commit 1bb258350092defd802cf6fbb94b1e8de96935cc
Author: George Hotz <geohot@gmail.com>
Date:   Sat Oct 17 22:57:01 2020 -0700

    start tinygrad

diff --git a/.gitignore b/.gitignore
new file mode 100644
index 0000000000..a783405e0c
--- /dev/null
+++ b/.gitignore
@@ -0,0 +1,3 @@
+__pycache__
+notebooks
+.*.swp
diff --git a/README b/README
new file mode 100644
index 0000000000..9da75a69c9
--- /dev/null
+++ b/README
@@ -0,0 +1,4 @@
+For something in between a grad and a karpathy/micrograd
+
+Requires numpy
+
diff --git a/tensor.py b/tensor.py
new file mode 100644
index 0000000000..4ca1312332
--- /dev/null
+++ b/tensor.py
@@ -0,0 +1,76 @@
+# inspired by https://github.com/karpathy/micrograd/blob/master/micrograd/engine.py
+from functools import partialmethod
+import numpy as np
+
+# **** start with three base classes ****
+
+class Context:
+  def __init__(self):
+    self.saved_tensors = []
+
+  def save_for_backward(self, *x):
+    self.saved_tensors.extend(x)
+
+class Tensor:
+  def __init__(self, data, _children=()):
+    self.data = data
+    self.grad = np.zeros_like(data)
+
+    # internal variables used for autograd graph construction
+    self._prev = set(_children)
+
+class Function:
+  def apply(self, arg, *x):
+    ctx = Context()
+    x = [self]+list(x)
+    ret = Tensor(arg.forward(ctx, *[t.data for t in x]))
+    return ret
+
+def register(name, fxn):
+  setattr(Tensor, name, partialmethod(fxn.apply, fxn))
+
+# **** implement a few functions ****
+    
+"""
+class ReLU(Function):
+  @staticmethod
+  def forward(ctx, input):
+    ctx.save_for_backward(input)
+    return np.maximum(input, 0)
+
+  @staticmethod
+  def backward(ctx, grad_output):
+    input, = ctx.saved_tensors
+    grad_input = grad_output.clone()
+    grad_input[input < 0] = 0
+    return grad_input
+setattr(Tensor, 'relu', partialmethod(run, ReLU))
+"""
+
+class Dot(Function):
+  @staticmethod
+  def forward(ctx, input, weight):
+    ctx.save_for_backward(input, weight)
+    return input.dot(weight)
+
+  @staticmethod
+  def backward(ctx, grad_output):
+    input, weight = ctx.saved_tensors
+    grad_input = grad_output.dot(weight.T)
+    grad_weight = grad_output.dot(input)
+    return grad_input, grad_weight
+register('dot', Dot)
+
+# may be wrong
+class Sum(Function):
+  @staticmethod
+  def forward(ctx, input):
+    ctx.save_for_backward(input)
+    return input.sum()
+
+  @staticmethod
+  def backward(ctx, grad_output):
+    input = ctx.saved_tensors
+    return grad_output * input
+register('sum', Sum)
+
diff --git a/test.py b/test.py
new file mode 100644
index 0000000000..c324238566
--- /dev/null
+++ b/test.py
@@ -0,0 +1,10 @@
+import torch
+from tensor import Tensor
+
+x = np.random.randn(1,3)
+W = np.random.randn(3,3)
+out = x.dot(W)
+print(out)
+
+
+