|
| 1 | +# Copyright (c) 2024 Intel Corporation |
| 2 | +# Licensed under the Apache License, Version 2.0 (the "License"); |
| 3 | +# you may not use this file except in compliance with the License. |
| 4 | +# You may obtain a copy of the License at |
| 5 | +# http://www.apache.org/licenses/LICENSE-2.0 |
| 6 | +# Unless required by applicable law or agreed to in writing, software |
| 7 | +# distributed under the License is distributed on an "AS IS" BASIS, |
| 8 | +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 9 | +# See the License for the specific language governing permissions and |
| 10 | +# limitations under the License. |
| 11 | + |
| 12 | +from pathlib import Path |
| 13 | +from typing import Dict, Tuple |
| 14 | + |
| 15 | +import openvino as ov |
| 16 | +import torch |
| 17 | +from ultralytics import YOLO |
| 18 | +from ultralytics.data.utils import check_det_dataset |
| 19 | +from ultralytics.engine.validator import BaseValidator as Validator |
| 20 | +from ultralytics.utils.torch_utils import de_parallel |
| 21 | + |
| 22 | +import nncf |
| 23 | +from nncf.torch import disable_patching |
| 24 | +from tests.post_training.pipelines.base import OV_BACKENDS |
| 25 | +from tests.post_training.pipelines.base import BackendType |
| 26 | +from tests.post_training.pipelines.base import PTQTestPipeline |
| 27 | + |
| 28 | + |
| 29 | +class UltralyticsDetection(PTQTestPipeline): |
| 30 | + """Pipeline for Yolo detection models from the Ultralytics repository""" |
| 31 | + |
| 32 | + def prepare_model(self) -> None: |
| 33 | + if self.batch_size != 1: |
| 34 | + raise RuntimeError("Batch size > 1 is not supported") |
| 35 | + |
| 36 | + model_path = f"{self.fp32_model_dir}/{self.model_id}" |
| 37 | + yolo = YOLO(f"{model_path}.pt") |
| 38 | + self.validator, self.data_loader = self._prepare_validation(yolo, "coco128.yaml") |
| 39 | + self.dummy_tensor = torch.ones((1, 3, 640, 640)) |
| 40 | + |
| 41 | + if self.backend in OV_BACKENDS + [BackendType.FP32]: |
| 42 | + onnx_model_path = Path(f"{model_path}.onnx") |
| 43 | + ir_model_path = self.fp32_model_dir / "model_fp32.xml" |
| 44 | + yolo.export(format="onnx", dynamic=True, half=False) |
| 45 | + ov.save_model(ov.convert_model(onnx_model_path), ir_model_path) |
| 46 | + self.model = ov.Core().read_model(ir_model_path) |
| 47 | + |
| 48 | + if self.backend == BackendType.FX_TORCH: |
| 49 | + pt_model = yolo.model |
| 50 | + # Run mode one time to initialize all |
| 51 | + # internal variables |
| 52 | + pt_model(self.dummy_tensor) |
| 53 | + |
| 54 | + with torch.no_grad(): |
| 55 | + with disable_patching(): |
| 56 | + self.model = torch.export.export(pt_model, args=(self.dummy_tensor,), strict=False).module() |
| 57 | + |
| 58 | + def prepare_preprocessor(self) -> None: |
| 59 | + pass |
| 60 | + |
| 61 | + @staticmethod |
| 62 | + def _validate_fx( |
| 63 | + model: ov.Model, data_loader: torch.utils.data.DataLoader, validator: Validator, num_samples: int = None |
| 64 | + ) -> Tuple[Dict, int, int]: |
| 65 | + for batch_i, batch in enumerate(data_loader): |
| 66 | + if num_samples is not None and batch_i == num_samples: |
| 67 | + break |
| 68 | + batch = validator.preprocess(batch) |
| 69 | + preds = model(batch["img"]) |
| 70 | + preds = validator.postprocess(preds) |
| 71 | + validator.update_metrics(preds, batch) |
| 72 | + stats = validator.get_stats() |
| 73 | + return stats, validator.seen, validator.nt_per_class.sum() |
| 74 | + |
| 75 | + @staticmethod |
| 76 | + def _validate_ov( |
| 77 | + model: ov.Model, data_loader: torch.utils.data.DataLoader, validator: Validator, num_samples: int = None |
| 78 | + ) -> Tuple[Dict, int, int]: |
| 79 | + model.reshape({0: [1, 3, -1, -1]}) |
| 80 | + compiled_model = ov.compile_model(model) |
| 81 | + output_layer = compiled_model.output(0) |
| 82 | + for batch_i, batch in enumerate(data_loader): |
| 83 | + if num_samples is not None and batch_i == num_samples: |
| 84 | + break |
| 85 | + batch = validator.preprocess(batch) |
| 86 | + preds = torch.from_numpy(compiled_model(batch["img"])[output_layer]) |
| 87 | + preds = validator.postprocess(preds) |
| 88 | + validator.update_metrics(preds, batch) |
| 89 | + stats = validator.get_stats() |
| 90 | + return stats, validator.seen, validator.nt_per_class.sum() |
| 91 | + |
| 92 | + def get_transform_calibration_fn(self): |
| 93 | + def transform_func(batch): |
| 94 | + return self.validator.preprocess(batch)["img"] |
| 95 | + |
| 96 | + return transform_func |
| 97 | + |
| 98 | + def prepare_calibration_dataset(self): |
| 99 | + self.calibration_dataset = nncf.Dataset(self.data_loader, self.get_transform_calibration_fn()) |
| 100 | + |
| 101 | + @staticmethod |
| 102 | + def _prepare_validation(model: YOLO, data: str) -> Tuple[Validator, torch.utils.data.DataLoader]: |
| 103 | + custom = {"rect": False, "batch": 1} # method defaults |
| 104 | + args = {**model.overrides, **custom, "mode": "val"} # highest priority args on the right |
| 105 | + |
| 106 | + validator = model._smart_load("validator")(args=args, _callbacks=model.callbacks) |
| 107 | + stride = 32 # default stride |
| 108 | + validator.stride = stride # used in get_dataloader() for padding |
| 109 | + validator.data = check_det_dataset(data) |
| 110 | + validator.init_metrics(de_parallel(model)) |
| 111 | + |
| 112 | + data_loader = validator.get_dataloader(validator.data.get(validator.args.split), validator.args.batch) |
| 113 | + |
| 114 | + return validator, data_loader |
| 115 | + |
| 116 | + def _validate(self): |
| 117 | + if self.backend == BackendType.FP32: |
| 118 | + stats, _, _ = self._validate_ov(self.model, self.data_loader, self.validator) |
| 119 | + elif self.backend in OV_BACKENDS: |
| 120 | + stats, _, _ = self._validate_ov(self.compressed_model, self.data_loader, self.validator) |
| 121 | + elif self.backend == BackendType.FX_TORCH: |
| 122 | + stats, _, _ = self._validate_fx(self.compressed_model, self.data_loader, self.validator) |
| 123 | + else: |
| 124 | + raise RuntimeError(f"Backend {self.backend} is not supported in UltralyticsDetection") |
| 125 | + |
| 126 | + self.run_info.metric_name = "mAP50(B)" |
| 127 | + self.run_info.metric_value = stats["metrics/mAP50(B)"] |
0 commit comments