A lightweight C library for artificial neural networks
# acquire source code and compile git clone https://github.com/attractivechaos/kann cd kann; make # or "make CBLAS=/path/to/openblas" for faster matrix multiplication # learn unsigned addition (30000 samples; numbers within 10000) seq 30000 | awk -v m=10000 '{a=int(m*rand());b=int(m*rand());print a,b,a+b}' \ | ./examples/rnn-bit -m7 -o add.kan - # apply the model (output 1138429, the sum of the two numbers) echo 400958 737471 | ./examples/rnn-bit -Ai add.kan -
KANN is a standalone and lightweight library in C for constructing and training small to medium artificial neural networks such as multi-layer perceptrons, convolutional neural networks and recurrent neural networks (including LSTM and GRU). It implements graph-based reverse-mode automatic differentiation and allows to build topologically complex neural networks with recurrence, shared weights and multiple inputs/outputs/costs. In comparison to mainstream deep learning frameworks such as TensorFlow, KANN is not as scalable, but it is close in flexibility, has a much smaller code base and only depends on the standard C library. In comparison to other lightweight frameworks such as tiny-dnn, KANN is still smaller, times faster and much more versatile, supporting RNN, VAE and non-standard neural networks that may fail these lightweight frameworks.
KANN could be potentially useful when you want to experiment small to medium neural networks in C/C++, to deploy no-so-large models without worrying about dependency hell, or to learn the internals of deep learning libraries.
Flexible. Model construction by building a computational graph with operators. Support RNNs, weight sharing and multiple inputs/outputs.
Efficient. Reasonably optimized matrix product and convolution. Support mini-batching and effective multi-threading. Sometimes faster than mainstream frameworks in their CPU-only mode.
Small and portable. As of now, KANN has less than 4000 lines of code in four source code files, with no non-standard dependencies by default. Compatible with ANSI C compilers.
CPU only. As such, KANN is not intended for training huge neural networks.
Lack of some common operators and architectures such as batch normalization.
Verbose APIs for training RNNs.
The KANN library is composed of four files:
kautodiff.{h,c}and
kann.{h,c}. You are encouraged to include these files in your source code tree. No installation is needed. To compile examples:
sh makeThis generates a few executables in the examples directory.
Comments in the header files briefly explain the APIs. More documentations can be found in the doc directory. Examples using the library are in the examples directory.
Working with neural networks usually involves three steps: model construction, training and prediction. We can use layer APIs to build a simple model:
c kann_t *ann; kad_node_t *t; t = kann_layer_input(784); // for MNIST t = kad_relu(kann_layer_dense(t, 64)); // a 64-neuron hidden layer with ReLU activation t = kann_layer_cost(t, 10, KANN_C_CEM); // softmax output + multi-class cross-entropy cost ann = kann_new(t, 0); // compile the network and collate variablesFor this simple feedforward model with one input and one output, we can train it with:
c int n; // number of training samples float **x; // model input, of size n * 784 float **y; // model output, of size n * 10 // fill in x and y here and then call: kann_train_fnn1(ann, 0.001f, 64, 25, 10, 0.1f, n, x, y);We can save the model to a file with
kann_save()or use it to classify a MNIST image:
c float *x; // of size 784 const float *y; // this will point to an array of size 10 // fill in x here and then call: y = kann_apply1(ann, x);
Working with complex models requires to use low-level APIs. Please see 01user.md for details.
This example learns to count the number of "1" bits in an integer (i.e. popcount): ```c // to compile and run: gcc -O2 this-prog.c kann.c kautodiff.c -lm && ./a.out
int main(void) { int i, k, maxbit = 20, nsamples = 30000, mask = (1<>1); x[i] = (float)calloc(max_bit, sizeof(float)); y[i] = (float)calloc(maxbit + 1, sizeof(float)); for (k = c = 0; k < maxbit; ++k) x[i][k] = (float)(a>>k&1), c += (a>>k&1); y[i][c] = 1.0f; // c is ranged from 0 to maxbit inclusive } // train kanntrainfnn1(ann, 0.001f, 64, 50, 10, 0.1f, nsamples, x, y); // predict x1 = (float*)calloc(maxbit, sizeof(float)); for (i = nerr = 0; i < nsamples; ++i) { int c, a = kadrand(0) & (mask>>1); // generating a new number const float *y1; for (k = c = 0; k < maxbit; ++k) x1[k] = (float)(a>>k&1), c += (a>>k&1); y1 = kannapply1(ann, x1); for (k = 0, maxk = -1, max = -1.0f; k <= maxbit; ++k) // find the max if (max < y1[k]) max = y1[k], maxk = k; if (maxk != c) ++nerr; } fprintf(stderr, "Test error rate: %.2f%%\n", 100.0 * nerr / nsamples); kanndelete(ann); // TODO: also to free x, y and x1 return 0; } ```
First of all, this benchmark only evaluates relatively small networks, but in practice, it is huge networks on GPUs that really demonstrate the true power of mainstream deep learning frameworks. Please don't read too much into the table.
