Back to Gallery

ILSVRC2012 Classification: 256x256x3 Convolutional Neural Network (57.56% top 1 accuracy, 80.71% top 5 accuracy, 380ms/frame on Raspberry Pi 3)

Download	dr_I256x256x3CMCMCMCMCMCMCMC1AS.ell.zip
Accuracy	ILSVRC2012: 80.71% (Top 5), 57.56% (Top 1)
Performance	Raspberry Pi 3 (Raspbian) @ 700MHz: 380ms/frame
Uncompressed Size	91MB
Input	256 x 256 x {B,G,R}
Architecture	Convolution ⇨ 256x256x16 size=3x3, stride=1, type=float32, activation=leaky relu Pooling ⇨ 130x130x16 size=2x2, stride=2, operation=max Convolution ⇨ 128x128x32 size=3x3, stride=1, type=float32, activation=leaky relu Pooling ⇨ 66x66x32 size=2x2, stride=2, operation=max Convolution ⇨ 64x64x64 size=3x3, stride=1, type=float32, activation=leaky relu Pooling ⇨ 34x34x64 size=2x2, stride=2, operation=max Convolution ⇨ 32x32x128 size=3x3, stride=1, type=float32, activation=leaky relu Pooling ⇨ 18x18x128 size=2x2, stride=2, operation=max Convolution ⇨ 16x16x256 size=3x3, stride=1, type=float32, activation=leaky relu Pooling ⇨ 10x10x256 size=2x2, stride=2, operation=max Convolution ⇨ 8x8x512 size=3x3, stride=1, type=float32, activation=leaky relu Pooling ⇨ 6x6x512 size=2x2, stride=2, operation=max Convolution ⇨ 4x4x1024 size=3x3, stride=1, type=float32, activation=leaky relu Pooling ⇨ 2x2x1024 size=2x2, stride=2, operation=max Convolution ⇨ 2x2x1000 size=1x1, stride=1, type=float32, activation=leaky relu Pooling ⇨ 1x1x1000 size=2x2, stride=1, operation=average Softmax ⇨ 1x1x1000
Output	ILSVRC2012 1000 classes
Notes	Trained by Juan Lema using CNTK 2.2