Complex Ops
Reduce¤
sum
¤
sum(
axis: Optional[Union[int, Sequence[int]]] = None,
keepdim=False,
acc_dtype: Optional[DTypeLike] = None,
)
Returns the sum of the elements of the tensor along the specified axis or axes.
You can pass in axis
and keepdim
keyword arguments to control the axis along
which the maximum is computed and whether the reduced dimensions are retained.
You can pass in acc_dtype
keyword argument to control the data type of the accumulation.
If not specified, the accumulation data type is chosen based on the input tensor's data type.
t = Tensor.arange(6).reshape(2, 3)
print(t.numpy())
[[0 1 2]
[3 4 5]]
print(t.sum().numpy())
15
print(t.sum(axis=0).numpy())
[3 5 7]
print(t.sum(axis=1).numpy())
[ 3 12]
Source code in tinygrad/tensor.py
1497 1498 1499 1500 1501 1502 1503 1504 1505 1506 1507 1508 1509 1510 1511 1512 1513 1514 1515 1516 1517 1518 1519 1520 1521 1522 |
|
prod
¤
prod(
axis: Optional[Union[int, Sequence[int]]] = None,
keepdim=False,
acc_dtype: Optional[DTypeLike] = None,
)
Returns the product of the elements of the tensor along the specified axis or axes.
You can pass in axis
and keepdim
keyword arguments to control the axis along
which the maximum is computed and whether the reduced dimensions are retained.
You can pass in acc_dtype
keyword argument to control the data type of the accumulation.
If not specified, the accumulation data type is chosen based on the input tensor's data type.
t = Tensor([-1, -2, -3, 1, 2, 3]).reshape(2, 3)
print(t.numpy())
[[-1 -2 -3]
[ 1 2 3]]
print(t.prod().numpy())
-36
print(t.prod(axis=0).numpy())
[-1 -4 -9]
print(t.prod(axis=1).numpy())
[-6 6]
Source code in tinygrad/tensor.py
1524 1525 1526 1527 1528 1529 1530 1531 1532 1533 1534 1535 1536 1537 1538 1539 1540 1541 1542 1543 1544 1545 1546 1547 1548 |
|
max
¤
Returns the maximum value of the tensor along the specified axis or axes.
You can pass in axis
and keepdim
keyword arguments to control the axis along
which the maximum is computed and whether the reduced dimensions are retained.
t = Tensor([[1, 0, 2], [5, 4, 3]])
print(t.numpy())
[[1 0 2]
[5 4 3]]
print(t.max().numpy())
5
print(t.max(axis=0).numpy())
[5 4 3]
print(t.max(axis=1, keepdim=True).numpy())
[[2]
[5]]
Source code in tinygrad/tensor.py
1550 1551 1552 1553 1554 1555 1556 1557 1558 1559 1560 1561 1562 1563 1564 1565 1566 1567 1568 1569 1570 1571 |
|
min
¤
Returns the minimum value of the tensor along the specified axis or axes.
You can pass in axis
and keepdim
keyword arguments to control the axis along
which the minimum is computed and whether the reduced dimensions are retained.
t = Tensor([[1, 0, 2], [5, 4, 3]])
print(t.numpy())
[[1 0 2]
[5 4 3]]
print(t.min().numpy())
0
print(t.min(axis=0).numpy())
[1 0 2]
print(t.min(axis=1, keepdim=True).numpy())
[[0]
[3]]
Source code in tinygrad/tensor.py
1573 1574 1575 1576 1577 1578 1579 1580 1581 1582 1583 1584 1585 1586 1587 1588 1589 1590 1591 1592 1593 1594 1595 |
|
any
¤
Tests if any element evaluates to True
along the specified axis or axes.
