DATA_DIR_VAL : /home/cfang/Downloads/dataset/RGB-D/OFFICIAL_SUNRGBD/data_in_class/conc_data/val
DATA_DIR_UNLABELED : /home/cfang/Downloads/dataset/nyud2/mix/conc_data/10k_conc_bak
SAMPLE_MODEL_PATH : None
CHECKPOINTS_DIR : ./checkpoints
ROOT_DIR : /home/cfang/works/RGBD/Translate-to-Recognize-Networks-master/
SUMMARY_DIR_ROOT : /home/cfang/works/RGBD/Translate-to-Recognize-Networks-master/summary/
LOG_PATH : summary
CONTENT_MODEL_PATH : /home/cfang/works/RGBD/Translate-to-Recognize-Networks-master/resnet18_places365.pth.tar
DATA_TYPE : pair
WHICH_DIRECTION : AtoB
NUM_CLASSES : 19
BATCH_SIZE : 40
LOAD_SIZE : 256
FINE_SIZE : 224
FLIP : True
UNLABELED : False
FIVE_CROP : False
FAKE_DATA_RATE : 0.3
LR : 0.0002
WEIGHT_DECAY : 0.0001
MOMENTUM : 0.9
LR_POLICY : lambda
PHASE : train
RESUME : False
RESUME_PATH : None
RESUME_PATH_A : /home/cfang/works/RGBD/Translate-to-Recognize-Networks-master/checkpoints/trecg/2019_09_11_10_59_55/trecg_BtoA_best.pth
RESUME_PATH_B : /home/cfang/works/RGBD/Translate-to-Recognize-Networks-master/checkpoints/trecg/2019_09_11_10_47_03/trecg_AtoB_best.pth
NO_FC : True
INIT_EPOCH : True
START_EPOCH : 1
ROUND : 1
MANUAL_SEED : 8790
NITER : 20
NITER_DECAY : 80
NITER_TOTAL : 100
LOSS_TYPES : ['CLS', 'SEMANTIC']
EVALUATE : True
USE_FAKE_DATA : False
CLASS_WEIGHTS_TRAIN : None
PRINT_FREQ : 100
NO_VIS : False
CAL_LOSS : True
SAVE_BEST : True
INFERENCE : False
ALPHA_CLS : 1
WHICH_CONTENT_NET : resnet18
CONTENT_LAYERS : 0,1,2,3,4
NITER_START_CONTENT : 1
NITER_END_CONTENT : 200
ALPHA_CONTENT : 10
NO_LSGAN : True
NITER_START_GAN : 1
NITER_END_GAN : 200
ALPHA_GAN : 1
NITER_START_PIX2PIX : 1
NITER_END_PIX2PIX : 200
ALPHA_PIX2PIX : 5
parse : <bound method DefaultConfig.parse of <config.default_config.DefaultConfig object at 0x7f5a832f25c0>>
device_ids: 2
fusion

请教下哪里设置得不对，万分感谢！🙏

Hi,

I have a question related to the weight estimation for handle the imbalance training. In your paper the following formula is showed: w(y) = (N(y) -N(c_min) + lambda)/(N(c_max) - N(c_min)). However as far as I understand this formula is giving a high weight to classes that have more representation in the dataset and a low weight to the classes that have less representation. Is it a mistake? Can be that the numerator and denominator are switched?

hello ,when i want to run the evalute.py,it's report has no attrbuite "set_data_loader",then,i click into "TRecgNet_Upsample_Resiual",i am also cann't find .could you tell me how to solve it?(我是**的学生)

感谢您的分享，您提供的数据集网页无法进入，能否提供类似于百度网盘的连接呢

hello ， i use resnet18 to train the split rgb image dataset directly， the result is 57.43%,even higer than use your way,is there me have problem? Or differently with you?