Worse performance in softmax_loss.py with inception model

[OysterEleven]

Hi, I'm trying to reproduce your benchmark on Market1501. For now, I successfully accomplished the Resnet50 both triplet and softmax, but when using inception module with softmax loss, I always get 10 percent lower than benchmark(Mean AP : 0.412233532517 Rank1 accuracy : 0.65053444). By the way, I'm using my own way to read data, but I think it's fine, because it worked well with Resnet50 , My training parameters are like below:

args {   
          weight_decay: 0.0005
          is_cuda: True
          workers: 4
          batch_size: 256
          height: 144
          epochs: 150
          epoch_step: 75 
          width: 56
          seed: 1
          lr: 0.1(also test with 0.01)
          log_interval: 10
          gpus: 2,3
          val_interval: 50
          dataset: market1501
          features: 128
  }  
pytorch version: 0.3.1
python version: 2.7.12

Any suggestion?

[liangbh6]

@OysterEleven Have you figured out why? I even get ~40% rank1. It is a little confusing.

[liangbh6]

@OysterEleven And I find that however I use 224x224 or 299x299 or 256x256, the size of the layer before classifier is 1536...

[liangbh6]

Well, I check out several versions of GoogLeNet and find the Inception here cannot match any of them.