Open msieb1 opened 6 years ago
Try using a resnet50
backbone instead of resnet101
Other than that, similarly to training, inference is quite hardware dependant, hence the reason that Nvidia are building specific cards for inference such as the Tesla P4 and the Jetson module. If you are running an AWS/Google Cloud instance or something like that for inference, try increasing core count and make sure that Keras is using multiprocessing.
What are the hyperparameters to look at that have the most influence on decreasing inference time without affecting detection performance too drastically?