Первый прототип нейросети готов

auto_diff/auto_diff.py

@@ -14,6 +14,9 @@ class Node:
 		name = f"{self.label}:" if self.label else ""
 		return f"<{name}{self.op or 'var'} val={self.val:.6g} grad={self.grad:.6g}>"
 	
+	def __float__(self):
+		return self.val
+	
 	@staticmethod
 	def _to_node(x):
 		return x if isinstance(x, Node) else Node(x)
@@ -63,8 +66,10 @@ class Node:
 	
 	def __rtruediv__(self, other): return Node._to_node(other) / self
 
-	def __pow__(self, other):
+	def pow_with_node(self, other):
 		other = Node._to_node(other)
+		if self.val <= 0:
+			raise ValueError("base must be > 0 for real-valued power")
 		out = Node(self.val**other.val, parents=[(self, lambda g: g * other.val * (self.val ** (other.val-1))), (other, lambda g: g * (self.val ** other.val) * math.log(self.val))], op=f"**{other.val}")
 		def _backward():
 			self.grad += out.grad * other.val * (self.val ** (other.val-1))
@@ -72,6 +77,14 @@ class Node:
 		out._backward = _backward
 		return out
 	
+	def __pow__(self, p: float):
+		# степень фиксируем скаляром p
+		out = Node(self.val**p, parents=[(self, lambda g: g * p * (self.val ** (p-1)))], op=f"**{p}")
+		def _backward():
+			self.grad += out.grad * p * (self.val ** (p-1))
+		out._backward = _backward
+		return out
+	
 	def __rpow__(self, p: float):
 		# основание фиксируем скаляром p
 		out = Node(p**self.val, parents=[(self, lambda g: g * p ** self.val * math.log(p))], op=f"{p}**")
@@ -113,14 +126,12 @@ def backward(loss: Node):
 				build(p)
 			topo.append(u)
 	build(loss)
-	print(topo, list(reversed(topo)), visited)
 	# 2) инициализируем dL/dL = 1 и идём в обратном порядке
 	for n in topo:
 		n.grad = 0.0
 	loss.grad = 1.0
 	for node in reversed(topo):
 		node._backward()
-		print(node)
 
 if __name__ == "__main__":
 

classes.py

@@ -1,18 +1,18 @@
 import random
 
 class DataSet:
-	def __init__(self, N=1000) -> None:
+	def __init__(self, func, N=1000) -> None:
 		self.train = []
 		self.train_answs = []
 		self.test = []
 		self.test_answs = []
 
 		for i in range(N//5*4):
-			x = random.uniform(-1000, 1000)
+			x = random.uniform(1, 9)
 			self.train.append(x)
-			self.train_answs.append(x+1)
+			self.train_answs.append(func(x))
 
 		for i in range(N//5):
-			x = random.uniform(-1000, 1000)
+			x = random.uniform(1, 9)
 			self.test.append(x)
-			self.test_answs.append(x+1)
+			self.test_answs.append(func(x))

generate.py

@@ -1,4 +1,7 @@
 import classes
 
 def generate_dataset(N=1000):
-	return classes.DataSet(N)
+	return classes.DataSet(lambda x: 1*x-12, N)
+
+if __name__ == "__main__":
+	print([[i.train, i.train_answs] for i in [generate_dataset(10)]])

main.py

@@ -4,7 +4,7 @@ import visual
 import neuro_defs
 
 
-dataset = generate.generate_dataset(1000)
+dataset = generate.generate_dataset(100)
 
 
 # Создаём и обучаем сеть
@@ -12,8 +12,9 @@ nn = neuro_defs.SimpleNN()
 nn.train(dataset.train, dataset.train_answs, epochs=100)
 
 # Проверяем на новой точке
-for dot in dataset.test[:10]:
-	print(nn.forward(dot), dot)
-
+for dot in range(len(dataset.test)):
+	print(nn.forward(dataset.test[dot]).val, dataset.test_answs[dot])
+print()
+print(nn.w_out.val, nn.b_out.val)
 # visual.plot_dataset(dataset)
 # visual.plt_show()

neuro_defs.py

@@ -1,48 +1,28 @@
 import math
 import random
+from auto_diff import auto_diff
 
-def sigmoid(x):
-	if x >= 0:
-		z = math.exp(-x)
-		return 1 / (1 + z)
-	else:
-		z = math.exp(x)
-		return z / (1 + z)
-
-def sigmoid_derivative(x):
-    s = sigmoid(x)
-    return s * (1 - s)
+def func_active(x):
+	return
 
 class SimpleNN:
 	def __init__(self):
-		# инициализация весов случайными числами
-		self.w1 = random.uniform(-1, 1)
-		self.b = random.uniform(-1, 1)   # смещение
-		self.w_out = random.uniform(-1, 1)
-		self.b_out = random.uniform(-1, 1)
-		self.lr = 0.001   # скорость обучения
+		self.w_out = auto_diff.Node(random.uniform(-1, 1), label="w_out")
+		self.b_out = auto_diff.Node(random.uniform(-1, 1), label="b_out")
+		self.lr = 0.02   # скорость обучения
 
-	def forward(self, x1):
+	def forward(self, x):
 		# прямой проход
-		self.z1 = self.w1 * x1 + self.b
-		self.a1 = sigmoid(self.z1)  # активация скрытого слоя
-		self.z2 = self.w_out * self.a1 + self.b_out
-		self.a2 = sigmoid(self.z2)  # выход сети
-		return self.a2
+		self.z1 = self.w_out * x + self.b_out
+		return self.z1
 
-	def backward(self, x1, y):
+	def backward(self, x, y):
 		# вычисляем ошибку
-		error = self.a2 - y  # dL/da2
+		error = (self.z1 - y)**2  # dL/da2
 
-		# производные для выходного слоя
-		d_out = error * sigmoid_derivative(self.z2)
-		self.w_out -= self.lr * d_out * self.a1
-		self.b_out -= self.lr * d_out
-
-		# производные для скрытого слоя
-		d_hidden = d_out * self.w_out * sigmoid_derivative(self.z1)
-		self.w1 -= self.lr * d_hidden * x1
-		self.b -= self.lr * d_hidden
+		auto_diff.backward(error)
+		self.w_out = auto_diff.Node(float(self.w_out) - self.lr * self.w_out.grad, label="w_out")
+		self.b_out = auto_diff.Node(float(self.b_out) - self.lr * self.b_out.grad, label="b_out")
 
 	def train(self, dataset, answs, epochs=1000):
 		for _ in range(epochs):