Fig 1.
Inverted residual block structure of MobileNetV2.
Fig 2.
PRNET [6].
Fig 3.
Data parallelism.
Fig 4.
Model parallelism.
Fig 5.
Cropping and aligning image.
Fig 6.
Generating depth image from RGB image.
Fig 7.
Example of horizontal flipping.
Fig 8.
RGBD input layer.
Fig 9.
RGB+D input layer.
Fig 10.
Example of extending the block (1).
Fig 11.
Example of extending the block (2).
Fig 12.
Example of weight transferring.
Fig 13.
Example of the workflow of automatic model finding on distributed training.
Fig 14.
Example of automatic model finding on distributed training.
Fig 15.
Example of the workflow of the automatic model finding on concurrent training.
Table 1.
Performance comparison between each model condition.
Fig 16.
Result of automatic model finding on distributed training.
Fig 17.
Result of the best layer replication positions.
Fig 18.
Results of the best layer replication position in Server1.
Fig 19.
Results of the best layer replication position in Server2.
Table 2.
Performance comparison between distributed training and concurrent training.