dgenerate-ultralytics-headless 8.3.155__tar.gz → 8.3.156__tar.gz

{dgenerate_ultralytics_headless-8.3.155 → dgenerate_ultralytics_headless-8.3.156}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: dgenerate-ultralytics-headless
-Version: 8.3.155
+Version: 8.3.156
 Summary: Automatically built Ultralytics package with python-opencv-headless dependency instead of python-opencv
 Author-email: Glenn Jocher <glenn.jocher@ultralytics.com>, Jing Qiu <jing.qiu@ultralytics.com>
 Maintainer-email: Ultralytics <hello@ultralytics.com>

{dgenerate_ultralytics_headless-8.3.155 → dgenerate_ultralytics_headless-8.3.156}/dgenerate_ultralytics_headless.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: dgenerate-ultralytics-headless
-Version: 8.3.155
+Version: 8.3.156
 Summary: Automatically built Ultralytics package with python-opencv-headless dependency instead of python-opencv
 Author-email: Glenn Jocher <glenn.jocher@ultralytics.com>, Jing Qiu <jing.qiu@ultralytics.com>
 Maintainer-email: Ultralytics <hello@ultralytics.com>

{dgenerate_ultralytics_headless-8.3.155 → dgenerate_ultralytics_headless-8.3.156}/ultralytics/__init__.py

@@ -1,6 +1,6 @@
 # Ultralytics 🚀 AGPL-3.0 License - https://ultralytics.com/license
 
-__version__ = "8.3.155"
+__version__ = "8.3.156"
 
 import os
 

{dgenerate_ultralytics_headless-8.3.155 → dgenerate_ultralytics_headless-8.3.156}/ultralytics/data/build.py

@@ -154,7 +154,7 @@ def build_grounding(cfg, img_path, json_file, batch, mode="train", rect=False, s
     )
 
 
-def build_dataloader(dataset, batch: int, workers: int, shuffle: bool = True, rank: int = -1):
+def build_dataloader(dataset, batch: int, workers: int, shuffle: bool = True, rank: int = -1, drop_last: bool = False):
     """
     Create and return an InfiniteDataLoader or DataLoader for training or validation.
 
@@ -164,6 +164,7 @@ def build_dataloader(dataset, batch: int, workers: int, shuffle: bool = True, ra
         workers (int): Number of worker threads for loading data.
         shuffle (bool, optional): Whether to shuffle the dataset.
         rank (int, optional): Process rank in distributed training. -1 for single-GPU training.
+        drop_last (bool, optional): Whether to drop the last incomplete batch.
 
     Returns:
         (InfiniteDataLoader): A dataloader that can be used for training or validation.
@@ -189,6 +190,7 @@ def build_dataloader(dataset, batch: int, workers: int, shuffle: bool = True, ra
         collate_fn=getattr(dataset, "collate_fn", None),
         worker_init_fn=seed_worker,
         generator=generator,
+        drop_last=drop_last,
     )
 
 

{dgenerate_ultralytics_headless-8.3.155 → dgenerate_ultralytics_headless-8.3.156}/ultralytics/engine/exporter.py

@@ -62,7 +62,6 @@ import shutil
 import subprocess
 import time
 import warnings
-from contextlib import contextmanager
 from copy import deepcopy
 from datetime import datetime
 from pathlib import Path
@@ -107,6 +106,7 @@ from ultralytics.utils.downloads import attempt_download_asset, get_github_asset
 from ultralytics.utils.export import export_engine, export_onnx
 from ultralytics.utils.files import file_size, spaces_in_path
 from ultralytics.utils.ops import Profile, nms_rotated
+from ultralytics.utils.patches import arange_patch
 from ultralytics.utils.torch_utils import TORCH_1_13, get_cpu_info, get_latest_opset, select_device
 
