--input_format

Description

Sets the model input format.

See Also

None

Argument

Argument:

• If the original framework is Caffe, the supported formats include NCHW (default) and ND (any format with N ≤ 4).
• If the original framework is ONNX, the supported formats include NCHW (default), NCDHW, and ND (any format with N ≤ 4).
• If the original framework is TensorFlow, the supported formats include NCHW, NHWC (default), ND, NCDHW, and NDHWC.
  • For a TensorFlow model converted using the ONNX model conversion tool, this option is required and the value must be NCHW.
  • If the original model contains an operator with the argument data_format, this option is required and ND is recommended. During model conversion, the format is inferred based on the operator with data_format. If the input data format cannot be determined, ND is recommended.
• If the original framework is MindSpore, only NCHW is supported. Other values are invalid, but the model will be successfully converted.

Default: NCHW for MindSpore, Caffe, and ONNX; NHWC for TensorFlow

Restrictions:

• As model inference only accepts NHWC format data. If AIPP is enabled during model conversion, the input node associated with AIPP will be forced with the NHWC format, which may be different from that specified by --input_format in the atc command.
• If --insert_op_conf is also included, --input_format must be set to NCHW or NHWC.
• This option does not support the scenario where a model has multiple inputs of different data formats. That is, only a single value is supported. Using a TensorFlow model as an example, this option can only be set to NCHW, NHWC, ND, NCDHW, or NDHWC.

Suggestions and Benefits

None

Example

--input_format=NCHW

Applicability

Atlas 200/300/500 Inference Product

Atlas Training Series Product

Dependencies and Restrictions

None