Model Zoo — Pre-trained Models

The Voyager Model Zoo is a collection of pre-trained AI models ready to run on Axelera hardware. When you run inference.py yolov5s-v7-coco usb:0, the model name (yolov5s-v7-coco) comes from the Model Zoo.

Listing available models​

From the SDK root directory (with environment activated):

make

This lists three categories:

How model names work​

Model names follow a pattern:

\<architecture\>-\<dataset\>[-\<variant\>]

Examples:

Task types​

Running a model​

# Object detection with USB camera
./inference.py yolov5s-v7-coco usb:0

# Classification with a video file
./inference.py resnet50-imagenet media/traffic1_1080p.mp4

# Headless benchmarking (no display)
./inference.py yolov8s-coco-onnx usb:0 --no-display --frames 1000

The first time you run a model, the SDK:

1. Downloads the pre-trained weights (if not cached)
2. Compiles the model for the AIPU
3. Caches the compiled model for subsequent runs
4. Runs inference

Subsequent runs skip steps 1-3 and start immediately.

Datasets​

Models are trained on specific datasets. The dataset name in the model identifier tells you what the model can recognize:

Non-redistributable datasets​

Most datasets download automatically when you first run a model. A few datasets require manual registration and download due to licensing restrictions. These must be downloaded by hand from the links below, then placed in the specified directory within your SDK installation. The SDK raises an error with the expected path if a required dataset is missing.

You are responsible for adhering to the terms and conditions of each dataset's license.

Performance characteristics​

The tables below list all Model Zoo models for this SDK release. Columns:

• Ref FP32 — accuracy of the original floating-point model
• Accuracy loss — FP32 accuracy minus quantized int8 accuracy (lower is better)
• Ref PCIe FPS — host throughput on Intel Core i9-13900K + Metis 1× PCIe card
• Ref M.2 FPS — host throughput on Intel Core i5-1145G7E + Metis 1× M.2 card

Accuracy is measured using:

./inference.py \<model\> dataset --pipe=torch-aipu --no-display

Throughput is measured using a 720p h.264 video file:

./inference.py \<model\> media/traffic2_720p.mp4 --pipe=gst --no-display

Image Classification​

Object Detection​

Semantic Segmentation​

Instance Segmentation​

Keypoint Detection​

Depth Estimation​

License Plate Recognition​

Image Enhancement (Super Resolution)​

Face Recognition​

Re-Identification​

Large Language Models​

For usage details see the LLM Inference guide.

Experimenting with optimized input shapes​

Most models are trained on square inputs (640×640), but real-world video is often rectangular (16:9). Standard pipelines pad the input ("letterboxing"), forcing the model to process empty pixels.

By switching to a rectangular input shape that matches your video's aspect ratio, you can often achieve significant speedups with minimal accuracy impact. This is especially effective for fixed-camera applications like surveillance or traffic monitoring.

How to test​

Export models with dynamic input shapes, then compare:

# Standard 640×640
./inference.py yolox-m-coco-onnx dataset --pipe=torch-aipu --no-display

# Rectangular 640×480
./inference.py yolox-m-coco-onnx-rect dataset --pipe=torch-aipu --no-display

Expected results​

Custom weights​

You can use your own trained weights with any model architecture. This involves updating the model's YAML configuration to point to your custom weight file. See Deploy Custom Weights for the full walkthrough.

See also​

• First Inference — run your first model
• Measure Accuracy — benchmark model performance
• inference.py — full command reference
• Glossary — definitions of COCO, YOLO, mAP, and other terms