prius12345678/gymnext-onnx

GymNeXt — Gym Equipment Classifier (ONNX)

ConvNeXt-Tiny fine-tuned on 32 classes of gym equipment. Exported to ONNX with external data for browser inference via onnxruntime-web.

Model Details

Classes (alphabetical, matching index order)

 0: ab_crunch_machine    16: lateral_raise
 1: arm_curl             17: leg_curl
 2: back_extension       18: leg_extension
 3: barbell              19: leg_press
 4: bench                20: leg_raise_tower
 5: chest_fly            21: pulley_machine
 6: chest_press          22: recumbent_bike
 7: chinning_dipping     23: seated_dip
 8: dumbbell             24: seated_row
 9: elliptical           25: shoulder_press
10: functional_trainer   26: smith_machine
11: hip_abduction        27: squat_rack
12: home_machine         28: stability_ball
13: incline_bench        29: stationary_bike
14: kettlebell           30: torso_rotation
15: lat_pulldown         31: treadmill

Preprocessing

Matches torchvision validation pipeline:

1. Resize short edge to 224 (bicubic)
2. Center crop to 224×224
3. Normalize: mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]

Files

• gym_convnext_ft.onnx — model graph (~707 KB)
• gym_convnext_ft.onnx.data — weights (~107 MB)

Both files are required. When using onnxruntime-web, pass the data file via externalData option:

const session = await ort.InferenceSession.create(modelUrl, {
  executionProviders: ['webgpu', 'wasm'],
  externalData: [{ path: 'gym_convnext_ft.onnx.data', data: dataUrl }],
});

Training

Fine-tuned from ImageNet-pretrained ConvNeXt-Tiny using a merged dataset of gym equipment images (32 classes). Training used AdamW optimizer with cosine LR schedule, EMA, and mixed-precision (AMP).