If you are using tutorial from deeplearning.net you might want to use your own set of data to train the model. you can easily create your mnist.pkl.gz dataset from scratch with your own images with this little tool.
This simple script converts your own images to mnist.pkl.gz dataset to train your model for DBN.py from deeplearning.net
step 1: First prepare your images in 3 sets in 3 folders like below:
train
test
valid
step 2: Then make 3 csv file containing your classess. I made them like this:
train.csv
test.csv
valid.csv
step 3: You have to make your images resized and in garyscale format, for doing so use the resize-script.sh script with the following commands:
sudo chmode +x resize-script.sh
./resize-script.shIt checks all images and converts them into grayscale 28x28 png images.
step 4: Finally run the image-to-mnist.py with the following command:
python image-to-mnist.py
I searched lots of websites and read a lot of comments to finnaly found this solution, hope you use it and save you a lot of headache and time.