comparison image neural networks (original) (raw)

comparison image neural networks Comparison of the architectures of LeNet and AlexNet by CMG Lee with data from http://d2l.ai/chapter\_convolutional-neural-networks/lenet.html and http://d2l.ai/chapter\_convolutional-modern/alexnet.html . LeNet Output: 1 of 10 classes sigmoid sigmoid flatten sigmoid sigmoid Image: 28 (height) × 28 (width) × 1 (channel) Convolution with 5×5 kernel+2 padding:28×28×6 Convolution with 5×5 kernel (no pad): 10×10×16 Pool with 2×2 average kernel+2 stride: 14×14×6 Pool with 2×2 average kernel+2 stride: 5×5×16 Dense: 120 fully connected neurons Dense: 84 fully connected neurons Dense: 10 fully connected neurons AlexNet Output: 1 of 1000 classes ReLu ReLu ReLu ReLu ReLu flatten ReLu, dropout p=0.5 ReLu, dropout p=0.5 Image: 224 (height) × 224 (width) × 3 (channels) Convolution with 11×11 kernel+4 stride:54×54×96 Pool with 3×3 max. kernel+2 stride: 26×26×96 Convolution with 5×5 kernel+2 pad:26×26×256 Pool with 3×3 max. kernel+2 stride: 12×12×256 Convolution with 3×3 kernel+1 pad:12×12×384 Convolution with 3×3 kernel+1 pad:12×12×384 Convolution with 3×3 kernel+1 pad:12×12×256 Pool with 3×3 max. kernel+2 stride: 5×5×256 Dense: 4096 fully connected neurons Dense: 4096 fully connected neurons Dense: 1000 fully connected neurons