Lots of stuff WIP

.gitignore

@@ -6,6 +6,7 @@ src/wandb
 # Files to ignore
 scratch.py
 data/mnist_py3_deskewed.pkl.gz
+data/mnist_py3_mini_deskewed.pkl.gz
 networks/*
 !networks/humans.txt
 

digits.py

@@ -3,6 +3,7 @@
 import numpy as np
 import wandb
 
+from src.gui import gui
 from src.network import NeuralNetwork
 from src.utils import deskew_data, load_data
 
@@ -13,34 +14,38 @@ def main():
 		if x.lower() == 'y':
 			deskew_data()
 
+	# TODO: Refactor these functions to make sense
 	training, validation, test = load_data()
-	train([784, 128, 10], 0.008, 0.2, 0.05, training, validation)
-
+	wandb.init(project='digits')
+	n = train([784, 128, 10], 0.008, 0.2, 0.05, training, validation)
+	try:
+		gui(n)
+	except:
+		wandb.run.save()
 
-# root = tk.Tk()
-# InputGUI(root, n)
-# root.mainloop()
 
 def train(size, eta, lmbda, alpha, training, validation):
-	wandb.init(project='digits')
 	n = NeuralNetwork(size, eta=eta, lmbda=lmbda, alpha=alpha)
 	n.train(
 		np.random.permutation(training)[:5000],
-		np.random.permutation(validation)[:500], epochs=20, batch_size=20
+		np.random.permutation(validation)[:500], epochs=5, batch_size=20
 	)
 
 	data_dir = os.path.join(os.getcwd(), 'networks')
 	i = get_save_index(data_dir)
-	n.save(os.path.join(data_dir, f'{i}.json'))
-	n.plot(os.path.join(data_dir, f'{i}.png'))
-	wandb.save(os.path.join(data_dir, f'{i}.json'))
-	wandb.save(os.path.join(data_dir, f'{i}.png'))
-	wandb.run.save()
+	data_dir = os.path.join(data_dir, str(i))
+	# TODO: Save to run folder
+	#  os.path.join(wandb.run.dir, '')
+	n.save(os.path.join(data_dir, 'network.json'))
+	n.plot(os.path.join(data_dir, 'accuracy.png'))
+	wandb.save(os.path.join(data_dir, 'network.json'))
+	wandb.save(os.path.join(data_dir, 'accuracy.png'))
+	return n
 
 
 def get_save_index(data_dir):
 	i = 1
-	while os.path.isfile(os.path.join(data_dir, f'{i}.json')):
+	while os.path.isdir(os.path.join(data_dir, str(i))):
 		i += 1
 	return i
 

src/gui.py

@@ -4,7 +4,7 @@
 import numpy as np
 
 from preprocessor import deskew_image, dots_to_image
-from utils import draw_digit
+from utils import draw_digit, save_digit
 
 
 class InputGUI:
@@ -56,16 +56,25 @@ def clear(self):
 		self.dots = []
 
 	def predict(self):
+		# TODO: Create new directory inside the run folder for this prediction
 		data = dots_to_image(self.dots, self.scale)
-		# draw_digit(data)
-
+		# save_digit(data, 'raw.png')
 		deskewed = deskew_image(data)
-		draw_digit(deskewed)
+		# TODO: Save images in prediction folder
+		# save_digit(data, 'deskewed.png')
+
+		# draw_digit(data)
+		# draw_digit(deskewed)
 
 		if self.n:
-			prediction = self.n.predict(data)
+			# TODO: Save raw as well as deskewed prediction as json
+			# prediction = self.n.predict(data.reshape((784,)))
+			prediction = self.n.predict(deskewed.reshape((784,)))
 			digit = np.argmax(prediction)
-			print(prediction, digit, prediction[digit])
+			print('Prediction: %s, confidence: %d%%' % (
+				digit, prediction[digit] * 100
+			))
+			print(prediction)
 
 	def draw(self, event):
 		x, y = event.x, event.y
@@ -78,9 +87,9 @@ def draw(self, event):
 			)
 
 
-def gui():
+def gui(n=None):
 	root = tk.Tk()
-	InputGUI(root)
+	InputGUI(root, n)
 	root.mainloop()
 
 

src/utils.py

@@ -8,7 +8,7 @@
 from preprocessor import deskew_image
 
 
-def load_data(deskew=True):
+def load_data(mini=True, deskew=True):
 	"""
 	Loads MNIST dataset from disk.
 
@@ -20,8 +20,11 @@ def load_data(deskew=True):
 	# The _py3 version of the dataset is a redumped version for Python 3
 	# which doesn't use Python 2's latin1 encoding
 	data_dir = os.path.join(os.getcwd(), 'data')
-	file_path = os.path.join(data_dir, 'mnist_py3_deskewed.pkl.gz') if deskew \
+	file_path = (
+		os.path.join(data_dir, 'mnist_py3_mini_deskewed.pkl.gz') if mini
+		else os.path.join(data_dir, 'mnist_py3_deskewed.pkl.gz') if deskew
 		else os.path.join(data_dir, 'mnist_py3.pkl.gz')
+	)
 
 	if not os.path.isfile(file_path):
 		raise FileNotFoundError(f'{file_path} not found!')
@@ -56,15 +59,25 @@ def deskew_data():
 	data_path = os.path.join(data_dir, 'mnist_py3.pkl.gz')
 	with gzip.open(data_path, 'rb') as f:
 		data = pickle.load(f)
-	processed_data = []
 
+	processed_data = []
 	for section in data:
-		xs = [
-			list(deskew_image(x.reshape((28, 28))).reshape(784, ))
-			for x in section[0]
-		]
+		xs = []
+		for x in section[0]:
+			xs.append(list(deskew_image(x.reshape((28, 28))).reshape(784, )))
 		processed_data.append((xs, section[1]))
 
+	# deskew_mini_path = os.path.join(data_dir,
+	# 'mnist_py3_mini_deskewed.pkl.gz')
+	# processed_data_mini = [
+	# 	(processed_data[0][0][:5000], processed_data[0][1][:5000]),
+	# 	(processed_data[0][0][:1000], processed_data[0][1][:1000]),
+	# 	(processed_data[0][0][:1000], processed_data[0][1][:1000])
+	# ]
+	# with gzip.open(deskew_mini_path, 'wb') as f:
+	# A protocol of -1 means the latest one
+	# pickle.dump(processed_data_mini, f, protocol=-1)
+
 	with gzip.open(deskew_path, 'wb') as f:
 		# A protocol of -1 means the latest one
 		pickle.dump(processed_data, f, protocol=-1)
@@ -87,6 +100,14 @@ def draw_digit(image):
 	Image.fromarray(arr).resize((256, 256), Image.ANTIALIAS).show()
 
 
+def save_digit(image, file_path):
+	"""
+	Renders and displays the image to disk.
+	"""
+	arr = (np.array(image).reshape((28, 28)) * 255).astype('uint8')
+	Image.fromarray(arr).resize((256, 256), Image.ANTIALIAS).save(file_path)
+
+
 def main():
 	deskew_data()