Move XPU/Opt related PRs from releases2.0 to develop (#3295)

* Enable training on XPU devices in OTX2.0 (#3094)

* add raising an error when metric is None

* added accelerators

* fix packages

* fix assigning model

* debug on MAX

* change precision

* update MixedPrecisionXPUPlugin

* debug

* added monkey patching

* minor

* minor

* added patch for mmengine

* fix OD and IS

* benchmark debug

* change device

* quick fix for instance seg

* fix pre-commit

* fix pre-commit

* clean the code

* added additional flag for mmcv

* added unit tests

* fixed unit test

* fix linter

* added unit tests and replied comments

* fix pre-commit

* minor fix

* added documentation

* fix unit test

* add workaround for semantic segmentation

* remove RoiAlignTest due to unstability

* minor

* remove strategy back

* try to patch SingleDeviceStrategy

* added auto xpu configuration

* patch strategy

* small fix

* reply to comments

* move patching xpu packages to accelerator

* fix test_xpu test

* remove do-not-install-mmcv

* fix pre-commit

* remove torch.xpu.optimize for segmentation

---------

Co-authored-by: Emily <emily.chun@intel.com>

* Add exporter/demo unit tests (#3218)

* added unit tests. Need to clean up

* move tests

* fix pre-commit

* return demo back

* minor

* delete unnecessery comments

* fix unit test

* fix pre-commit

* fix pre-commit 2

* fix test_postprocess_openvino_model

* fix unit tests

* test_precommit

* Fix a bug that engine.test doesn't work with XPU (#3293)

* fix bug

* align with pre-commit

---------

Co-authored-by: Emily <emily.chun@intel.com>

* fix merge conflicts for pre-commit

* fix precommit 2

* fix unit test

* fix pre-commit

* fix export tests

* fix pre-commit

* fix tox

* fix pre-commit

---------

Co-authored-by: Emily <emily.chun@intel.com>
Co-authored-by: Eunwoo Shin <eunwoo.shin@intel.com>

src/otx/algo/__init__.py

@@ -3,6 +3,24 @@
 #
 """Module for OTX custom algorithms, e.g., model, losses, hook, etc..."""
 
-from . import action_classification, classification, detection, segmentation, visual_prompting
+from . import (
+    accelerators,
+    action_classification,
+    classification,
+    detection,
+    plugins,
+    segmentation,
+    strategies,
+    visual_prompting,
+)
 
-__all__ = ["action_classification", "classification", "detection", "segmentation", "visual_prompting"]
+__all__ = [
+    "action_classification",
+    "classification",
+    "detection",
+    "segmentation",
+    "visual_prompting",
+    "strategies",
+    "accelerators",
+    "plugins",
+]

src/otx/algo/accelerators/__init__.py

@@ -0,0 +1,8 @@
+# Copyright (C) 2024 Intel Corporation
+# SPDX-License-Identifier: Apache-2.0
+#
+"""Lightning accelerator for XPU device."""
+
+from .xpu import XPUAccelerator
+
+__all__ = ["XPUAccelerator"]

