explain the training output

[liangshi036]

Hi, , the output seems interesting.

at one epoch,the test output is :

Test: [ 0/68] Time: 11.811 (11.811) Loss: 0.1038 (0.1038) Acc@1: 99.3137 (99.3137)
Test: [ 50/68] Time: 0.224 (0.514) Loss: 0.2778 (0.3847) Acc@1: 93.3333 (89.9769)
Test: [ 68/68] Time: 3.108 (0.467) Loss: 0.3384 (0.3697) Acc@1: 90.4589 (90.2704)

at another epoch ,the test output:

Test: [ 0/68] Time: 11.161 (11.161) Loss: 0.1021 (0.1021) Acc@1: 99.5098 (99.5098)
Test: [ 50/68] Time: 0.181 (0.524) Loss: 0.2095 (0.4677) Acc@1: 95.5882 (88.1546)
Test: [ 68/68] Time: 0.527 (0.437) Loss: 0.3604 (0.4326) Acc@1: 90.3382 (88.7958)

my question is , the second one has higher Acc@1 than the first one at every step,

but it got lower accuracy output (90.2704 vs 88.7958 ) in the average?

[rwightman]

each of the step values is for 1 batch, so you see 3 batch values of 68, the avg over 68 can differ quite a bit from the sample of 3.