LibreYOLOXs Qualcomm-Optimized Model Artifacts

This directory is the shared artifact location for the notebooks in ../notebooks/. The repository tracks the documentation and ignore rules here; model binaries are downloaded or generated locally when you run the workflows.

The same source model is used for three deployment paths:

QAIRT local conversion and quantization
SNPE local conversion and quantization
QAI Hub cloud compilation, quantization, and profiling

Model overview

Property	Value
Architecture	YOLOX-Small (LibreYOLOXs)
Framework	LibreYOLO
Task	Object detection
Base checkpoint	LibreYOLO/LibreYOLOXs
Input tensor	`images`
Input shape	`1 x 3 x 640 x 640`
Input format	BGR `float32`, NCHW, values in `0-255`

Artifact flow

LibreYOLOXs.pt
    |
    `---> LibreYOLOXs.onnx
             |\
             | `---> qairt/LibreYOLOXs_fp32.dlc
             |        `---> qairt/LibreYOLOXs_int8.dlc
             |
             | `---> snpe/LibreYOLOXs_fp32.dlc
             |        `---> snpe/LibreYOLOXs_int8.dlc
             |
             `---> qaihub/LibreYOLOXs_fp32.dlc
                      `---> qaihub/LibreYOLOXs_int8.dlc

All three notebooks reuse the same exported ONNX model.

Expected files

Path	Produced by	Description
`LibreYOLOXs.pt`	All notebooks	Base PyTorch checkpoint, downloaded on demand
`LibreYOLOXs.onnx`	All notebooks	Shared ONNX export used by all backend flows
`qairt/LibreYOLOXs_fp32.dlc`	QAIRT notebook	FP32 DLC produced by `qairt-converter`
`qairt/LibreYOLOXs_int8.dlc`	QAIRT notebook	INT8 DLC produced by `qairt-quantizer`
`snpe/LibreYOLOXs_fp32.dlc`	SNPE notebook	FP32 DLC produced by `snpe-onnx-to-dlc`
`snpe/LibreYOLOXs_int8.dlc`	SNPE notebook	INT8 DLC produced by `snpe-dlc-quantize`
`qaihub/LibreYOLOXs_fp32.dlc`	QAI Hub notebook	FP32 DLC downloaded from a cloud compile job
`qaihub/LibreYOLOXs_int8.dlc`	QAI Hub notebook	INT8 DLC downloaded from cloud quantize + compile jobs

Tensor contract

Input

Property	Value
Name	`images`
Shape	`1 x 3 x 640 x 640`
Layout	NCHW
Data type	`float32`
Color order	BGR

Preprocessing used by the notebooks:

Letterbox-resize to 640 x 640.
Pad with value 114.
Keep BGR channel order.
Keep values in the 0-255 range.
Transpose from HWC to CHW before serialization / inference input assembly.

Outputs

The exported ONNX model is validated in each notebook before backend-specific compilation. The documentation in this directory assumes the shared decoded output layout used by the notebooks:

Output	Shape	Description
`bboxes`	`1 x 8400 x 4`	Bounding-box coordinates
`scores`	`1 x 8400 x 81`	Objectness plus 80 COCO class scores

Calibration data

INT8 flows use representative data prepared by the notebooks from COCO 2017 validation images:

source images are staged under docker/notebooks/dataset/
calibration tensors are written to docker/notebooks/subset/calib/
validation tensors and metadata are written to docker/notebooks/subset/val/
manifests are written to subset/calib/filenames.txt and subset/val/filenames.txt

The local QAIRT and SNPE quantizers read those manifests directly. The QAI Hub notebook loads the same .raw tensors into NumPy arrays and submits them to the cloud quantization job.

Downloading published artifacts

Published files are hosted on Hugging Face:

fabricionarcizo/LibreYOLOXs

Example download:

huggingface-cli download fabricionarcizo/LibreYOLOXs \
    --local-dir docker/models \
    --exclude "*.gitattributes" "README.md"

You can also generate the same artifact layout locally by running the notebooks in ../notebooks/.

Model tree for fabricionarcizo/LibreYOLOXs

Base model

LibreYOLO/LibreYOLOXs

Quantized

(1)

this model

fabricionarcizo
/

LibreYOLOXs