This is an implementation of a Convolutional Neural Network in Rust, which is specifically written to process the raw MNIST dataset, as well as other input images in the same format, and classify those images as digits in the range from 0 to 9. You can download the dataset from that link.
This repository is discussed in this blog post.
This neural network is built from scratch, leveraging only the ndarray crate for linear algebraic operations. The program can also be configured to test against a directory of custom jpeg images, formatted as 28x28 pixel grayscale squares with an inverted colors (255 is black, 0 is white). For example, I tested its accuracy against these:
The cnn-rs binary is a simple CNN with just three layers; a convolutional, max-pooling, and softmax-based fully-connected classification layer.
Based on empirical testing, this model can reach a 98%+ accuracy rate against the MNIST test corpus.
Information about program arguments and flags:
Usage: cnn-rs [OPTIONS]
Options:
--training_images <path> Path to MNIST training image file
--training_labels <path> Path to MNIST image labels file
--test_images <path> Path to MNIST test image file
--test_labels <path> Path to MNIST test labels file
--epochs <num> Number of epochs to run over the training dataset; defaults to 3
--learning_rate <float> Learning rate; defaults to 0.01
--save if true, will save computed model weights to ./cnn.data, overwriting any existing local file
--load <FILE_PATH> path to a previously-written file containing model data
--input_dir <FILE_PATH> path to a directory containing 28x28 px jpeg files
--cheat if true, will use the contents of files under the input_dir argument to fine-tune the model before predicting outputs
-h, --help Print help
-V, --version
If you modify the code, there are quite a few tests you can use to verify behavior:
$ cargo test
...
running 30 tests
test approx::tests::test_equal ... ok
test approx::tests::test_not_equal ... ok
test convolution::tests::test_init_dimensions ... ok
...
test pooling::tests::test_back_propagation_output_shape ... ok
An example run, which:
- trains based on the MNIST training images and label files
- specifies two full training epochs over that training set
- loads previously-trained model data from disk
- overwrites the updated model data to disk
- specifies a learning rate of 1e-6
looks like this:
~/code/cnn-rs ~>> ./target/release/cnn-rs --training_images ./train-images.idx3-ubyte --training_labels ./train-labels.idx1-ubyte --epochs 2 --save --load ./cnn.model --learning_rate 0.000001
loaded model data from ./cnn.model
>> epoch 1 / 2
1000 examples, avg loss 0.0881 accuracy 98.3% avg batch time 781ms
2000 examples, avg loss 0.0732 accuracy 98.6% avg batch time 790ms
3000 examples, avg loss 0.0683 accuracy 98.9% avg batch time 797ms
4000 examples, avg loss 0.0962 accuracy 98.3% avg batch time 792ms
...
60000 examples, avg loss 0.0604 accuracy 98.6% avg batch time 779ms
model data saved to cnn.model
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent