- Image Classification

- Image Classification

All of the original and compressed models can be downloaded easily on the Model Compressor Model Zoo.

You can get Compressed results with Automatic Compression and Compressed (Adv.) results with Advanced Compression.


PyTorch

ModelBest PracticeTypeDatasetAccuracy (%)FLOPs (M)Params (M)Latency (ms)Model Size (MB)
VGG16OriginalCIFAR10074.00629.7615.3071.6559.65
VGG16Google ColabCompressed-1CIFAR10072.22 (-1.78)431.84 (1.46x)5.16 (2.96x)24.52 (2.91x)20.22 (2.95x)
VGG16Google ColabCompressed-2CIFAR10068.01 (-5.99)213.06 (2.96x)1.25 (12.26x)11.34 (6.32x)4.93 (12.10x)
MobileNetV2OriginalCIFAR10074.29189.302.3546.268.98
MobileNetV2Google ColabCompressedCIFAR10073.68 (-0.61)119.09 (1.59x)0.82 (2.88x)24.50 (1.89x)3.38 (2.66x)
RepVGGOriginalCIFAR10076.441715.7012.94248.1050.33
RepVGGGoogle ColabCompressed-1CIFAR10074.92 (-1.52)1644.88 (1.04x)10.64 (1.22x)113.35 (2.19x)41.81 (1.20x)
RepVGGGoogle ColabCompressed-2CIFAR10069.84 (-4.60)721.77 (2.38x)2.95 (4.39x)51.69 (4.80x)11.71 (4.30x)
ViTOriginalCIFAR10094.4233725.7685.801396.53327.43
ViTGoogle ColabCompressedCIFAR10093.30 (-1.12)14804.95 (2.28x)37.78 (2.27x)737.11 (1.89x)144.32 (2.27x)
  • The model’s latency is measured on Raspberry Pi 4B (1.5GHz ARM Cortex).
  • Options: FP32, ONNX runtime

TensorFlow-Keras

ModelBest PracticeTypeDatasetAccuracy (%)FLOPs (M)Params (M)Latency (ms)Model Size (MB)
VGG19OriginalCIFAR-10072.28796.7920.09189.3178.69
VGG19Google ColabCompressedCIFAR-10071.13 (-1.15)132.20 (6.03x)1.17 (17.13x)12.85 (14.73x)4.98 (15.80x)
VGG19Compressed (Adv.)CIFAR-10071.14 (-1.14)100.09 (7.96x)0.66 (30.38x)4.5 (42.06x)5.68 (13.85x)
ResNet50OriginalCIFAR-10078.032596.0623.71450.1493.31
ResNet50Google ColabCompressedCIFAR-10076.92 (-1.11)613.43 (4.23x)2.64 (8.99x)130.39 (3.45x)9.83 (9.49x)
ResNet50Compressed (Adv.)CIFAR-10076.63 (-1.4)224.70 (11.55x)2.17 (10.91x)48.37 (9.31x)18.35 (5.09x)
MobileNet V1OriginalCIFAR-10066.6892.903.3135.6113.28
MobileNet V1Google ColabCompressedCIFAR-10066.32 (-0.36)26.09 (3.56x)0.53 (6.24x)3.66 (9.73x)2.38 (5.58x)
MobileNet V1Compressed (Adv.)CIFAR-10066.11 (-0.57)17.90 (5.19x)0.35 (9.35x)2.08 (17.12x)3.3 (4.02x)
  • The model’s latency is measured on Raspberry Pi 4B (1.5GHz ARM Cortex).
  • Options: FP32, TensorFlow Lite