Create a "datasets" folder in the directory and extract these datasets in it. https://drive.google.com/open?id=1pmsn73Wp7-R1CRfJsY08la7HZCceS_8o
Then open train.py and change the learning_rate and in nn.py change the p variables to use different dropout rates and prevent overfitting. You can also play around with the number of trainnings and number of hidden nodes in each layer to find the best combination.
To create a network run: python train.py <dataset_name> <save_name> <iterations>
dataset_name = Name of the dataset i.e. mnist/letter/balanced/digits/byclass/bymerge
save_name = Nework will saved with this name in the networks folder
iteration = number of iterations for the training
To test a network run: python test.py <dataset_name> <save_name>
dataset_name = Name of the dataset i.e. mnist or letters
save_name = Name of the saved network in the networks folder without the extension
To use your own data run: python pic_to_letter <folder_location>
folder_location = Folder that includes pictures of letters in png format i.e. C:\Users\user\Documents\GitHub\Computer-Vision\examle_letters
Example letters can be found in the example_letters folder.
You can also change the network used for recognition in the pic_to_letter.py file.
To extract individual characters from an image and save as their own file: python find_letter.py <image_name>
image_name = Name of the image file. Must be in the pictures directory.
Example: python find_letter.py test.png