Environment:VainF Torch Pruning YOLO Pruning Dependencies

Knowledge Sources	Torch-Pruning Ultralytics YOLOv8
Domains	Computer_Vision, Object_Detection, Model_Compression
Last Updated	2026-02-08 12:00 GMT

Overview

Ultralytics and related dependencies required for YOLO object detection model pruning workflows.

Description

This environment extends the core PyTorch environment with the Ultralytics library and related packages needed for the YOLO pruning examples. The examples/yolov8/yolov8_pruning.py script uses the Ultralytics YOLO API for model loading, training, validation, and export. The pruning pipeline replaces standard C2f modules with a pruning-compatible C2f_v2 variant and iteratively prunes and fine-tunes the model.

For YOLOv5 and YOLOv7 examples, the respective upstream repositories must be cloned separately, as they are not installable via pip.

Usage

Use this environment for the Object Detection Pruning workflow, specifically when:

• Pruning YOLOv8 models with iterative prune-finetune loops
• Exporting pruned models to ONNX format
• Generating pruning performance graphs (requires matplotlib)

System Requirements

Category	Requirement	Notes
OS	Linux (recommended)	Ultralytics supports Linux, macOS, Windows
Hardware	NVIDIA GPU	Training and fine-tuning require GPU
Disk	Varies	COCO dataset requires ~20GB

Dependencies

Python Packages

• torch >= 2.0
• numpy
• ultralytics (YOLOv8 framework)
• torch-pruning
• matplotlib (for performance graphs)
• onnx (for model export)

Credentials

No credentials required. COCO dataset is downloaded automatically by Ultralytics.

Quick Install

pip install torch>=2.0 numpy torch-pruning ultralytics matplotlib onnx

Code Evidence

Ultralytics imports from examples/yolov8/yolov8_pruning.py:1-15:

from ultralytics import YOLO, __version__
from ultralytics.nn.modules import C2f, Conv, Bottleneck, Detect
from ultralytics.engine.trainer import BaseTrainer
from ultralytics.nn.tasks import attempt_load_one_weight

YOLO model loading from examples/yolov8/yolov8_pruning.py:277:

model = YOLO(args.model)

ONNX export from examples/yolov8/yolov8_pruning.py:384:

model.export(format='onnx')

Common Errors

Error Message	Cause	Solution
ModuleNotFoundError: No module named 'ultralytics'	Ultralytics not installed	pip install ultralytics
AttributeError: 'C2f' has no attribute 'cv0'	C2f_v2 replacement not applied	Ensure replace_c2f_with_c2f_v2() is called before pruning
FileNotFoundError: ... coco.yaml	COCO dataset not found	Set data path in pruning config or let Ultralytics download automatically

Compatibility Notes

• YOLOv8 only: The pruning script uses Ultralytics-specific APIs. YOLOv5 and YOLOv7 have separate scripts that depend on their respective upstream repositories.
• C2f_v2 replacement: The standard Ultralytics C2f module fuses operations in a way that is incompatible with structural pruning. The C2f_v2 class splits the fused cv1 into separate cv0 and cv1 convolutions to allow proper channel pruning.
• train_v2: The script patches the YOLO training loop to avoid saving in half-precision (which would lose the pruned architecture) and to support early stopping based on mAP drop.

Related Pages

• Implementation:VainF_Torch_Pruning_C2f_V2_Replacement