reimplement with python3.6 and pytorch 1.0.1, Is this problem because of batch_size?

Hallo tekin,

I have just reimplement this project with python3.6 and pytorch1.0.1, and because of the limited computation of the GPU(GTX 970m, memory: 3GB) on my laptop. I changed the batch_size from 32 to only 2 and trained for a whole night(epoch:700 remains).

after that I test my model, but with results below:

-----------------------------------
  tensor to cuda : 0.000402
         predict : 0.004263
get_region_boxes : 0.063616
            eval : 0.009278
           total : 0.077559
-----------------------------------
2019-04-11 13:23:42 Results of ape
2019-04-11 13:23:42    Acc using 5 px 2D Projection = 0.00%
2019-04-11 13:23:42    Acc using 10% threshold - 0.0103 vx 3D Transformation = 0.00%
2019-04-11 13:23:42    Acc using 5 cm 5 degree metric = 0.00%
2019-04-11 13:23:42    Mean 2D pixel error is 2978.765381, Mean vertex error is 2.461958, mean corner error is 409.030334
2019-04-11 13:23:42    Translation error: 2.461311 m, angle error: 143.237230 degree, pixel error:  2978.765628 pix

all the Acc are Zero!!

I have checked the code, but everything is fine. Do you think it is because of batch_size is too small?? Or still because of the version of python and pytorch.

Looking forward to your reply!

[gaobaoding]

@btekin Hi, after change another GPU with 4GB memory and change the batch_size from 2 to 8. The ACC is not zero any more. It is really related with batch_size. I’m still a bit confused with this issue. So do you have any idea why the batch_size cannot be so small just like 2?

[btekin]

With very small batch sizes, the learning rate is also set to a higher value in the current implementation. See https://github.com/microsoft/singleshotpose/blob/master/train.py#L386 . Therefore, with small batch sizes you might want to adapt your learning rate as well for better convergence.