Command-Line Options

Before using the ATC tool to convert a model, check the tool restrictions and use the parameter overview function provided in this section to quickly preview related parameters.

General Restrictions

Before model conversion, pay attention to the following restrictions:

To convert a network model such as Faster R-CNN to an offline model supported by the Ascend AI Processor, you must modify the model file (.prototxt). For details, see Custom Caffe Network Modification.
Model conversion is supported for Caffe, TensorFlow, MindSpore, and ONNX frameworks.
- For MindSpore, Caffe, and ONNX, the input data type is FP32, FP16, or UINT8 (implemented by configuring --insert_op_conf for data preprocessing).
- If the original framework type is TensorFlow, the input data type can be FP16, FP32, UINT8, INT32, INT64, or BOOL.
For a Caffe model, the input data is up to 4D and operators such as Reshape and ExpandDims do not support 5D output.
dim! = 0 must be true for the inputs and outputs at all layers except const in a model.
Only the operators in CANN Operator Specifications are supported, and the operator restrictions must be met.
Due to software restrictions (the input data cannot be of the DT_INT8 type in the dynamic shape scenario), when the ATC tool is used to convert the quantized deployable model, parameters related to the dynamic shape cannot be used, such as --dynamic_batch_size and --dynamic_image_size. Otherwise, the model conversion fails.
When the ATC tool is used to convert a deployable model quantized by the AMCT tool, the high-precision feature cannot be used. For example, force_fp32 or must_keep_origin_dtype (fp32 input of the original graph) cannot be configured through --precision_mode, origin cannot be configured through --precision_mode_v2, and high_precision cannot be configured through --op_precision_mode. Setting quantization parameters in high-precision mode does not provide any performance benefits of quantization nor that of the high precision mode.

Command-Line Options

If options queried with the atc --help command are not described in Table 1, they are reserved or applicable to other products. You do not need to pay attention to such options.
You can use either of the following methods to convert a model by running the atc command. This document uses the first method as an example.
- atc param1=value1 param2=value2 ... (No space is allowed before value. Otherwise, it will be clipped, and the value of param is empty.)
- atc param1 value1 param2 value2 ...
ATC options can be prefixed with -- (for example, --help) or - (for example, -help). When - is used as the prefix, it will be automatically converted to -- during running of atc commands. This document uses the prefix --.
In an ATC option, you can use an underscore (_) or hyphen (-) to connect two strings (for example, soc_version or soc-version). This document uses underscores.
The values in the Required column in Table 1 are determined by the value of --mode.

