Convolutional neural network to recognize handwritten digits (MNIST dataset).
MNIST Dataset is downloaded from https://s3.amazonaws.com/torch7/data/mnist.t7.tgz which contains images (32x32)
with classes from 1 to 10 divided into 60'000 training images and 10'000 test images.
To use this script you have to download Torch first.
Below you can see example usage of this convolutional neural network.
You can also use this model to predict output for Kaggle Digit Recognizer challenge.
Unfortunately Kaggle MNIST Dataset have different images geometry (28x28) and different file format what makes code a little bit uglier :C
Kaggle dataset consist of 42'000 training images with classes from 0 to 9 and 28'000 images to predict.
In cnn_model.lua you can find exact model which is used in learning.
To use Kaggle model type:
$ th train.lua -kaggle 1 -train_size 42000 -test_size 28000
Tip: after each iteration script saves new predictions into data/submission.csv file.
Max score I have achieved on Kaggle (with about 10 minutes of learning): 0.98686
You can set parameters which will help you to better control the network.
| Parameter | Optional usage | Description | Default |
|---|---|---|---|
| -gpuid | --enable_gpu | Enables CUDA (use only if you have NVIDIA GPU) | -1 |
| -kaggle | --enable_kaggle | Switches to Kaggle challenge prediction | nil |
| -train_size | --traing_data_size | Size of training sets | 60000 |
| -test_size | --test_data_size | Size of testing sets | 10000 |
| -rate | --learning_rate | Learning rate | 0.05 |
| -batch | --batch_size | Numbers of sets in batch | 10 |
| -t | --threads | Number of threads used during usage | 2 |
$ th train.lua --learning_rate 0.01 --batch_size 100
$ th train.lua --learning_rate 0.02 --threads 8 --traing_data_size 2000 --test_data_size 1000
$ th train.lua -kaggle 1 -train_size 42000 -test_size 28000
$ th train.lua -kaggle 1 -train_size 42000 -test_size 28000 --learning_rate 0.005 --batch_size 50