main.py

import sys
import json
sys.path.extend(["./src/"])
from tools import get_args, error_exit
from preprocess import preprocess
from visualize import visualize
from neural_network import *
# from validation_metrics import *

def main():
	
	try:	
		args = get_args()

		data = preprocess(args)

		if args.visualize_data:
			visualize(data)
			sys.exit(1) 

		train_set, test_set = split(data)
		
		num_examples = train_set.shape[0]
		num_features = train_set.shape[1] - 1
		if args.mini_batch:
			batch_size = 32		# or 64
			epochs = 1500
		else:
			batch_size = num_examples
			epochs = 30000

		nn = NeuralNetwork(num_features, batch_size, epochs)

		if args.train:
			nn.train(data, train_set, test_set, num_examples, args.quiet)

			if args.evaluation:
				y_pred = probability_to_class( nn.output.T)
				get_validation_metrics(y_pred[:, 0],  nn.y.T[:, 0])

			# mini-batch learning is noisy, so we don't plot it 
			if not args.mini_batch:
				plot_learning(nn.train_losses, nn.test_losses)

			# save network params
			if args.save_model:
				W1, W2, W3, W4 =  nn.weights1.tolist(),  nn.weights2.tolist(),  nn.weights3.tolist(),  nn.weights4.tolist()
				B1, B2, B3, B4 =  nn.bias1.tolist(),  nn.bias2.tolist(),  nn.bias3.tolist(),  nn.bias4.tolist()
				model = dict(weights1=W1, weights2=W2, weights3=W3, weights4=W4, bias1=B1, bias2=B2, bias3=B3, bias4=B4)
				with open("model.json", "w") as f:
					json.dump(model, f, separators=(',', ':'), indent=4)

		if args.predict and (args.predict == "model.json"):
			try:
				with open(args.predict) as file:
					model = json.load(file)
			except: 		
				error_exit("please provide a valid model")
			nn.load_model(model)
			nn.predict(test_set, epochs)

	except:
		pass

if __name__ == '__main__':
	main()


# to visualize the data:
# python3 main.py data.csv -v

# to train the model:
# python3 main.py data.csv -t -s {optional -b -e -q}

# to load a trained model and test:
# python3 main.py data.csv -p model.json