src/otx/algo/accelerators/xpu.py

@@ -0,0 +1,88 @@
+"""Lightning accelerator for XPU device."""
+# Copyright (C) 2023 Intel Corporation
+# SPDX-License-Identifier: Apache-2.0
+#
+from __future__ import annotations
+
+from typing import Any, Union
+
+import numpy as np
+import torch
+from lightning.pytorch.accelerators import AcceleratorRegistry
+from lightning.pytorch.accelerators.accelerator import Accelerator
+from mmcv.ops.nms import NMSop
+from mmcv.ops.roi_align import RoIAlign
+from mmengine.structures import instance_data
+
+from otx.algo.detection.utils import monkey_patched_nms, monkey_patched_roi_align
+from otx.utils.utils import is_xpu_available
+
+
+class XPUAccelerator(Accelerator):
+    """Support for a XPU, optimized for large-scale machine learning."""
+
+    accelerator_name = "xpu"
+
+    def setup_device(self, device: torch.device) -> None:
+        """Sets up the specified device."""
+        if device.type != "xpu":
+            msg = f"Device should be xpu, got {device} instead"
+            raise RuntimeError(msg)
+
+        torch.xpu.set_device(device)
+        self.patch_packages_xpu()
+
+    @staticmethod
+    def parse_devices(devices: str | list | torch.device) -> list:
+        """Parses devices for multi-GPU training."""
+        if isinstance(devices, list):
+            return devices
+        return [devices]
+
+    @staticmethod
+    def get_parallel_devices(devices: list) -> list[torch.device]:
+        """Generates a list of parrallel devices."""
+        return [torch.device("xpu", idx) for idx in devices]
+
+    @staticmethod
+    def auto_device_count() -> int:
+        """Returns number of XPU devices available."""
+        return torch.xpu.device_count()
+
+    @staticmethod
+    def is_available() -> bool:
+        """Checks if XPU available."""
+        return is_xpu_available()
+
+    def get_device_stats(self, device: str | torch.device) -> dict[str, Any]:
+        """Returns XPU devices stats."""
+        return {}
+
+    def teardown(self) -> None:
+        """Cleans-up XPU-related resources."""
+        self.revert_packages_xpu()
+
+    def patch_packages_xpu(self) -> None:
+        """Patch packages when xpu is available."""
+        # patch instance_data from mmengie
+        long_type_tensor = Union[torch.LongTensor, torch.xpu.LongTensor]
+        bool_type_tensor = Union[torch.BoolTensor, torch.xpu.BoolTensor]
+        instance_data.IndexType = Union[str, slice, int, list, long_type_tensor, bool_type_tensor, np.ndarray]
+
+        # patch nms and roi_align
+        self._nms_op_forward = NMSop.forward
+        self._roi_align_forward = RoIAlign.forward
+        NMSop.forward = monkey_patched_nms
+        RoIAlign.forward = monkey_patched_roi_align
+
+    def revert_packages_xpu(self) -> None:
+        """Revert packages when xpu is available."""
+        NMSop.forward = self._nms_op_forward
+        RoIAlign.forward = self._roi_align_forward
+
+
+AcceleratorRegistry.register(
+    XPUAccelerator.accelerator_name,
+    XPUAccelerator,
+    description="Accelerator supports XPU devices",
+)

src/otx/algo/detection/utils/__init__.py

@@ -1,3 +1,8 @@
 # Copyright (C) 2024 Intel Corporation
 # SPDX-License-Identifier: Apache-2.0
-"""Data structures for detection task."""
+#
+"""utils for detection task."""
+
+from .mmcv_patched_ops import monkey_patched_nms, monkey_patched_roi_align
+
+__all__ = ["monkey_patched_nms", "monkey_patched_roi_align"]

