LibreEfficientDetd3

EfficientDet-D3 with an EfficientNet-B3 backbone, a weighted BiFPN, 12.0M parameters, and a fixed 896x896 input, repackaged for LibreYOLO.

from libreyolo import LibreYOLO

model = LibreYOLO("LibreEfficientDetd3.pt")
results = model.predict("image.jpg")

LibreYOLO ships this family for detection inference and validation. Training is not implemented. ONNX, TorchScript, OpenVINO, and TensorRT exports have trained-checkpoint prediction-parity coverage on the D0 family representative.

Source

Converted from the official tf_efficientdet_d3_47-0b525f35.pth release asset from rwightman/efficientdet-pytorch 0.4.1 at commit c6dff775a36cea0bf9b76c58e59f936411c5ce01. Copyright 2020 Ross Wightman. Licensed under the Apache License 2.0.

The native EfficientNet block implementation also follows huggingface/pytorch-image-models v1.0.28 at commit 8ef73809f622e0031bd7f4940265734aef8b9978 (Apache-2.0). The original EfficientDet design is from Google Research's Apache-2.0 google/automl implementation.

Source checkpoint SHA-256: 0b525f352fea3c768fd1a3cc885f1016d24d1b99cf299e8861e784a4ec2f0eff.

Modifications

Checkpoint metadata wrap only. Learned parameter names, dtypes, and tensor values are unchanged. See weights/convert_efficientdet_weights.py in the LibreYOLO source repository.

Strict loading succeeds with no missing or unexpected keys. Against effdet 0.4.1, identical fixed inputs produce bit-exact FP32 class and box feature tensors, anchors, and all 5,000 decoded candidate rows (max_abs_diff == 0.0). LibreYOLO maps the sparse 90-slot COCO head to a contiguous COCO-80 interface.

Benchmarks

The pinned upstream model zoo reports 47.1 COCO val2017 box AP for this checkpoint. Independent accuracy and speed results will appear at visionanalysis.org/model/efficientdet-d3 after EfficientDet is added to the Vision Analysis model catalog.

Limitations

• Detection uses the fixed 896x896 evaluation canvas and COCO-80 classes.
• The focal-loss and anchor-matching training recipe is not included.
• TensorRT uses 3,840 pre-NMS candidates because TensorRT 10.x limits TopK; native, ONNX, TorchScript, and OpenVINO retain the upstream 5,000 candidates.

License

Apache License 2.0. See the LICENSE and NOTICE files in this repository. The upstream release does not attach a separate checkpoint-specific license object; this mirror relies on Apache-2.0 as implied by the releasing project, not a publisher-confirmed asset-specific grant.