| Calibration types | Description |
|---|---|
| index | file: calibration-index.txt (containing index) The file provided contains the indexes to be picked from the inputlist file for doing calibration. Example calibration index file: The index file can have indexes in multiple lines. |
| raw | file: calibration-raw.txt (containing paths to raw files) The user can provide a file with entries of the preprocessed raw files (one input set per row) to be used for calibration. |
| dataset | file: calibration-dataset.txt (containing paths to dataset files) In this case, the calibration images provided in the file are processed separately and passed to the inference stage. Note: The file paths provided may or may not be present in the inputlist file. |
| type | One of inbuilt comparator types (avg | abs | l1norm | l2norm | rme) or custom comparator. See Custom comparators on custom comparator configuration. |
| tol | Tolerance value – Depends on per comparator described below:
For all other inbuilt comparators – This defines the number of decimal places to round to before evaluating the comparator. |
| fetch_top | Out of all mismatched files, order by maximum % mismatch and write the fetch_top number of preprocessed input file paths in a file (<plat_name>_mm.txt) per platform. Default value: 1 |
| enabled | By default, enabled (True). |
| onnxrt | Runs on ONNX runtime platform. The version installed in the qaic-env is used ( default 1.8.0). This platform currently does not take any other parameter |
| tensorflow | Runs on TensorFlow (CPU). Supports TensorFlow frozen graphs. |
| caffe | Runs on Caffe framework (CPU). Only supported in qaic-docker environment. This platform currently does not take any other parameter. |
| aic | Runs on AIC cards installed in the system. The device to be used is selected from a pool (see aic-device-ids). |
| onnx_custom_op_lib | (Optional) Path to a custom-op library file that should be loaded to the ONNX runtime session. |
| precision | int8 | fp16 | fp32 / default |
| use_precompiled | Path to precompiled binaries. |
| reuse_pgq | Used to reuse the generated PGQ profile across the search space for compilation options for a given platform. Default is True. If set to False then the PGQ profile will be generated for all the generated platforms added after scanning over the search space for compilation options. |
| pgq_group | Can be any string. The same PGQ group can be provided to different AIC int8 platforms to indicate that the PGQ profile can be reused across these platforms. If reuse_pgq is set to False, then pgq_group is ignored. |
| params | Used only if use_precompiled is not provided or blank. Under params, the user can provide the compilation options. Details regarding these parameters can be referred from qaic-exec compiler documentation. Each line should contain one compilation parameter. Example below: params:
Users can also provide a search space for compilation options by separating values by a pipe operator. The example configuration below will result in four different platform configurations.
|
| is_ref | Optional parameter. Boolean value used to indicate the reference platform to be used by the comparator to compare with the model outputs. Generally, any non-aic platform would be chosen as reference platform unless explicitly specified. If no non-aic platform is present then first aic platform is chosen as reference platform. |
| use_gpu | Default False; flag used to determine if the user wants to run a non-aic platform on a GPU device. |
| gpu_ids | A list of GPU device IDs to use. Example: Note the tool automatically fetches available GPUs if use_gpu is True and gpu_ids are not provided by the user. |