feat: model_converter support for models from training_extensions

[mgumowsk]

What does this PR do?

Here's a summary of the most important features and changes in PR #594:

Architecture Refactor — Converter Registry

• Monolithic cli.py was split into a converters/ package with a BaseConverter, a CONVERTER_REGISTRY, and dedicated subclasses: TorchvisionConverter, TimmConverter, YoloConverter, and the new GetituneConverter
• The standalone yolo/yolo.py CLI entry point was removed and absorbed into the new architecture

Getitune (OpenVINO Training Extensions) Support

• Added GetituneConverter to convert models exported by the getitune (formerly OTX) framework, covering classification, multi-label classification, detection, segmentation, and rotated detection tasks

Multiplie Datasets support

• DatasetRegistry class loads a datasets.json config file mapping dataset type identifiers (e.g. imagenet-1k, coco-detection, ade20k) to local filesystem paths, decoupling model configs from environment-specific paths
• Model configs now declare a dataset_type field; INT8 quantization is skipped if absent
• datasets/ package with readers for: ClassFolderReader (ImageNet-style), CocoImagesReader (COCO detection/segmentation), and Ade20kReader (ADE20K segmentation)
• reader_for() factory selects the correct reader by dataset_type

Task-Aware Metrics Module

• metrics/ package with a Metric abstract base and concrete implementations:
• TopOneAccuracy — classification top-1
• MultilabelMAP — multi-label mean AP via torchmetrics
• CocoDetectionMAP — object detection mAP via pycocotools
• SemSegMIoU — semantic segmentation mIoU via confusion matrix
• metric_for() factory dispatches to the right metric based on (dataset_type, model_type, task)

Conversion Summary Reporting

• --report [PATH] CLI flag generates a console table + Markdown report with per-model accuracy (Original, FP32, FP16, INT8) and a status (OK, ACCURACY DROP >5%, FAILED: conversion, SKIPPED, etc.)
• Rolling upsert_result() updates the report after each model so partial runs are preserved

Timm model fixes

• TimmConverter now reads preprocessing parameters (mean, std, input size) directly from timm.data.resolve_data_config, overriding hand-maintained config values to avoid silent accuracy regressions
• Pinned huggingface_revision is now required for reproducibility

Config & Presets

• examples/config.json replaced by a comprehensive presets/config.json covering torchvision, timm, YOLO11, and getitune models, plus a presets/datasets.json template
• New dataset_type field added to model config schema
• Extensive new unit tests covering the dataset readers, all metric implementations, the dataset registry, the converter classes, the reporting module, and CLI integration

