CS144 Project: Simple neural net that recognizes handwriting.
I've recently started using Emojis in a lot of my commit messages, READMEs, and wherever I can on GitHub. Here's what each emoji I use represents. It may be subject to change:
- 🎉 - Commits that include something so amazing that I have to celebrate 🎉
- 🔧 - Commits that are relatively small to medium in size
⚠️ - Commits that introduce configs or code that break things
Thank you to Joseph Redmon for his data sets.
Inside the repo are a couple of data sets provided by Joseph Redmon. You can
run the neural network by running build/project and piping one of the .csv
files located in
data/.
The test data set contains 10,000 labeled inputs while the training data set contains 60,000 labeled inputs.
To build the project, run:
$ makeYou can also run make clean or make lint.
The built binary is located at build/project. You can get a help message by running:
$ build/project -h
Usage: build/project [-v|-d|-l] < data/mnist_test.csv
Flags:
-v - Enable verbose output.
-d - Dump network weights after training.
-l - Load network weights from previous training.The neural network uses the file weights.data in the current directory for
dumping and loading to and from a file. When verbose is enabled, extra messages
during training and network prediction matching are printed.
I've already taken the liberty of generating the weights and saving them to
files for later use. In order to use these weights, you'll have to
copy/move/symlink the files to weights.data in your current directory:
$ cp weights/test.data weights.data
$ build/project -v -l < data/mnist_test.csvThis will create a copy of the weights. When the network is run with the -l
flag, the weights will be fed into the network and training will be skipped
unless the weights file can't be read or parsed.
The weights can end up being unparseable if some data fails to write to the file or if the network parameters or test data are changed, but the weights aren't.
The following results were computed inside an Arch Linux virtual machine running in VMWare Workstation 14. The VM has 1 vCPU and 2 vCores. My host desktop has an i7-5820K overclocked to 4.2GHz. Thus, computation times may vary.
$ cp weights/test.data weights.data
$ build/project -l < data/mnist_test.csv
Neural Network Stats:
Matches: 9129 / 10000 (91.29%)
Parsing time: 674ms
Training time: 841ms
Matching time: 52030ms$ cp weights/test.data weights.data
$ build/project -l < data/mnist_train.csv
Neural Network Stats:
Matches: 53447 / 60000 (89.08%)
Parsing time: 3831ms
Training time: 694ms
Matching time: 320944ms$ cp weights/train.data weights.data
$ build/project -l < data/mnist_test.csv
Neural Network Stats:
Matches: 9422 / 10000 (94.22%)
Parsing time: 641ms
Training time: 730ms
Matching time: 51796ms$ cp weights/train.data weights.data
$ build/project -l < data/mnist_train.csv
Neural Network Stats:
Matches: 56940 / 60000 (94.90%)
Parsing time: 3806ms
Training time: 1027ms
Matching time: 317144msMIT License
Copyright (c) 2017 Jeremy Asuncion
Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.