Compound model scaling for neural networks

A computer-implemented method comprising:

• receiving a baseline neural network architecture comprising a convolutional neural network layer followed by a plurality of FusedConv neural network layers followed by a plurality of MBConv neural network layers;
• and
• generating an output neural network from the received baseline neural network architecture, comprising:
• linearly scaling a network width dimension, a network depth dimension, and a resolution dimension of the baseline neural network architecture based on a corresponding width coefficient, a depth coefficient, and a resolution coefficient, respectively, and based on a compound coefficient;
• wherein the output neural network is configured to run on one or more special-purpose hardware accelerators that are configured on a device.

One or more non-transitory computer-readable storage media storing instructions that, when executed by one or more computers, cause the one or more computers to perform operations comprising:

• receiving a baseline neural network architecture comprising a convolutional neural network layer followed by a plurality of FusedConv neural network layers followed by a plurality of MBConv neural network layers;
• and
• generating an output neural network from the received baseline neural network architecture, comprising:
• linearly scaling a network width dimension, a network depth dimension, and a resolution dimension of the baseline neural network architecture based on a corresponding width coefficient, a depth coefficient, and a resolution coefficient, respectively;
• wherein the output neural network is configured to run on one or more special-purpose hardware accelerators that are configured on a device.