src/otx/algo/detection/utils/mmcv_patched_ops.py

@@ -0,0 +1,73 @@
+# Copyright (C) 2024 Intel Corporation
+# SPDX-License-Identifier: Apache-2.0
+#
+"""utils for detection task."""
+
+from __future__ import annotations
+
+from typing import TYPE_CHECKING
+
+import torch
+from mmcv.utils import ext_loader
+from torchvision.ops import nms as tv_nms
+from torchvision.ops import roi_align as tv_roi_align
+
+if TYPE_CHECKING:
+    from mmcv.ops.nms import NMSop
+    from mmcv.ops.roi_align import RoIAlign
+
+ext_module = ext_loader.load_ext("_ext", ["nms", "softnms", "nms_match", "nms_rotated", "nms_quadri"])
+
+
+def monkey_patched_nms(
+    ctx: NMSop,
+    bboxes: torch.Tensor,
+    scores: torch.Tensor,
+    iou_threshold: float,
+    offset: float,
+    score_threshold: float,
+    max_num: int,
+) -> torch.Tensor:
+    """Runs MMCVs NMS with torchvision.nms, or forces NMS from MMCV to run on CPU."""
+    _ = ctx
+    is_filtering_by_score = score_threshold > 0
+    if is_filtering_by_score:
+        valid_mask = scores > score_threshold
+        bboxes, scores = bboxes[valid_mask], scores[valid_mask]
+        valid_inds = torch.nonzero(valid_mask, as_tuple=False).squeeze(dim=1)
+
+    if bboxes.dtype == torch.bfloat16:
+        bboxes = bboxes.to(torch.float32)
+    if scores.dtype == torch.bfloat16:
+        scores = scores.to(torch.float32)
+
+    if offset == 0:
+        inds = tv_nms(bboxes, scores, float(iou_threshold))
+    else:
+        device = bboxes.device
+        bboxes = bboxes.to("cpu")
+        scores = scores.to("cpu")
+        inds = ext_module.nms(bboxes, scores, iou_threshold=float(iou_threshold), offset=offset)
+        bboxes = bboxes.to(device)
+        scores = scores.to(device)
+
+    if max_num > 0:
+        inds = inds[:max_num]
+    if is_filtering_by_score:
+        inds = valid_inds[inds]
+    return inds
+
+
+def monkey_patched_roi_align(self: RoIAlign, _input: torch.Tensor, rois: torch.Tensor) -> torch.Tensor:
+    """Replaces MMCVs roi align with the one from torchvision.
+
+    Args:
+        self: patched instance
+        _input: NCHW images
+        rois: Bx5 boxes. First column is the index into N. The other 4 columns are xyxy.
+    """
+    if "aligned" in tv_roi_align.__code__.co_varnames:
+        return tv_roi_align(_input, rois, self.output_size, self.spatial_scale, self.sampling_ratio, self.aligned)
+    if self.aligned:
+        rois -= rois.new_tensor([0.0] + [0.5 / self.spatial_scale] * 4)
+    return tv_roi_align(_input, rois, self.output_size, self.spatial_scale, self.sampling_ratio)

src/otx/algo/plugins/__init__.py

@@ -0,0 +1,8 @@
+# Copyright (C) 2024 Intel Corporation
+# SPDX-License-Identifier: Apache-2.0
+#
+"""Plugin for mixed-precision training on XPU."""
+
+from .xpu_precision import MixedPrecisionXPUPlugin
+
+__all__ = ["MixedPrecisionXPUPlugin"]

