MobileOne#

The ResNet model is based on the “An Improved One millisecond Mobile Backbone” paper.

Architecture overview#

This architecture optimizes the model for inference speed at inference time on mobile device.

https://github.com/frgfm/Holocron/releases/download/v0.2.1/mobileone.png

The key takeaways from the paper are the following:

  • reuse the reparametrization concept of RepVGG while adding overparametrization in the block branches.

  • each block is composed of two consecutive reparametrizeable blocks (in a similar fashion than RepVGG): a depth-wise convolutional block, a point-wise convolutional block.

Model builders#

The following model builders can be used to instantiate a MobileOne model, with or without pre-trained weights. All the model builders internally rely on the holocron.models.classification.mobileone.MobileOne base class. Please refer to the source code for more details about this class.

mobileone_s0([pretrained, checkpoint, progress])

MobileOne-S0 from "An Improved One millisecond Mobile Backbone"

mobileone_s1([pretrained, checkpoint, progress])

MobileOne-S1 from "An Improved One millisecond Mobile Backbone"

mobileone_s2([pretrained, checkpoint, progress])

MobileOne-S2 from "An Improved One millisecond Mobile Backbone"

mobileone_s3([pretrained, checkpoint, progress])

MobileOne-S3 from "An Improved One millisecond Mobile Backbone"