 
@@ -199,27 +199,6 @@ def try_export(inner_func):
     return outer_func
 
 
-@contextmanager
-def arange_patch(args):
-    """
-    Workaround for ONNX torch.arange incompatibility with FP16.
-
-    https://github.com/pytorch/pytorch/issues/148041.
-    """
-    if args.dynamic and args.half and args.format == "onnx":
-        func = torch.arange
-
-        def arange(*args, dtype=None, **kwargs):
-            """Return a 1-D tensor of size with values from the interval and common difference."""
-            return func(*args, **kwargs).to(dtype)  # cast to dtype instead of passing dtype
-
-        torch.arange = arange  # patch
-        yield
-        torch.arange = func  # unpatch
-    else:
-        yield
-
-
 class Exporter:
     """
     A class for exporting YOLO models to various formats.
@@ -345,8 +324,6 @@ class Exporter:
             LOGGER.warning("half=True only compatible with GPU export, i.e. use device=0")
             self.args.half = False
         self.imgsz = check_imgsz(self.args.imgsz, stride=model.stride, min_dim=2)  # check image size
-        if self.args.int8 and engine:
-            self.args.dynamic = True  # enforce dynamic to export TensorRT INT8
         if self.args.optimize:
             assert not ncnn, "optimize=True not compatible with format='ncnn', i.e. use optimize=False"
             assert self.device.type == "cpu", "optimize=True not compatible with cuda devices, i.e. use device='cpu'"
@@ -555,8 +532,6 @@ class Exporter:
         """Build and return a dataloader for calibration of INT8 models."""
         LOGGER.info(f"{prefix} collecting INT8 calibration images from 'data={self.args.data}'")
         data = (check_cls_dataset if self.model.task == "classify" else check_det_dataset)(self.args.data)
-        # TensorRT INT8 calibration should use 2x batch size
-        batch = self.args.batch * (2 if self.args.format == "engine" else 1)
         dataset = YOLODataset(
             data[self.args.split or "val"],
             data=data,
@@ -564,7 +539,7 @@ class Exporter:
             task=self.model.task,
             imgsz=self.imgsz[0],
             augment=False,
-            batch_size=batch,
+            batch_size=self.args.batch,
         )
         n = len(dataset)
         if n < self.args.batch:
@@ -574,7 +549,7 @@ class Exporter:
             )
         elif n < 300:
             LOGGER.warning(f"{prefix} >300 images recommended for INT8 calibration, found {n} images.")
-        return build_dataloader(dataset, batch=batch, workers=0)  # required for batch loading
+        return build_dataloader(dataset, batch=self.args.batch, workers=0, drop_last=True)  # required for batch loading
 
     @try_export
     def export_torchscript(self, prefix=colorstr("TorchScript:")):

{dgenerate_ultralytics_headless-8.3.155 → dgenerate_ultralytics_headless-8.3.156}/ultralytics/models/yolo/classify/train.py

@@ -228,10 +228,9 @@ class ClassificationTrainer(BaseTrainer):
             batch (Dict[str, torch.Tensor]): Batch containing images and class labels.
             ni (int): Number of iterations.
         """
+        batch["batch_idx"] = torch.arange(len(batch["img"]))  # add batch index for plotting
         plot_images(
-            images=batch["img"],
-            batch_idx=torch.arange(len(batch["img"])),
-            cls=batch["cls"].view(-1),  # warning: use .view(), not .squeeze() for Classify models
+            labels=batch,
             fname=self.save_dir / f"train_batch{ni}.jpg",
             on_plot=self.on_plot,
         )

{dgenerate_ultralytics_headless-8.3.155 → dgenerate_ultralytics_headless-8.3.156}/ultralytics/models/yolo/detect/train.py

@@ -13,6 +13,7 @@ from ultralytics.engine.trainer import BaseTrainer
 from ultralytics.models import yolo
 from ultralytics.nn.tasks import DetectionModel
 from ultralytics.utils import LOGGER, RANK
+from ultralytics.utils.patches import override_configs
 from ultralytics.utils.plotting import plot_images, plot_labels, plot_results
 from ultralytics.utils.torch_utils import de_parallel, torch_distributed_zero_first
 