"Linux" has 48 cores on two Xeno E5-2697 CPUs at 2.7GHz. MKL, NumPy-1.12.0 and Theano-0.8.2 were installed with Conda; Keras-1.2.2 installed with pip. The official TensorFlow-1.0.0 wheel does not work with Cent OS 6 on this machine, due to glibc. This machine has one Tesla K40c GPU installed. We are using by CUDA-7.0 and cuDNN-4.0 for training on GPU.
"Mac" has 4 cores on a Core i7-3667U CPU at 2GHz. MKL, NumPy and Theano came with Conda, too. Keras-1.2.2 and Tensorflow-1.0.0 were installed with pip. On both machines, Tiny-DNN was acquired from github on March 1st, 2017.
mnist-mlp implements a simple MLP with one layer of 64 hidden neurons. mnist-cnn applies two convolutional layers with 32 3-by-3 kernels and ReLU activation, followed by 2-by-2 max pooling and one 128-neuron dense layer. mul100-rnn uses two GRUs of size 160. Both input and output are 2-D binary arrays of shape (14,2) -- 28 GRU operations for each of the 30000 training samples.
|Task |Framework |Machine|Device |Real |CPU |Command line | |:----------|:------------|:------|--------:|--------:|-------:|:------------| |mnist-mlp |KANN+SSE |Linux |1 CPU | 31.3s | 31.2s |mlp -m20 -v0| | | |Mac |1 CPU | 27.1s | 27.1s || | |KANN+BLAS |Linux |1 CPU | 18.8s | 18.8s || | |Theano+Keras |Linux |1 CPU | 33.7s | 33.2s |keras/mlp.py -m20 -v0| | | | |4 CPUs | 32.0s |121.3s || | | |Mac |1 CPU | 37.2s | 35.2s || | | | |2 CPUs | 32.9s | 62.0s || | |TensorFlow |Mac |1 CPU | 33.4s | 33.4s |tensorflow/mlp.py -m20| | | | |2 CPUs | 29.2s | 50.6s |tensorflow/mlp.py -m20 -t2| | |Tiny-dnn |Linux |1 CPU | 2m19s | 2m18s |tiny-dnn/mlp -m20| | |Tiny-dnn+AVX |Linux |1 CPU | 1m34s | 1m33s || | | |Mac |1 CPU | 2m17s | 2m16s || |mnist-cnn |KANN+SSE |Linux |1 CPU |57m57s |57m53s |mnist-cnn -v0 -m15| | | | |4 CPUs |19m09s |68m17s |mnist-cnn -v0 -t4 -m15| | |Theano+Keras |Linux |1 CPU |37m12s |37m09s |keras/mlp.py -Cm15 -v0| | | | |4 CPUs |24m24s |97m22s || | | | |1 GPU |2m57s | |keras/mlp.py -Cm15 -v0| | |Tiny-dnn+AVX |Linux |1 CPU |300m40s |300m23s |tiny-dnn/mlp -Cm15| |mul100-rnn |KANN+SSE |Linux |1 CPU |40m05s |40m02s |rnn-bit -l2 -n160 -m25 -Nd0| | | | |4 CPUs |12m13s |44m40s |rnn-bit -l2 -n160 -t4 -m25 -Nd0| | |KANN+BLAS |Linux |1 CPU |22m58s |22m56s |rnn-bit -l2 -n160 -m25 -Nd0| | | | |4 CPUs |8m18s |31m26s |rnn-bit -l2 -n160 -t4 -m25 -Nd0| | |Theano+Keras |Linux |1 CPU |27m30s |27m27s |rnn-bit.py -l2 -n160 -m25| | | | |4 CPUs |19m52s |77m45s ||
In the single thread mode, Theano is about 50% faster than KANN probably due to efficient matrix multiplication (aka.
sgemm) implemented in MKL. As is shown in a previous micro-benchmark, MKL/OpenBLAS can be twice as fast as the implementation in KANN.
KANN can optionally use the
sgemmroutine from a BLAS library (enabled by macro
HAVE_CBLAS). Linked against OpenBLAS-0.2.19, KANN matches the single-thread performance of Theano on Mul100-rnn. KANN doesn't reduce convolution to matrix multiplication, so MNIST-cnn won't benefit from OpenBLAS. We observed that OpenBLAS is slower than the native KANN implementation when we use a mini-batch of size 1. The cause is unknown.
KANN's intra-batch multi-threading model is better than Theano+Keras. However, in its current form, this model probably won't get alone well with GPUs.