You can pass in axis
and keepdim
keyword arguments to control the reduce axis and whether the reduced dimensions are retained.
t = Tensor([[True, True], [True, False], [False, False]])
print(t.numpy())
[[ True True]
[ True False]
[False False]]
print(t.any().numpy())
True
print(t.any(axis=0).numpy())
[ True True]
print(t.any(axis=1, keepdim=True).numpy())
[[ True]
[ True]
[False]]
Source code in tinygrad/tensor.py
1597 1598 1599 1600 1601 1602 1603 1604 1605 1606 1607 1608 1609 1610 1611 1612 1613 1614 1615 1616 1617 |
|
all
¤
Tests if all element evaluates to True
along the specified axis or axes.
You can pass in axis
and keepdim
keyword arguments to control the reduce axis and whether the reduced dimensions are retained.
t = Tensor([[True, True], [True, False], [False, False]])
print(t.numpy())
[[ True True]
[ True False]
[False False]]
print(t.all().numpy())
False
print(t.all(axis=0).numpy())
[False False]
print(t.all(axis=1, keepdim=True).numpy())
[[ True]
[False]
[False]]
Source code in tinygrad/tensor.py
1619 1620 1621 1622 1623 1624 1625 1626 1627 1628 1629 1630 1631 1632 1633 1634 1635 1636 1637 1638 1639 |
|
mean
¤
Returns the mean value of the tensor along the specified axis or axes.
You can pass in axis
and keepdim
keyword arguments to control the axis along
which the mean is computed and whether the reduced dimensions are retained.
Tensor.manual_seed(42)
t = Tensor.normal(2, 3, mean=2.5, std=0.5)
print(t.numpy())
[[2.9889 2.7339 2.7763]
[2.3356 2.0722 2.6376]]
print(t.mean().numpy())
2.5907674
print(t.mean(axis=0).numpy())
[2.6623 2.4031 2.707 ]
print(t.mean(axis=1).numpy())
[2.833 2.3485]
Source code in tinygrad/tensor.py
1641 1642 1643 1644 1645 1646 1647 1648 1649 1650 1651 1652 1653 1654 1655 1656 1657 1658 1659 1660 1661 1662 1663 1664 1665 |
|
var
¤
Returns the variance of the tensor along the specified axis or axes.
You can pass in axis
, keepdim
, and correction
keyword arguments to control the axis along
which the variance is computed, whether the reduced dimensions are retained, and the Bessel's correction applied.
Tensor.manual_seed(42)
t = Tensor.normal(2, 3, mean=2.5, std=0.5)
print(t.numpy())
[[2.9889 2.7339 2.7763]
[2.3356 2.0722 2.6376]]
print(t.var().numpy())
0.109925404
print(t.var(axis=0).numpy())
[0.2134 0.2189 0.0096]
print(t.var(axis=1).numpy())
[0.0187 0.08 ]
Source code in tinygrad/tensor.py
1667 1668 1669 1670 1671 1672 1673 1674 1675 1676 1677 1678 1679 1680 1681 1682 1683 1684 1685 1686 1687 1688 1689 1690 1691 |
|
std
¤
Returns the standard deviation of the tensor along the specified axis or axes.
You can pass in axis
, keepdim
, and correction
keyword arguments to control the axis along
which the standard deviation is computed, whether the reduced dimensions are retained, and the Bessel's correction applied.
Tensor.manual_seed(42)
t = Tensor.normal(2, 3, mean=2.5, std=0.5)
print(t.numpy())
[[2.9889 2.7339 2.7763]
[2.3356 2.0722 2.6376]]
print(t.std().numpy())
0.33155
print(t.std(axis=0).numpy())
[0.462 0.4679 0.0981]
print(t.std(axis=1).numpy())
[0.1367 0.2829]
Source code in tinygrad/tensor.py
1693 1694 1695 1696 1697 1698 1699 1700 1701 1702 1703 1704 1705 1706 1707 1708 1709 1710 1711 1712 1713 1714 1715 |
|
std_mean
¤
Calculates the standard deviation and mean over the dimensions specified by dim.
Syntactic sugar around Tensor.std
and Tensor.mean
to match torch.std_mean
.