@@ -210,6 +211,8 @@ class DetectionTrainer(BaseTrainer):
         Returns:
             (int): Optimal batch size.
         """
-        train_dataset = self.build_dataset(self.data["train"], mode="train", batch=16)
+        with override_configs(self.args, overrides={"cache": False}) as self.args:
+            train_dataset = self.build_dataset(self.data["train"], mode="train", batch=16)
         max_num_obj = max(len(label["cls"]) for label in train_dataset.labels) * 4  # 4 for mosaic augmentation
+        del train_dataset  # free memory
         return super().auto_batch(max_num_obj)

{dgenerate_ultralytics_headless-8.3.155 → dgenerate_ultralytics_headless-8.3.156}/ultralytics/utils/export.py

@@ -143,11 +143,12 @@ def export_engine(
         for inp in inputs:
             profile.set_shape(inp.name, min=min_shape, opt=shape, max=max_shape)
         config.add_optimization_profile(profile)
+        if int8:
+            config.set_calibration_profile(profile)
 
     LOGGER.info(f"{prefix} building {'INT8' if int8 else 'FP' + ('16' if half else '32')} engine as {engine_file}")
     if int8:
         config.set_flag(trt.BuilderFlag.INT8)
-        config.set_calibration_profile(profile)
         config.profiling_verbosity = trt.ProfilingVerbosity.DETAILED
 
         class EngineCalibrator(trt.IInt8Calibrator):
@@ -181,7 +182,11 @@ def export_engine(
                 trt.IInt8Calibrator.__init__(self)
                 self.dataset = dataset
                 self.data_iter = iter(dataset)
-                self.algo = trt.CalibrationAlgoType.MINMAX_CALIBRATION
+                self.algo = (
+                    trt.CalibrationAlgoType.ENTROPY_CALIBRATION_2  # DLA quantization needs ENTROPY_CALIBRATION_2
+                    if dla is not None
+                    else trt.CalibrationAlgoType.MINMAX_CALIBRATION
+                )
                 self.batch = dataset.batch_size
                 self.cache = Path(cache)
 

{dgenerate_ultralytics_headless-8.3.155 → dgenerate_ultralytics_headless-8.3.156}/ultralytics/utils/patches.py

@@ -2,8 +2,10 @@
 """Monkey patches to update/extend functionality of existing functions."""
 
 import time
+from contextlib import contextmanager
+from copy import copy
 from pathlib import Path
-from typing import List, Optional
+from typing import Any, Dict, List, Optional
 
 import cv2
 import numpy as np
@@ -139,3 +141,48 @@ def torch_save(*args, **kwargs):
             if i == 3:
                 raise e
             time.sleep((2**i) / 2)  # Exponential backoff: 0.5s, 1.0s, 2.0s
+
+
+@contextmanager
+def arange_patch(args):
+    """
+    Workaround for ONNX torch.arange incompatibility with FP16.
+
+    https://github.com/pytorch/pytorch/issues/148041.
+    """
+    if args.dynamic and args.half and args.format == "onnx":
+        func = torch.arange
+
+        def arange(*args, dtype=None, **kwargs):
+            """Return a 1-D tensor of size with values from the interval and common difference."""
+            return func(*args, **kwargs).to(dtype)  # cast to dtype instead of passing dtype
+
+        torch.arange = arange  # patch
+        yield
+        torch.arange = func  # unpatch
+    else:
+        yield
+
+
+@contextmanager
+def override_configs(args, overrides: Optional[Dict[str, Any]] = None):
+    """
+    Context manager to temporarily override configurations in args.
+
+    Args:
+        args (IterableSimpleNamespace): Original configuration arguments.
+        overrides (Dict[str, Any]): Dictionary of overrides to apply.
+
+    Yields:
+        (IterableSimpleNamespace): Configuration arguments with overrides applied.
+    """
+    if overrides:
+        original_args = copy(args)
+        for key, value in overrides.items():
+            setattr(args, key, value)
+        try:
+            yield args
+        finally:
+            args.__dict__.update(original_args.__dict__)
+    else:
+        yield args