src/otx/algo/plugins/xpu_precision.py

@@ -0,0 +1,117 @@
+# Copyright (C) 2024 Intel Corporation
+# SPDX-License-Identifier: Apache-2.0
+#
+"""Plugin for mixed-precision training on XPU."""
+
+from __future__ import annotations
+
+from contextlib import contextmanager
+from typing import TYPE_CHECKING, Any, Callable, Generator
+
+import torch
+from lightning.pytorch.plugins.precision.precision import Precision
+from lightning.pytorch.utilities import GradClipAlgorithmType
+from lightning.pytorch.utilities.exceptions import MisconfigurationException
+from torch import Tensor
+from torch.optim import LBFGS, Optimizer
+
+if TYPE_CHECKING:
+    import lightning.pytorch as pl
+    from lightning_fabric.utilities.types import Optimizable
+
+
+class MixedPrecisionXPUPlugin(Precision):
+    """Plugin for Automatic Mixed Precision (AMP) training with ``torch.xpu.autocast``.
+
+    Args:
+        scaler: An optional :class:`torch.cuda.amp.GradScaler` to use.
+    """
+
+    def __init__(self, scaler: torch.cuda.amp.GradScaler | None = None) -> None:
+        self.scaler = scaler
+
+    def pre_backward(self, tensor: Tensor, module: pl.LightningModule) -> Tensor:
+        """Apply grad scaler before backward."""
+        if self.scaler is not None:
+            tensor = self.scaler.scale(tensor)
+        return super().pre_backward(tensor, module)
+
+    def optimizer_step(  # type: ignore[override]
+        self,
+        optimizer: Optimizable,
+        model: pl.LightningModule,
+        closure: Callable,
+        **kwargs: dict,
+    ) -> None | dict:
+        """Make an optimizer step using scaler if it was passed."""
+        if self.scaler is None:
+            # skip scaler logic, as bfloat16 does not require scaler
+            return super().optimizer_step(
+                optimizer,
+                model=model,
+                closure=closure,
+                **kwargs,
+            )
+        if isinstance(optimizer, LBFGS):
+            msg = "Native AMP and the LBFGS optimizer are not compatible."
+            raise MisconfigurationException(
+                msg,
+            )
+        closure_result = closure()
+
+        if not _optimizer_handles_unscaling(optimizer):
+            # Unscaling needs to be performed here in case we are going to apply gradient clipping.
+            # Optimizers that perform unscaling in their `.step()` method are not supported (e.g., fused Adam).
+            # Note: `unscale` happens after the closure is executed, but before the `on_before_optimizer_step` hook.
+            self.scaler.unscale_(optimizer)
+
+        self._after_closure(model, optimizer)
+        skipped_backward = closure_result is None
+        # in manual optimization, the closure does not return a value
+        if not model.automatic_optimization or not skipped_backward:
+            # note: the scaler will skip the `optimizer.step` if nonfinite gradients are found
+            step_output = self.scaler.step(optimizer, **kwargs)
+            self.scaler.update()
+            return step_output
+        return closure_result
+
+    def clip_gradients(
+        self,
+        optimizer: Optimizer,
+        clip_val: int | float = 0.0,
+        gradient_clip_algorithm: GradClipAlgorithmType = GradClipAlgorithmType.NORM,
+    ) -> None:
+        """Handle grad clipping with scaler."""
+        if clip_val > 0 and _optimizer_handles_unscaling(optimizer):
+            msg = f"The current optimizer, {type(optimizer).__qualname__}, does not allow for gradient clipping"
+            " because it performs unscaling of gradients internally. HINT: Are you using a 'fused' optimizer?"
+            raise RuntimeError(
+                msg,
+            )
+        super().clip_gradients(optimizer=optimizer, clip_val=clip_val, gradient_clip_algorithm=gradient_clip_algorithm)
+
+    @contextmanager
+    def forward_context(self) -> Generator[None, None, None]:
+        """Enable autocast context."""
+        with torch.xpu.autocast(True):
+            yield
+
+    def state_dict(self) -> dict[str, Any]:
+        """Returns state dict of the plugin."""
+        if self.scaler is not None:
+            return self.scaler.state_dict()
+        return {}
+
+    def load_state_dict(self, state_dict: dict[str, torch.Tensor]) -> None:
+        """Loads state dict to the plugin."""
+        if self.scaler is not None:
+            self.scaler.load_state_dict(state_dict)
+
+
+def _optimizer_handles_unscaling(optimizer: torch.optim.Optimizer) -> bool:
+    """Determines if a PyTorch optimizer handles unscaling gradients in the step method ratherthan through the scaler.
+
+    Since, the current implementation of this function checks a PyTorch internal variable on the optimizer, the return
+    value will only be reliable for built-in PyTorch optimizers.
+    """
+    return getattr(optimizer, "_step_supports_amp_scaling", False)

src/otx/algo/strategies/__init__.py

@@ -0,0 +1,8 @@
+# Copyright (C) 2024 Intel Corporation
+# SPDX-License-Identifier: Apache-2.0
+#
+"""Lightning strategy for single XPU device."""
+
+from .xpu_single import SingleXPUStrategy
+
+__all__ = ["SingleXPUStrategy"]