**Table 1** ATC command-line options
Option	Description	Required (--mode = 0 or 3)	Default
--help or --h	Displays help information.	No	N/A
--mode	Sets the run mode.	No	0
--framework	Sets the framework of the original model.	Yes	N/A
--model	Sets the model file directory, including the file name.	Yes	N/A
--weight	Sets the weight file directory, including the file name.	No	N/A
--input_format	Sets the input format.	No	For MindSpore, Caffe, and ONNX: NCHW; For TensorFlow: NHWC
--input_shape	Sets the input shape.	No	N/A
--dynamic_batch_size	Sets dynamic batch size profiles. Applies to the scenario where image count per inference batch is unfixed.	No	N/A
--dynamic_image_size	Sets dynamic image size profiles. Applies to the scenario where the resolution of images input for inference is unfixed.	No	N/A
--dynamic_dims	Sets dynamic dimension profiles in ND format. Applies to the scenario where the dimensions for inference are unfixed.	No	N/A
--om	Sets the directory (including the file name) of the offline model or original model file to be converted to JSON format.	No	N/A
--singleop	Sets the single-operator definition file (JSON), which will be converted to an offline model supported by the Ascend AI Processor.	No	N/A
--output	Sets the directory (including the file name) of the offline model converted from a network from open-source framework. Sets the directory of the single-operator model converted from a single-operator description file.	Yes	N/A
--output_type	Sets the output data type of a network or an output node.	No	N/A
--check_report	Sets the precheck result file directory, including the file name.	No	check_result.json generated in the current path where the atc command is executed
--json	Sets the directory (including the file name) of the JSON file converted from the offline model or original model file.	No	N/A
--host_env_os	If the OS and its architecture of the model build environment are inconsistent with those of the model operating environment, set this option to the OS type of the model operating environment.	No	Run the atc --help command to view the default value of the --host_env_os option or view the value in the ${INSTALL_DIR}/opp/scene.info file.
--host_env_cpu	If the OS and its architecture of the model build environment are inconsistent with those of the model operating environment, set this option to the OS architecture of the model operating environment.	No	Run the atc --help command to view the default value of the --host_env_cpu option or view the value in the ${INSTALL_DIR}/opp/scene.info file.
--soc_version	Sets the target SoC version.	Yes	N/A
--core_type	Sets the Core type used for a model. If a model contains a Cube operator, only AICore is supported.	No	AiCore
--aicore_num	Sets the number of AI Cores for model build.	No	The maximum
--virtual_type	Allows an offline model to run on a virtual device generated by the computing power group.	No	0
--out_nodes	Sets the output nodes.	No	N/A
--input_fp16_nodes	Sets the name of the FP16 input node.	No	N/A
--insert_op_conf	Sets the configuration file directory (including the file name) of an operator to be inserted, for example, the AIPP operator for data preprocessing.	No	N/A
--external_weight	Sets whether to save the weights of the Const/Constant nodes on the original network in a separate file and converts the node type to FileConstant when the OM offline model is generated.	No	0
--op_name_map	Sets the directory (including the file name) of the mapping configuration file of a custom operator. The function of a custom operator varies according to the network. You can specify the mapping between the custom operator and the actual custom operator running on the network.	No	N/A
--is_input_adjust_hw_layout	Sets the data type and format of the network inputs to FP16 and NC1HWC0, respectively.	No	false
--is_output_adjust_hw_layout	Sets the data type and format of the network outputs to FP16 and NC1HWC0, respectively.	No	false
--buffer_optimize	Enables buffer optimization.	No	l2_optimize
--disable_reuse_memory	Enables memory reuse.	No	0
--enable_scope_fusion_passes	Enables specific fusion patterns at build time.	No	N/A
--fusion_switch_file	Sets the fusion switch configuration file directory, including the file name.	No	N/A
--enable_single_stream	Sets whether to enable single-stream serial execution of model inference.	No	false
--enable_small_channel	Enables small channel optimization. If it is enabled, performance benefits are generated at the first convolutional layer with channel <= 4.	No	0
--quant_dumpable	Collects the dump data of the quantization operator.	No	0
--compression_optimize_conf	Configures the path including the name of the compression optimization configuration file.	No	N/A
--mdl_bank_path	Sets the directory of the custom repository generated after Auto Tune tuning.	No	${HOME}/Ascend/latest/data/aoe/custom/graph/<soc_version>
--sparsity	Enables global sparsity.	No	0
--ac_parallel_enable	Allows the AI CPU operator and AI Core operator to run in parallel in the dynamic shape graph.	No	0
--oo_level	Specifies the optimization level for graph build.	No	O3
--oo_constant_folding	Enables constant folding optimization.	No	true
--oo_dead_code_elimination	Enables dead-edge elimination optimization.	No	true
--tiling_schedule_optimize	Sets whether to enable the optimization for tiling offload scheduling.	No	0
--topo_sorting_mode	Specifies the traversal mode when you compile operators in graph mode.	No	1
--precision_mode	Sets the precision mode of a model.	No	force_fp16
--precision_mode_v2	Sets the precision mode of a model. This option cannot be used together with --precision_mode. The recommended option is --precision_mode_v2.	No	fp16
--op_precision_mode	Sets the precision mode of an operator. You can use this option to set different precision modes for different operators.	No	N/A
--modify_mixlist	Sets the operators on the mixed precision list.	No	N/A
--optypelist_for_implmode	Sets the implementation mode of an operator on the optype list. The operator implementation mode can be either high_precision or high_performance.	No	N/A
--keep_dtype	Specifies whether the data type of a specific operator in the original model remains unchanged during model build by using a configuration file.	No	N/A
--customize_dtypes	Customizes operator computing accuracy during model build.	No	N/A
--op_bank_path	Sets the path of the custom repository generated after AOE tuning.	No	Default custom repository path: $HOME/Ascend/latest/data/aoe/custom/op
--allow_hf32	Enables the function of automatically replacing the FP32 data type with the HF32 data type.	No	Enable FP32-to-HF32 conversion for Conv operators; disable FP32-to-HF32 conversion for Matmul operators.
--is_weight_clip	Crops the weight data of the floating-point type.	No	1
--dump_mode	Sets whether to generate a JSON file with shape information.	No	0
--log	Sets the level of logs to be printed during ATC model conversion.	No	null
--display_model_info	Queries model information at build time or information about an existing offline model. The information includes model's usage of key resources, information about the build environment and operating environment, and more.	No	0
--op_compiler_cache_mode	Sets the disk cache mode for operator build.	No	disable
--op_compiler_cache_dir	Sets the disk cache directory for operator build.	No	$HOME/atc_data
--op_debug_level	Enables or disables TBE operator debug at operator build time.	No	0
--op_debug_config	Sets the path (including the file name) of the configuration file for enabling global memory (DDR) detection.	No	N/A
--debug_dir	Sets the path for storing debugging information files generated after operator build during model conversion and network migration, including the .o, .json, and .cce files.	No	./kernel_meta
--atomic_clean_policy	Sets whether to collectively clean up the memory occupied by all operators with the memset attribute (memset operators) on the network.	No	0
--deterministic	Enables or disables deterministic computing.	No	0
--deterministic_level	Deterministic computing level.	No	0
--status_check	Determines whether to add overflow detection logic during operator build.	No	0
--export_compile_stat	Sets whether to generate the fusion_result.json result file of operator fusion information (including graph fusion and UB fusion) during graph build.	No	1
--input_shape_range	Sets the input shape range. This option is deprecated. Avoid using it.	No	N/A
--op_select_implmode	The function of this option does not evolve and will be deprecated in later versions. You are advised to use --op_precision_mode. Sets the operator implementation mode.	No	high_performance
--shape_generalized_build_mode	Sets the shape build mode during graph build. Do not use this option because it will be deprecated in later versions.	No	shape_precise

Parent topic: Command Line Options