FasterAI-Labs/resnet50-imagenet1k-m53-int8
resnet50 — ImageNet-1k, -53.1 % MACs (Conv2d/Linear only), INT8
ImageNet-1k validation, 50000 images, one run (seed 42). Every published file is scored after reload, inside pinned numerics; the top-1 below is this artifact's own, not a comparison. The weights derive from the published checkpoint torchvision.models.resnet50 (IMAGENET1K_V2), which is the comparison point for size, memory and MACs only. The reload on a machine holding only the public package scores a frozen 2048-image subset of the same split. Seeds 42, 43, 44 of this point, one recipe and one calibration draw, gave 76.88 / 76.85 / 76.72 % on the TorchScript form and 76.73 / 77.00 / 76.76 % on the ONNX form top-1 — a range of 0.16 and 0.27 pt across 3 runs. No pair of seeds is distinguished on the same images.
Criteria
Top-1 on the evaluation set named above; size on disk, peak live activations and multiply-accumulates — the last two for one image of (3, 224, 224) at batch 1 — each against torchvision.models.resnet50 (IMAGENET1K_V2).
INT8 TorchScript — load with fastermodels.load(repo) or torch.jit.load (model.torchscript.pt)
Top-1: 76.9 % [76.5, 77.2] on 50000 images
Quantization step: -0.56 pt [-0.74, -0.38] of top-1 against the FP32 weights this file was built from, measured on the same images; the two agree on 93.0 % of images. The TorchScript and ONNX forms agree on 95.6 % of images and are scored separately.
Post-training static quantization (x86 backend), calibrated on 1000 class-balanced, unaugmented training images; BatchNorm folded.
INT8 rows store the FP32 form's weights as int8 with BatchNorm folded, so their parameter count is not comparable to an FP32 one and is left unmeasured. MACs and peak activation bytes of the INT8 rows are read off the PyTorch FX graph the TorchScript and ONNX files were written from.
INT8 ONNX — load with any ONNX runtime, scored here through onnxruntime (model.onnx)
Top-1: 76.7 % [76.4, 77.1] on 50000 images
Quantization step: -0.70 pt [-0.87, -0.52] of top-1 against the FP32 weights this file was built from, measured on the same images; the two agree on 93.1 % of images. The TorchScript and ONNX forms agree on 95.6 % of images and are scored separately.
Post-training static quantization (x86 backend), calibrated on 1000 class-balanced, unaugmented training images; BatchNorm folded.
INT8 rows store the FP32 form's weights as int8 with BatchNorm folded, so their parameter count is not comparable to an FP32 one and is left unmeasured. MACs and peak activation bytes of the INT8 rows are read off the PyTorch FX graph the TorchScript and ONNX files were written from.
Top-1 brackets give the 95 % interval over the evaluation images; size, memory and MACs gaps are against the reference.
Variants
Other points on the same ladder, from the same source model:
Latency
not measured
Provenance
- Source model: torchvision.models.resnet50 (IMAGENET1K_V2)
- fasterai:
0.4.1 - fastermodels:
0.1.0 - torch:
2.9.1+cu128 - measured_on:
2026-09-21 - sourcestatehash:
cf450479235c
Publication checks: 10/10 structural checks passed.