Remove tensor prefix

pygrad/nn.py

@@ -1,38 +1,43 @@
 # Neural networks from scratch with numpy.
 
 import numpy as np
-import pygrad.tensor as tensor
+from pygrad.tensor import tensor, Tensor
 
 def mean_absolute_error(x, y):
   return np.mean(np.abs(x - y))
 
-def mean_squared_error(x, y):
+def mean_squared_error(x: Tensor, y: Tensor):
-  return np.mean(np.power(x - y, 2))
+  return x.sub(y).expt(2).div(tensor([[2.0]]))
 
-def cross_entropy_loss(x, y):
+def cross_entropy_loss(x: Tensor, y: Tensor):
-  return -np.log(np.exp(y) / np.sum(np.exp(x)))
+  return y.exp().div(np.sum(x.exp())).log().neg()
 
 # prepare inputs and outputs
-x = np.array([[1, 0]])
+x = tensor(np.array([[1, 0]]))
-y = np.array([[1]])
+y = tensor(np.array([[1]]))
 
 # we're doing xavier initialisation - see <http://proceedings.mlr.press/v9/glorot10a/glorot10a.pdf>
-w1 = np.random.randn(2, 3) / np.sqrt(2)
+w1 = tensor(np.random.randn(2, 3) / np.sqrt(2))
-w2 = np.random.randn(3, 1) / np.sqrt(3)
+w2 = tensor(np.random.randn(3, 1) / np.sqrt(3))
 
 def single_pass():
+  global w1, w2
+
   # forward pass
-  h = np.matmul(x, w1)
+  h = x.mul(w1)
-  h_hat = np.tanh(h)
+  h_hat = h.tanh()
-  j = np.matmul(h_hat, w2)
+  j = h_hat.mul(w2)
-  print("prediction {}".format(j))
+  print(f"prediction {j}")
 
   # loss calculation
-  loss = cross_entropy_loss(j, y)
+  loss = mean_squared_error(j, y)
-  print("loss {}".format(loss))
+  print(f"loss {loss}")
 
-  # TODO Backward pass.
+  loss.backward()
-  return
+  print(w1.grad, w2.grad)
+
+  w1.value -= 0.1 * w1.grad
+  w2.value -= 0.1 * w2.grad
 
 # initialise layers
 #   self.lin1 = nn.Linear(2, 3)
@@ -48,7 +53,7 @@ def single_pass():
 # Small test to see if autograd works.
 def test():
   # Input tensors.
-  x, y, z = tensor.Tensor(np.array([[1, 2, 3]])), tensor.Tensor(np.array([[2, 3, 4]])), tensor.Tensor(np.array([[1], [2], [3]]))
+  x, y, z = Tensor(np.array([[1, 2, 3]])), Tensor(np.array([[2, 3, 4]])), Tensor(np.array([[1], [2], [3]]))
   
   # Forward pass.
   q = x.add(y)

pygrad/tensor.py

@@ -54,7 +54,7 @@ class Tensor:
 
     def back(upstream):
       a, b = tensor._parents
-      return np.dot(b.value, upstream), np.dot(a.value.T, upstream)
+      return np.dot(upstream, b.value.T), np.dot(a.value.T, upstream)
 
     tensor._back = back
     return tensor
@@ -131,7 +131,7 @@ class Tensor:
     def back(upstream):
       # dtanh(x)/dx = 1 - tanh2(x)
       a, = tensor._parents
-      return [1 - np.dot(np.tanh(a.value) ** 2, upstream)]
+      return [np.ones_like(self.value) - np.dot(upstream, (np.tanh(a.value) ** 2).T)]
 
     tensor._back = back
     return tensor