Example conversion summary report

Conversion Summary Report

Model Full Name	Model Type	Model Library	Original URL	Original Accuracy	FP32 Accuracy	FP16 Accuracy	INT8 Accuracy	Status
EfficientNet-B0	Classification	torchvision	https://download.pytorch.org/models/efficientnet_b0_rwightman-3dd342df.pth	81.60%	81.60%	81.60%	81.80%	OK
EfficientNet-B1	Classification	torchvision	https://download.pytorch.org/models/efficientnet_b1_rwightman-bac287d4.pth	85.20%	85.20%	85.00%	83.60%	OK
EfficientNet-B2	Classification	torchvision	https://download.pytorch.org/models/efficientnet_b2_rwightman-c35c1473.pth	86.80%	86.80%	86.80%	86.00%	OK
EfficientNet-B3	Classification	torchvision	https://download.pytorch.org/models/efficientnet_b3_rwightman-b3899882.pth	89.40%	89.40%	89.40%	89.60%	OK
EfficientNet-B4	Classification	torchvision	https://download.pytorch.org/models/efficientnet_b4_rwightman-23ab8bcd.pth	89.00%	89.00%	89.00%	87.60%	OK
EfficientNet-B5	Classification	torchvision	https://download.pytorch.org/models/efficientnet_b5_lukemelas-1a07897c.pth	90.20%	90.20%	90.20%	89.40%	OK
EfficientNet-B6	Classification	torchvision	https://download.pytorch.org/models/efficientnet_b6_lukemelas-24a108a5.pth	90.40%	90.40%	90.40%	90.20%	OK
EfficientNet-B7	Classification	torchvision	https://download.pytorch.org/models/efficientnet_b7_lukemelas-c5b4e57e.pth	91.40%	91.40%	91.40%	91.20%	OK
EfficientNetV2-S	Classification	torchvision	https://download.pytorch.org/models/efficientnet_v2_s-dd5fe13b.pth	90.60%	90.60%	90.60%	91.20%	OK
EfficientNetV2-M	Classification	torchvision	https://download.pytorch.org/models/efficientnet_v2_m-dc08266a.pth	90.20%	90.20%	90.20%	89.80%	OK
EfficientNetV2-L	Classification	torchvision	https://download.pytorch.org/models/efficientnet_v2_l-59c71312.pth	12.80%	12.80%	12.80%	0.00%	ACCURACY DROP >5%
MobileNetV2	Classification	torchvision	https://download.pytorch.org/models/mobilenet_v2-b0353104.pth	74.00%	74.00%	74.00%	73.40%	OK
MobileNetV3-Small	Classification	torchvision	https://download.pytorch.org/models/mobilenet_v3_small-047dcff4.pth	45.00%	45.00%	45.20%	39.80%	ACCURACY DROP >5%
MobileNetV3-Large	Classification	torchvision	https://download.pytorch.org/models/mobilenet_v3_large-8738ca79.pth	75.60%	75.60%	75.60%	75.00%	OK
ResNet-18	Classification	torchvision	https://download.pytorch.org/models/resnet18-f37072fd.pth	71.60%	71.60%	71.60%	71.80%	OK
ResNet-34	Classification	torchvision	https://download.pytorch.org/models/resnet34-b627a593.pth	76.40%	76.40%	76.40%	76.20%	OK
ResNet-50	Classification	torchvision	https://download.pytorch.org/models/resnet50-0676ba61.pth	79.20%	79.20%	79.20%	79.60%	OK
ResNet-101	Classification	torchvision	https://download.pytorch.org/models/resnet101-63fe2227.pth	81.80%	81.80%	81.80%	81.40%	OK
ResNet-152	Classification	torchvision	https://download.pytorch.org/models/resnet152-394f9c45.pth	82.20%	82.20%	82.20%	82.20%	OK
Mask R-CNN ResNet-50 FPN	MaskRCNN	torchvision	https://download.pytorch.org/models/maskrcnn_resnet50_fpn_coco-bf2d0c1e.pth	N/A	0.01%	0.01%	0.01%	OK
MobileNetV2-0.5 LAMB ImageNet-1k	Classification	timm	https://huggingface.co/timm/mobilenetv2_050.lamb_in1k	71.60%	71.60%	71.40%	68.00%	OK
MobileNetV2-1.0 RandAugment ImageNet-1k	Classification	timm	https://huggingface.co/timm/mobilenetv2_100.ra_in1k	81.00%	81.00%	81.00%	79.00%	OK
MobileNetV2-1.1 RandAugment ImageNet-1k	Classification	timm	https://huggingface.co/timm/mobilenetv2_110d.ra_in1k	81.40%	81.40%	81.40%	75.60%	ACCURACY DROP >5%
MobileNetV2-1.2 RandAugment ImageNet-1k	Classification	timm	https://huggingface.co/timm/mobilenetv2_120d.ra_in1k	83.40%	83.40%	83.60%	83.00%	OK
MobileNetV2-1.4 RandAugment ImageNet-1k	Classification	timm	https://huggingface.co/timm/mobilenetv2_140.ra_in1k	81.80%	81.80%	81.80%	81.20%	OK
EfficientNet-B0 RandAugment ImageNet-1k	Classification	timm	https://huggingface.co/timm/efficientnet_b0.ra_in1k	83.80%	83.80%	83.80%	82.00%	OK
EfficientNet-B0 RA4 3600 Epochs ImageNet-1k	Classification	timm	https://huggingface.co/timm/efficientnet_b0.ra4_e3600_r224_in1k	85.20%	85.20%	85.20%	58.80%	ACCURACY DROP >5%
EfficientNet-B1 Fine-tuned ImageNet-1k	Classification	timm	https://huggingface.co/timm/efficientnet_b1.ft_in1k	84.20%	84.20%	84.20%	83.80%	OK
EfficientNet-B1 Pruned ImageNet-1k	Classification	timm	https://huggingface.co/timm/efficientnet_b1_pruned.in1k	84.80%	84.80%	84.80%	82.80%	OK
EfficientNet-B1 RA4 3600 Epochs ImageNet-1k	Classification	timm	https://huggingface.co/timm/efficientnet_b1.ra4_e3600_r240_in1k	87.80%	87.80%	87.80%	75.60%	ACCURACY DROP >5%
ViT-Tiny Patch16 224 AugReg ImageNet-21k	Classification	timm	https://huggingface.co/timm/vit_tiny_patch16_224.augreg_in21k	37.60%	37.60%	37.80%	37.60%	OK
DINOv2-Small Patch14 518 LVD-142M	Classification	timm	https://huggingface.co/timm/vit_small_patch14_dinov2.lvd142m	0.20%	0.20%	0.20%	0.20%	OK
YOLO11n	YOLO11	yolo	N/A	36.85%	N/A	37.01%	36.52%	OK
YOLO11s	YOLO11	yolo	N/A	44.71%	N/A	44.88%	43.95%	OK
YOLO11m	YOLO11	yolo	N/A	48.22%	N/A	48.65%	47.69%	OK
YOLO11l	YOLO11	yolo	N/A	51.00%	N/A	50.11%	49.90%	OK
YOLO11x	YOLO11	yolo	N/A	52.09%	N/A	52.42%	51.81%	OK

