(args)
| 29 | |
| 30 | |
| 31 | def main(args): |
| 32 | qnn_config = QnnConfig.load_config(args.config_file if args.config_file else args) |
| 33 | |
| 34 | # ensure the working directory exist. |
| 35 | os.makedirs(args.artifact, exist_ok=True) |
| 36 | |
| 37 | instance = timm.create_model("fbnetc_100", pretrained=True).eval() |
| 38 | |
| 39 | data_num = 100 |
| 40 | inputs, targets = get_imagenet_dataset( |
| 41 | dataset_path=f"{args.dataset}", |
| 42 | data_size=data_num, |
| 43 | image_shape=(299, 299), |
| 44 | ) |
| 45 | |
| 46 | pte_filename = "fbnet_qnn" |
| 47 | |
| 48 | quant_dtype = { |
| 49 | QnnExecuTorchBackendType.kGpuBackend: None, |
| 50 | QnnExecuTorchBackendType.kHtpBackend: QuantDtype.use_8a8w, |
| 51 | }[qnn_config.backend] |
| 52 | build_executorch_binary( |
| 53 | model=instance, |
| 54 | qnn_config=qnn_config, |
| 55 | file_name=f"{args.artifact}/{pte_filename}", |
| 56 | dataset=inputs, |
| 57 | quant_dtype=quant_dtype, |
| 58 | ) |
| 59 | |
| 60 | adb = SimpleADB( |
| 61 | qnn_config=qnn_config, |
| 62 | pte_path=f"{args.artifact}/{pte_filename}.pte", |
| 63 | workspace=f"/data/local/tmp/executorch/{pte_filename}", |
| 64 | ) |
| 65 | adb.push(inputs=inputs) |
| 66 | adb.execute() |
| 67 | |
| 68 | # collect output data |
| 69 | output_data_folder = f"{args.artifact}/outputs" |
| 70 | make_output_dir(output_data_folder) |
| 71 | |
| 72 | output_raws = [] |
| 73 | |
| 74 | def post_process(): |
| 75 | for f in sorted( |
| 76 | os.listdir(output_data_folder), key=lambda f: int(f.split("_")[1]) |
| 77 | ): |
| 78 | filename = os.path.join(output_data_folder, f) |
| 79 | if re.match(r"^output_[0-9]+_[1-9].raw$", f): |
| 80 | os.remove(filename) |
| 81 | else: |
| 82 | output = np.fromfile(filename, dtype=np.float32) |
| 83 | output_raws.append(output) |
| 84 | |
| 85 | adb.pull(host_output_path=args.artifact, callback=post_process) |
| 86 | |
| 87 | # top-k analysis |
| 88 | predictions = [] |
no test coverage detected