slf12 / grlmodel Goto Github PK

the step prepare data using spindle net to generate the regions of body part, i confused about that how to realize it ?
please tell me how to using spindlenet to generate the format you mentioned!!
thanks!

Can you provide the dataset download address? like Baidu cloud?
Because I've read the code and found out that you've been manipulating the dataset.
think you.

您好，我看到在测试代码里默认的阈值设置为0.5：softmax_or_not(predict, is_softmax=False, threshold=[0.5]*num_classes)。请问论文里在PETA数据集上的测得的各个指标的性能都是基于这个默认阈值吗？还是说对于每一个属性分类都在最佳阈值下测试？谢谢

Hi, @slf12 , I'm interested in your work, and I noticed that these scripts are to predict a dataset and calculate the accuracy instead of one image. Could you write some scripts to show how to predict one (or several) image(s)? (Sorry, I'm new to tensorflow...)

Hi , I have tried many times, and the trained, using the model before I trained is always AP=0.5 ,but I cana achieve a high AP by using your model .Why? Do you have special trick?

请问代码里各个属性的权重是跟预测属性的顺序一致的吗？我根据PETA数据集按照给定的预测顺序统计出来的正样本比例跟您的代码里不一致，然而按照另一个顺序计算出来的正样本比例跟您给出的权重顺序比较一致，如下：
1、论文给定的最优预测顺序：
['personalMale', 'personalLess30', 'personalLess45', 'personalLess60', 'personalLarger60','hairLong', 'accessoryMuffler', 'accessorySunglasses', 'accessoryHat', 'accessoryNothing', 'upperBodyShortSleeve', 'upperBodyThinStripes', 'upperBodyTshirt', 'upperBodyOther', 'upperBodyVNeck', 'upperBodyLogo', 'upperBodyPlaid', 'upperBodyJacket', 'upperBodyFormal', 'upperBodyCasual', 'lowerBodyTrousers', 'lowerBodyShortSkirt', 'lowerBodyJeans', 'lowerBodyShorts', 'lowerBodyCasual', 'lowerBodyFormal', 'footwearSneakers', 'footwearLeatherShoes', 'footwearSandals', 'footwearShoes', 'carryingBackpack', 'carryingMessengerBag', 'carryingNothing', 'carryingPlasticBags', 'carryingOther']
代码里给出的权重：[0.5004, 0.3231, 0.1031, 0.0639, 0.1984, 0.1976, 0.8607, 0.8520, 0.1380, 0.1343, 0.1016, 0.0700, 0.3058, 0.2945, 0.0395, 0.2394, 0.5511, 0.2932, 0.0816, 0.7465, 0.2757, 0.0265, 0.0774, 0.0199, 0.3627, 0.0336, 0.1440, 0.0443, 0.2198, 0.0163, 0.0317, 0.5201, 0.0842, 0.4526, 0.0136]；
实际统计出来的权重：
[0.5485, 0.5009, 0.3207, 0.1049, 0.0627, 0.2348, 0.0849, 0.0305, 0.0982, 0.7497, 0.1429, 0.0169, 0.0845, 0.4534,0.0119, 0.0389, 0.0269, 0.0697, 0.1325, 0.8535, 0.5172, 0.0474, 0.3055, 0.0348, 0.8643, 0.134, 0.2149, 0.292,0.0209, 0.3695, 0.1962, 0.2975, 0.2759, 0.0771, 0.1991]
2、另一种顺序：
['personalLess30', 'personalLess45', 'personalLess60', 'personalLarger60', 'carryingBackpack', 'carryingOther', 'lowerBodyCasual', 'upperBodyCasual', 'lowerBodyFormal', 'upperBodyFormal', 'accessoryHat', 'upperBodyJacket', 'lowerBodyJeans', 'footwearLeatherShoes', 'upperBodyLogo', 'hairLong', 'personalMale', 'carryingMessengerBag', 'accessoryMuffler', 'accessoryNothing', 'carryingNothing', 'upperBodyPlaid', 'carryingPlasticBags', 'footwearSandals', 'footwearShoes', 'lowerBodyShorts', 'upperBodyShortSleeve', 'lowerBodyShortSkirt', 'footwearSneakers', 'upperBodyThinStripes', 'accessorySunglasses', 'lowerBodyTrousers', 'upperBodyTshirt', 'upperBodyOther', 'upperBodyVNeck']；
计算出的权重：
[0.4914, 0.3342, 0.1016, 0.0617, 0.1932, 0.2058, 0.8586, 0.8501, 0.1402, 0.1368, 0.1017, 0.0694, 0.3027, 0.3057, 0.0393, 0.2362, 0.5457, 0.2933, 0.0836, 0.7464, 0.2753, 0.0257, 0.0774, 0.0188, 0.3605, 0.0354, 0.1396, 0.0454, 0.2101, 0.0163, 0.0306, 0.5188, 0.0822, 0.4579, 0.0125]
按照第二种顺序统计出来的结果跟您在代码中提供的结果基本一致，我根据原始数据集生成的标签文件地址为：https://drive.google.com/drive/folders/1g2CMkGHxh7Jekt2pVcHXlFkv9kM2ldp-?usp=sharing
其中名字中含'order'的文件为按照最优预测顺序生成的标签文件，另一个为按照非最优顺序生成的标签文件，可否帮忙确认一下标签是否有误吗？谢谢。

I had trained DeepMAR in PETA with ResNet50 as backbone, and my mA can achieve nearly 85% in pytorch. Have anyone trained baseline with InceptionV3 or ResNet50? The result in paper looks much lower than Resnet50, so I am not sure if it is my fault or author's.