--framework

Applicability

Description

Sets the framework of the original network model.

Atlas A3 training products/Atlas A3 inference products: The Caffe framework is not supported.

Atlas A2 training products/Atlas A2 inference products: The Caffe framework is not supported.

See Also

None

Argument

Argument:

• 0: Caffe
• 1: Model file in .air format of the MindSpore framework or standard .air file exported from TorchAir through export
• 3: TensorFlow
• 5: ONNX

Restrictions:

• This option is optional when --mode is set to 1. The TensorFlow, Caffe, and ONNX frameworks are supported. To include this option, ensure that the --om and --framework options are consistent. If this option is not included, conversion from the offline model to a JSON file is performed by default. See the following example:

  --mode=1 --framework=0 --om=$HOME/module/resnet50.prototxt
  --mode=1 --framework=3 --om=$HOME/module/resnet50_tensorflow*.pb
  --mode=1 --framework=5 --om=$HOME/module/resnet50.onnx

• This option is required when --mode is set to 0 or 3. The Caffe, TensorFlow, MindSpore, and ONNX frameworks are supported.
• If this option is set to 0, the Caffe network model is used. The model includes a .prototxt model file and a .caffemodel weight file. The op name and op type fields of the two files must be the same (case-sensitive).
• If this option is set to 3, the TensorFlow network model is used. Only the FrozenGraphDef format is supported, that is, the model file with a .pb file name extension is supported. The .pb file is stored in Protobuf format. The network model and weight data are stored in the same file.
• If this option is set to 5, the ONNX network model is used. Operators of opset v11 to v18 in the ai.onnx operator domain are supported. The PTH model of the PyTorch framework can be converted into an ONNX model or exported as a standard *.air file by using TorchAir export. Then, the model can be converted.
• If this option is set to 1 and the MindSpore network model is used, note the following restrictions:
  • Only model files in *.air format are convertible.
  • --mode can only be set to 0.
  • --input_format can only be set to NCHW. Other values are invalid, but the model will be successfully converted.
  • Under the MindSpore framework, --input_shape, --out_nodes, --is_output_adjust_hw_layout, --input_fp16_nodes, --is_input_adjust_hw_layout, and --op_name_map do not take effect, but the model will be successfully converted.
  • If the model size exceeds 2 GB, the *.air file, weight folder, and weight file in the folder are generated when the model is saved in the MindSpore framework. Ensure that the weight folder is stored in the same directory as the *.air file when converting the model. Otherwise, errors are reported during model conversion.

Suggestions and Benefits

None

Example

• ONNX:

  --mode=0 --framework=5 --model=$HOME/module/resnet50.onnx --output=$HOME/module/out/onnx_resnet50 --soc_version=<soc_version>  

• TensorFlow:

  --mode=0 --framework=3 --model=$HOME/module/resnet50_tensorflow*.pb --output=$HOME/module/out/tf_resnet50 --soc_version=<soc_version>  

• Caffe:

  --mode=0 --framework=0 --model=$HOME/module/resnet50.prototxt --weight=$HOME/module/resnet50.caffemodel  --output=$HOME/module/out/caffe_resnet50 --soc_version=<soc_version>   

• .air model files:

  --mode=0 --framework=1 --model=$HOME/module/ResNet50.air --output=$HOME/module/out/ResNet50_mindspore --soc_version=<soc_version> 

Restrictions

• Certain Caffe models (such as Faster R-CNN, YOLOv3, YOLOv2, and SSD) contain operators that are not defined in the original Caffe framework, such as ROIPooling, Normalize, PSROI Pooling, and Upsample. To support these models on the Ascend AI Processor, customization is needed to reduce the complexities of custom operator development and postprocessing programming. For details, see Custom Caffe Network Modification.
• Original TensorFlow models containing control flow operators (such as Switch, Merge, LoopCond, Case, and While) are not directly convertible by ATC. You need to convert such a model with control flow operators to an intermediate model with function operators, and then convert the intermediate model to an offline model adapted to the Ascend AI Processor by using ATC. For details, see Custom Network Modification (TensorFlow).