CCC-123
How can I get the results of paper?
Thanks for your excellent work to reproduce the original paper.Could you share your tricks for the improvements?
Such as:
1.How many epoch do you use in the train code,is it 1000?My computer run it very slow(i5-6400,GTX 1060,8G running memory),about 3 hours per 5 epoch.My trainning command is python train.py -m MNet
2.How to combine the model of AF1,AF2,AF3,HP and MNet?I can produce the model separately.In the paper we can learn about :
Initially,a plain M-net is trained to learn the fundamental pedestrian features. Then the M-net is duplicated three times to construct the AF-net with adjacent MDA modules...
How do the nine blocks in the paper show up in your program?
3.I wonder the 'test' or 'show' script only need single model can produce good results?
Female : 0.632039 AgeOver60 : 0.500000 Age18-60 : 0.499689 AgeLess18 : 0.499394 Front : 0.667503 Side : 0.626780 Back : 0.698474 Hat : 0.502970 Glasses : 0.528432 HandBag : 0.501559 ShoulderBag : 0.500210 Backpack : 0.504379 HoldObjectsInFront : 0.500000 ShortSleeve : 0.757856 LongSleeve : 0.757543 UpperStride : 0.524863 UpperLogo : 0.522764 UpperPlaid : 0.500000 UpperSplice : 0.539431 LowerStripe : 0.500000 LowerPattern : 0.500000 LongCoat : 0.500000 Trousers : 0.790902 Shorts : 0.647391 Skirt&Dress : 0.555429 boots : 0.500000 path: MNet/checkpoint_epoch_0 mA: 0.567600 ACC: 0.525607 Prec: 0.800878 Rec: 0.591070 F1: 0.680162 The above result is under the command of
python test.py -m MNet -p MNet/checkpoint_epoch_0,and the trainning epoch is just 5.
Do u get the paper results now? I get ma :0.68 now but still not reach the paper results.
@wzzwh110 How can u get mA :0.68? By train MNet for 1000 epoch or other methods? I just got mA :0.61 by train MNet for 280 epoch. Any help is appreciated. Thx!
