abhinema/handwritten_digit_recognition_mnist

Handwritten Digit Recognition using MNIST data base

Handwritten Digit Recognition using PyTorch & MNIST database

handwritten digit recognition using mnist data base for training.

Prerequisites

1. Handwritten digit recognition project work done on Ubuntu 18.04 with anaconda package installed.
2. Images used for prediction are generated with white background and black font, refer test_fonts.png. In program images converted to inverted colour for further processing.

Installing

Install components

1. To install pytorch:

conda install pytorch torchvision -c pytorch

2. To install cv2:

conda install -c conda-forge opencv conda install -c conda-forge/label/broken opencv

3. To install matplotlib

conda install -c conda-forge matplotlib conda install -c conda-forge/label/broken matplotlib conda install -c conda-forge/label/testing matplotlib conda install -c conda-forge/label/rc matplotlib

OR use below:

To install dependencies in anaconda envirnment, please use below command: conda create --name --file requirments.txt

Running the tests

1. How to use train.py (This will generate checkpoint model file)

python train.py --no-cuda (use CPU) --epochs= --checkpoint=<checkpoint_file_name.pth>

Example: To use GPU:

python train.py --epochs=10 --checkpoint=./sample_checkpoint_model_epoch10.pth

To use CPU:

python train.py --epochs=10 --checkpoint=./sample_checkpoint_model_epoch10.pth --no-cuda

2. predict.py

python predict.py --input=image= --checkpoint=<checkpoint_file_name.pth>

Input Image File

Example:

Output with model file with 1 epochs

python predict_UserImages.py --input-image=./test_fonts.png --checkpoint=./sample_checkpoint_model_epoch1.pth

Output File

Output with model file with 10 epochs

in above output 7 is predicted as 3. lets try with higher epochs model file.

python predict_UserImages.py --input-image=./test_fonts.png --checkpoint=./sample_checkpoint_model_epoch10.pth

Output File

Here 7 is predicted as 7.

handwritten_digit_recognition_mnist's Issues

should I resize my dataset's input into (28,28)？

Error when running train.py

I met "RuntimeError: output with shape [1, 28, 28] doesn't match the broadcast shape [3, 28, 28]" when running with code
"python train.py --epochs=5 --checkpoint=./checkpoint_model_epoch5.pth" in a Colab environment.
I found train.py can be worked if I change the transform code to "transform = transforms.Compose([transforms.ToTensor(), transforms.Normalize([0.5], [0.5])])", but the "predict_UserImages.py" have an error like:

Could you please kindly help me with this?