Edit model card

ENOT-AutoDL pruning benchmark on ImageNet-1k

This repository contains models accelerated with ENOT-AutoDL framework. Models from Torchvision are used as a baseline. Evaluation code is also based on Torchvision references.

ResNet-50

Model Latency (MMACs) Accuracy (%)
ResNet-50 Torchvision 4144.85 76.14
ResNet-50 ENOT (x2) 2057.61 (x2.01) 75.48 (-0.66)
ResNet-50 ENOT (x4) 867.94 (x4.77) 73.58 (-2.57)

ViT-B/32

Model Latency (MMACs) Accuracy (%)
ViT-B/32 Torchvision 4413.99 75.91
ViT-B/32 ENOT (x4.8) 911.80 (x4.84) 75.68 (-0.23)
ViT-B/32 ENOT (x9) 490.78 (x8.99) 73.72 (-2.19)

MobileNetV2

Model Latency (MMACs) Accuracy (%)
MobileNetV2 Torchvision 334.23 71.88
MobileNetV2 ENOT (x1.6) 209.24 (x1.6) 71.38 (-0.5)
MobileNetV2 ENOT (x2.1) 156.80 (x2.13) 69.90 (-1.98)

Validation

To validate results, follow this steps:

  1. Install all required packages:
    pip install -r requrements.txt
    
  2. Calculate model latency:
    python measure_mac.py --model-ckpt path/to/model.pth
    
  3. Measure accuracy of ONNX model:
    python test.py --data-path path/to/imagenet --model-onnx path/to/model.onnx --batch-size 1
    
  4. Measure accuracy of PyTorch (.pth) model:
    python test.py --data-path path/to/imagenet --model-ckpt path/to/model.pth
    

If you want to book a demo, please contact us: enot@enot.ai .

Downloads last month
1
This model does not have enough activity to be deployed to Inference API (serverless) yet. Increase its social visibility and check back later, or deploy to Inference Endpoints (dedicated) instead.

Dataset used to train ENOT-AutoDL/imagenet-benchmark