Simple Pipeline

This is a encapsulation where it is possible to train different kind of neural networks on the same training logic which you have came up previously. You need 3 things

• torch.nn.Module implemented neural network
• Trainer object which you have to implement extending the provided abstract class
• torch.util.data.DataLoader object to load the data batch wise

You need to implement the above two classes and then create a script where you combine everything together with a dictionary. A sample script is given below.

model = CaltechClassifier(102)

trns = transforms.Compose([transforms.Resize([224, 224]), transforms.RandomHorizontalFlip(), transforms.RandomVerticalFlip(), transforms.ToTensor()])

image_data = ImageFolder('101_ObjectCategories', transform=trns)
dataset_size = {'train': 7000, 'val': 2144}
batch_size = 20
data = {}
data['train'], data['val'] = random_split(image_data, (dataset_size['train'], dataset_size['val']))
loader = {phase: DataLoader(data[phase],batch_size=batch_size) for phase in ['train','val']}

optimizer = Adam(model.parameters())

loss_function = torch.nn.CrossEntropyLoss()

device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')

def training_parameters():
	parameters = {
		'model': model,
		'optimizer': Adam(model.parameters()),
		'scheduler': StepLR(optimizer, step_size=7),
		'loader': loader,
		'loss_function': loss_function,
		'device': device,
		'epochs': 25
	}

	return parameters

if __name__ == '__main__':
	training_parameters()

Now this script can be invoked with our pipeline to encapsulate training.

python Pipeline.py parameters --trainer trainer

Even though the parameter and trainer files are python files, you are supposed to omit the file name. Here Caltech101 dataset has been used as an example, I have implemented a CNN classifier (CaltechClassifier) and the training logic (CaltechTrainer) which we are going to train the model on. Therefore my pipeline call would look like this,

python Pipeline.py parameters --trainer CaltechTrainer

Notice here that I have not changed the parameter files name

To do

• Create Pipeline class
  • Add functionality so that it can load a parameter python file to load all the required modules
  • Can take command line argument of the file name to load the parameters
  • Can generate a skeleton parameter file
• Create sample code
• Pause and Continue training
• Give a list of parameter files to select best model out of them