machine-learningneural-networkimagenetmobilenetimage-classification

Size of mobilenet's training data (ImageNet)

Is there any information available on how many images/classes from ImageNet were used to train the MobileNet architecture? I have read the paper "MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications" but there's no such information available over there.

Any leads?

Solution

  • Well, I think the training data configurations change by changing comparison. At Table 8, they compare mobilenet architecture with GoogLeNet and VGG 16 where both of them are submitted to ILSVRC 2014 competition, so in that comparison I assume that they use the same training data as 1.2m training, 50k validation and 150k test set.

    And, at Table 9, they compare MobileNet with AlexNet and SqueezeNet, where both of them uses 1.2m training, 50k validation and 150k test set as well (ILSVRC 2012 challenge)

    You can find more info about the challenge and dataset info on ILSVRC 2014 and ILSVRC 2012.