Convolutional Neural Network with CUDA
- Linear
- Conv2D
- MaxPool2D
- ReLU
- Softmax
- Sigmoid
- NLLLoss
- RMSProp
- CMake 3.8+
- MSVC14.00/GCC6+
- CUDA 10.x [Not compatible with CUDA 11.x]
mkdir build && cd build
cmake .. -DCMAKE_BUILD_TYPE=Release
make -j10
mkdir mnist_data && cd mnist_data
wget -c https://yann.lecun.com/exdb/mnist/train-images-idx3-ubyte.gz
wget -c https://yann.lecun.com/exdb/mnist/train-labels-idx1-ubyte.gz
wget -c https://yann.lecun.com/exdb/mnist/t10k-images-idx3-ubyte.gz
wget -c https://yann.lecun.com/exdb/mnist/t10k-labels-idx1-ubyte.gz
gunzip train-images-idx3-ubyte.gz
gunzip train-labels-idx1-ubyte.gz
gunzip t10k-labels-idx1-ubyte.gz
gunzip t10k-images-idx3-ubyte.gz
cd .. && ./mnist
conv 1 32 5 relu
maxpool 2
conv 32 64 5 relu
maxpool 2
conv 64 128 3 relu
fc 4 * 128 128 relu
fc 128 10 relu
softmax
shuffle = true
batch_size = 128
learning_rate = 0.003
L2 = 0.0001
beta = 0.99
- 1 epoch 93%
- 10 epochs 99.12%
- 30 epochs 99.23%
- 10s / epoch(GTX1070)
- Faster matmul kernel function
- CUDA Streams
- High Performance Convolutional Neural Networks for Document Processing
- 卷积神经网络(CNN)反向传播算法
- 矩阵求导术
- Caffe
- CUDA Toolkit Documents