Fix mobilenet_v2 GPU utilization. (#544)
Summary:
# mobilenet_v2
### Eval
## batch size analysis
best batch size is 16
<google-sheets-html-origin>
Batch Size | GPU Time | CPU Dispatch Time | Walltime | GPU Delta
-- | -- | -- | -- | --
1 | 7.973 | 7.906 | 7.986 | -
2 | 7.578 | 7.515 | 7.587 | -0.04954220494
4 | 7.678 | 7.455 | 7.682 | 0.01319609396
8 | 11.774 | 9.156 | 11.778 | 0.5334722584
16 | 21.273 | 9.116 | 21.288 | 0.8067776457
32 | 40.116 | 7.314 | 40.12 | 0.8857706952
64 | 78.941 | 9.311 | 78.96 | 0.9678183269
128 | 155.825 | 8.057 | 155.838 | 0.9739425647
### non-idleness analysis (bs=16)
## Train
### batch size analysis
best batch size is 32
<google-sheets-html-origin>
Batch Size | GPU Time | CPU Dispatch Time | Walltime | GPU Delta
-- | -- | -- | -- | --
1 | 90.233 | 87.626 | 90.235 | -
2 | 105.711 | 104.197 | 105.714 | 0.1715336961
4 | 118.667 | 117.172 | 118.669 | 0.1225605661
8 | 152.167 | 149.58 | 152.168 | 0.2823025778
16 | 238.466 | 237.094 | 238.468 | 0.5671334783
32 | 423.704 | 421.364 | 423.699 | 0.7767899826
64 | 803.82 | 801.466 | 803.808 | 0.8971262957
128 | 1562.212 | 1559.868 | 1562.179 | 0.9434848598
### non-idleness analysis (bs=32)
Pull Request resolved: https://github.com/pytorch/benchmark/pull/544
Reviewed By: aaronenyeshi
Differential Revision: D32147757
Pulled By: xuzhao9
fbshipit-source-id: 663d91421c7aca0ff263523b877c93839d575db0