Getitune models, preliminary results before fixes in training_extensions repository

Conversion Summary Report

Model Full Name	Model Type	Model Library	Original URL	Original Accuracy	FP32 Accuracy	FP16 Accuracy	INT8 Accuracy	Status
DINOv2 Classification	Classification	getitune	N/A	N/A	1.00%	1.00%	1.00%	OK
EfficientNet-B0 Classification	Classification	getitune	N/A	N/A	0.00%	0.00%	0.00%	OK
EfficientNet-V2 Classification	Classification	getitune	N/A	N/A	0.00%	0.00%	0.00%	OK
MobileNet-V3 Large Classification	Classification	getitune	N/A	N/A	0.00%	0.00%	0.20%	OK
EfficientNet-B3 Classification	Classification	getitune	N/A	N/A	0.00%	0.00%	0.00%	OK
ViT-Tiny Classification	Classification	getitune	N/A	N/A	1.00%	1.00%	1.00%	OK
EfficientNet-B0 Multi-Label Classification	Classification	getitune	N/A	N/A	0.23%	0.23%	0.17%	OK
EfficientNet-V2 Multi-Label Classification	Classification	getitune	N/A	N/A	N/A	N/A	N/A	OK (no accuracy data)
DINOv2 Multi-Label Classification	Classification	getitune	N/A	N/A	0.20%	0.20%	0.27%	OK
EfficientNet-B3 Multi-Label Classification	Classification	getitune	N/A	N/A	0.21%	0.21%	0.23%	OK
MobileNet-V3-Large Multi-Label Classification	Classification	getitune	N/A	N/A	0.22%	0.22%	0.22%	OK
ViT-Tiny Multi-Label Classification	Classification	getitune	N/A	N/A	0.24%	0.24%	0.24%	OK
YOLOX-S	YOLOX	getitune	N/A	N/A	33.99%	33.99%	32.98%	OK
YOLOX-L	YOLOX	getitune	N/A	N/A	43.77%	43.77%	42.62%	OK
YOLOX-X	YOLOX	getitune	N/A	N/A	45.58%	45.58%	44.53%	OK
SSD MobileNet-V2	SSD	getitune	N/A	N/A	0.04%	0.04%	0.03%	OK
ATSS MobileNet-V2	SSD	getitune	N/A	N/A	0.00%	0.00%	0.00%	OK
ATSS ResNet-50 FPN	SSD	getitune	N/A	N/A	N/A	N/A	N/A	FAILED: conversion
YOLOX-Tiny	YOLOX	getitune	N/A	N/A	0.00%	0.00%	0.00%	OK
DEIM D-FINE-L	SSD	getitune	N/A	N/A	0.00%	0.00%	0.00%	OK
DEIM D-FINE-M	SSD	getitune	N/A	N/A	0.00%	0.00%	0.00%	OK
DEIM D-FINE-X	SSD	getitune	N/A	N/A	0.00%	0.00%	0.00%	OK
DEIMv2-L	SSD	getitune	N/A	N/A	0.00%	0.00%	0.00%	OK
DEIMv2-M	SSD	getitune	N/A	N/A	0.00%	0.00%	0.00%	OK
DEIMv2-S	SSD	getitune	N/A	N/A	0.00%	0.00%	0.00%	OK
D-FINE-X	SSD	getitune	N/A	N/A	0.00%	0.00%	0.00%	OK
RF-DETR Large	SSD	getitune	N/A	N/A	0.00%	0.00%	0.00%	OK
RF-DETR Medium	SSD	getitune	N/A	N/A	0.00%	0.00%	0.00%	OK
RF-DETR Small	SSD	getitune	N/A	N/A	0.00%	0.00%	0.00%	OK
RT-DETR ResNet-50	SSD	getitune	N/A	N/A	0.00%	0.00%	0.00%	OK

Fixes # (issue)

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you make sure to update the documentation with your changes?
• Did you write any new necessary tests?