Single-crop validation performance about places365 HOT 7 CLOSED

second this.

from places365.

Sure. Here is the accuracy on the validation directly from caffe training.
VGG16:
I0416 10:03:46.482185 2637 solver.cpp:404] Test net output #0: accuracy = 0.539833
I0416 10:03:46.482353 2637 solver.cpp:404] Test net output #1: loss = 1.73844 (* 1 = 1.73844 loss)

CaffNet:
I0407 01:58:51.494807 11471 solver.cpp:404] Test net output #0: accuracy = 0.51652
I0407 01:58:51.494874 11471 solver.cpp:404] Test net output #1: loss = 1.84729 (* 1 = 1.84729 loss)

ResNet152 finetuned:
I0601 23:04:56.067314 19006 solver.cpp:404] Test net output #0: accuracy = 0.55
I0601 23:04:56.067584 19006 solver.cpp:404] Test net output #1: loss = 1.76827 (* 1 = 1.76827 loss)

from places365.

@metalbubble something strange here.
by using resnet 152, the averaged score over 10 crops is 54.74/85.08%.
how can it be 55% with only single crop?

from places365.

Yes, indeed it is kind of weird to me. Maybe it is due to the fine-tuning of the resnet. I will manage to trian a resnet from scratch using torch and update the numbers.

from places365.

@metalbubble

I will manage to trian a resnet from scratch using torch and update the numbers.

Could you show the numbers for the ResNet152-places365 torch model you trained from scratch?
(I ported the Torch model to PyTorch and want to know if I ported it successfully.)

from places365.

Here is the final state of the training log:

| Test: [120][141/143] Time 0.134 Data 0.000 top1 48.047 ( 44.815) top5 13.281 ( 14.702)
| Test: [120][142/143] Time 0.121 Data 0.000 top1 42.578 ( 44.799) top5 16.797 ( 14.717)
| Test: [120][143/143] Time 0.121 Data 0.000 top1 48.047 ( 44.822) top5 15.234 ( 14.721)

• Finished epoch # 120 top1: 44.822 top5: 14.721

from places365.

Verify your pytorch model(alexnet),so big gap with your Caffe model. Here is the precision of the your pytorch model @zhoubolei

• Prec@1 47.551 1-crop
• Prec@1 49.151 10-crop

from places365.