Using anaconda as a distribution of python 2.7 is recommended.

Running the file setup_traning_env.sh will install all the required dependencies to start the training process.

NOTE: I have an commented code for installing pytorch on a CPU, make sure to uncomment it and remove the line above to run code on local machine.

Additionally requirements.txt contains the list of all the dependencies but due to the complexity involved in some packages using the shell scripts is recommended.

Images are 28x28 dimension with 60,000 samples for train and 10,000 for test.

torchvision is the library which is used to download the MNIST data.

Here are parameters that can be passed running the training script:

• --batch, -b : input batch size for training (default: 32)
• --test_batch_size, -test_b : input batch size for testing (default: 512)
• --epochs, -e : number of epochs to train (default: 10)
• --learning_rate, -lr : learning rate (default: 0.01)
• --decay_gamma, -dg : decay parameter to add this amount of decay (default:0.1)
• --decay, -d :how many epochs before the learning rate decays (default: 1000)
• --optimizer, -opt : Optimizer of choice (default: sgd)
• --momentum, -m : SGD momentum (default: 0.9)
• --weight_decay, -wd : SGD weight decay (default: 5e-4)
• --log_dir, -log : SGD weight decay (default: 5e-4)
• --resume, -r : Checkpoint file name. It ends with .ckpt (default: None)
• --help, -h: Get a help menu.

To kick off training: python train.py with any of the above paramters.

To evaluate the performance there are two tools that were built:

This can be seen in the terminal after kicking off the model training.

Using tensorboard: tensorboard --logdir runs/ in terminal and go to http://localhost:6006.

Currently the performance is measured by the loss and accuracy of train and accuracy of test.

There is a graph of the network also which is interactive to step through each block in the Network.

NOTE: Using tensorboard all the different runs' accuracy can be compared.

Running the file setup_infer_env.sh will install all the required dependencies.

Uses onnx and Caffe2 to create prodcution models.

NOTE: This approach is relatively new with pytorch and is being added as a module in the next release of pytorch 1.0

I have some sample data in the data folder that is not used for training. (./data/infer_data)

I also froze a trained model by:

                                python create_prod_model.py -m {THE NAME OF THE MODEL}

Run inference of new images:

                                python inference.py

NOTE: Make sure the input images are of 28x28.