Tensor.manual_seed(42)
t = Tensor.normal(2, 3, mean=2.5, std=0.5)
print(t.numpy())
[[2.9889 2.7339 2.7763]
[2.3356 2.0722 2.6376]]
std, mean = t.std_mean()
print(std.numpy(), mean.numpy())
0.33155 2.5907674
Source code in tinygrad/tensor.py
1717 1718 1719 1720 1721 1722 1723 1724 1725 1726 1727 1728 1729 1730 1731 1732 |
|
softmax
¤
softmax(axis=-1, dtype: Optional[DTypeLike] = None)
Applies the softmax function to the tensor along the specified axis.
Rescales the elements of the tensor such that they lie in the range [0, 1] and sum to 1.
You can pass in the axis
keyword argument to control the axis along which the softmax is computed.
Tensor.manual_seed(42)
t = Tensor.randn(2, 3)
print(t.numpy())
[[ 0.9779 0.4678 0.5526]
[-0.3288 -0.8555 0.2753]]
print(t.softmax().numpy())
[[0.4436 0.2664 0.29 ]
[0.2924 0.1727 0.5349]]
print(t.softmax(axis=0).numpy())
[[0.787 0.7897 0.5689]
[0.213 0.2103 0.4311]]
Source code in tinygrad/tensor.py
1740 1741 1742 1743 1744 1745 1746 1747 1748 1749 1750 1751 1752 1753 1754 1755 1756 1757 1758 1759 1760 1761 |
|
log_softmax
¤
log_softmax(axis=-1, dtype: Optional[DTypeLike] = None)
Applies the log-softmax function to the tensor along the specified axis.
The log-softmax function is a numerically stable alternative to the softmax function in log space.
You can pass in the axis
keyword argument to control the axis along which the log-softmax is computed.
Tensor.manual_seed(42)
t = Tensor.randn(2, 3)
print(t.numpy())
[[ 0.9779 0.4678 0.5526]
[-0.3288 -0.8555 0.2753]]
print(t.log_softmax().numpy())
[[-0.8127 -1.3228 -1.238 ]
[-1.2297 -1.7564 -0.6256]]
print(t.log_softmax(axis=0).numpy())
[[-0.2396 -0.2361 -0.564 ]
[-1.5463 -1.5594 -0.8414]]
Source code in tinygrad/tensor.py
1763 1764 1765 1766 1767 1768 1769 1770 1771 1772 1773 1774 1775 1776 1777 1778 1779 1780 1781 1782 1783 1784 |
|
logsumexp
¤
logsumexp(axis=None, keepdim=False)
Computes the log-sum-exp of the tensor along the specified axis or axes.
The log-sum-exp function is a numerically stable way to compute the logarithm of the sum of exponentials.
You can pass in axis
and keepdim
keyword arguments to control the axis along
which the log-sum-exp is computed and whether the reduced dimensions are retained.
Tensor.manual_seed(42)
t = Tensor.randn(2, 3)
print(t.numpy())
[[ 0.9779 0.4678 0.5526]
[-0.3288 -0.8555 0.2753]]
print(t.logsumexp().numpy())
2.1347282
print(t.logsumexp(axis=0).numpy())
[1.2174 0.7039 1.1167]
print(t.logsumexp(axis=1).numpy())
[1.7906 0.9009]
Source code in tinygrad/tensor.py
1786 1787 1788 1789 1790 1791 1792 1793 1794 1795 1796 1797 1798 1799 1800 1801 1802 1803 1804 1805 1806 1807 1808 1809 1810 1811 |
|
logcumsumexp
¤
logcumsumexp(axis=0)
Computes the log-cumsum-exp of the tensor along the specified axis or axes.
The log-cumsum-exp function is a numerically stable way to compute the logarithm of the cumulative sum of exponentials.
You can pass in the axis
keyword argument to control the axis along which
the log-cum-sum-exp is computed.
