文档

Supported Framework Types

Table 1 lists the original framework types, input data types, and model files supported by model conversion and tuning.

Table 1 Supported framework types

Original Framework Type

Input Data Type

Model File Description

Caffe
  • FP32
  • FP16: implemented by setting the input option --input_fp16_nodes.
  • UINT8: implemented by configuring data preprocessing.
    NOTE:

    The input can be up to 4-dimensional. Operators involving dimension changes (such as reshape and expanddim) cannot output five dimensions.
  • Model file: xxx.prototxt
  • Weight file: xxx.caffemodel

The op_name and op_type in the model file must be the same as those in the weight file (case sensitive).

TensorFlow
  • FP16
  • FP32
  • UINT8
  • INT32
  • INT64
  • BOOL
    NOTE:

    The output data type cannot be INT64. You need to change the data type from INT64 to INT32.

Model file: xxx.pb

Only .pb models in FrozenGraphDef format can be converted and tuned.

ONNX
  • FP32
  • FP16: implemented by setting the input option --input_fp16_nodes.
  • UINT8: implemented by configuring data preprocessing.

Model file: xxx.onnx

MindSpore
  • FP32
  • UINT8: implemented by configuring data preprocessing.

Model file: xxx.air