Optimization Tips

The model needs to be int8/uint8 quantized. Operators that take in other types are not supported on the DLPU.

Tensor sizes must be constant at compile time. Dynamic sizes are not supported.

All tensors must be 1-, 2-, or 3-dimensional. If a tensor has more than 3 dimensions, only the innermost 3 dimensions may have a size greater than 1. For example, shape [1, 128, 128, 3] is OK, but shape [3, 128, 128, 1] is not.

The Edge-TPU compiler cannot partition the model more than once, so if the model has an operator that can't be run on the DLPU and instead must run on the CPU, the graph will be split into two partitions where the unsupported operator and every operator after it runs on the CPU. Therefore, avoid unsupported operators, especially toward the beginning of the model.