Tensor.manual_seed(42)
t = Tensor.randn(2, 3)
print(t.numpy())
[[ 0.9779 0.4678 0.5526]
[-0.3288 -0.8555 0.2753]]
print(t.logcumsumexp().numpy())
[[0.9779 0.4678 0.5526]
[1.2174 0.7039 1.1167]]
print(t.logcumsumexp(axis=0).numpy())
[[0.9779 0.4678 0.5526]
[1.2174 0.7039 1.1167]]
print(t.logcumsumexp(axis=1).numpy())
[[ 0.9779 1.4481 1.7906]
[-0.3288 0.1353 0.9009]]
Source code in tinygrad/tensor.py
1813 1814 1815 1816 1817 1818 1819 1820 1821 1822 1823 1824 1825 1826 1827 1828 1829 1830 1831 1832 1833 1834 1835 1836 1837 1838 |
|
argmax
¤
argmax(axis=None, keepdim=False)
Returns the indices of the maximum value of the tensor along the specified axis.
You can pass in axis
and keepdim
keyword arguments to control the axis along
which the maximum is computed and whether the reduced dimensions are retained.
t = Tensor([[1, 0, 2], [5, 4, 3]])
print(t.numpy())
[[1 0 2]
[5 4 3]]
print(t.argmax().numpy()) # Returns the index of the maximum value in the flattened tensor.
3
print(t.argmax(axis=0).numpy()) # Returns the indices of the maximum values along axis 0.
[1 1 1]
print(t.argmax(axis=1).numpy()) # Returns the indices of the maximum values along axis 1.
[2 0]
Source code in tinygrad/tensor.py
1840 1841 1842 1843 1844 1845 1846 1847 1848 1849 1850 1851 1852 1853 1854 1855 1856 1857 1858 1859 1860 1861 1862 1863 1864 1865 |
|
argmin
¤
argmin(axis=None, keepdim=False)
Returns the indices of the minimum value of the tensor along the specified axis.
You can pass in axis
and keepdim
keyword arguments to control the axis along
which the minimum is computed and whether the reduced dimensions are retained.
t = Tensor([[1, 0, 2], [5, 4, 3]])
print(t.numpy())
[[1 0 2]
[5 4 3]]
print(t.argmin().numpy()) # Returns the index of the minimum value in the flattened tensor.
1
print(t.argmin(axis=0).numpy()) # Returns the indices of the minimum values along axis 0.
[0 0 0]
print(t.argmin(axis=1).numpy()) # Returns the indices of the minimum values along axis 1.
[1 2]
Source code in tinygrad/tensor.py
1867 1868 1869 1870 1871 1872 1873 1874 1875 1876 1877 1878 1879 1880 1881 1882 1883 1884 1885 1886 1887 1888 |
|
Processing¤
avg_pool2d
¤
avg_pool2d(
kernel_size=(2, 2),
stride=None,
dilation=1,
padding=0,
count_include_pad=True,
)
Applies average pooling over a tensor.
Note
unlike PyTorch, this implementation is not limited to only 2d pooling and instead works for any number of dimensions.
See: https://paperswithcode.com/method/average-pooling
t = Tensor.arange(25).reshape(1, 1, 5, 5)
print(t.avg_pool2d().numpy())
[[[[ 3. 5.]
[13. 15.]]]]
print(t.avg_pool2d(padding=1).numpy())
[[[[ 0. 0.75 1.75]
[ 3.75 9. 11. ]
[ 8.75 19. 21. ]]]]
Source code in tinygrad/tensor.py
2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021 2022 2023 2024 |
|
max_pool2d
¤
max_pool2d(
kernel_size=(2, 2), stride=None, dilation=1, padding=0
)
Applies max pooling over a tensor.
Note
unlike PyTorch, this implementation is not limited to only 2d pooling and instead works for any number of dimensions.
See: https://paperswithcode.com/method/max-pooling
t = Tensor.arange(25).reshape(1, 1, 5, 5)
print(t.max_pool2d().numpy())
[[[[ 6 8]
[16 18]]]]
print(t.max_pool2d(padding=1).numpy())
[[[[ 0 2 4]
[10 12 14]
[20 22 24]]]]
Source code in tinygrad/tensor.py
2026 2027 2028 2029 2030 2031 2032 2033 2034 2035 2036 2037 2038 2039 2040 2041 2042 2043 |
|
conv2d
¤
conv2d(
weight: Tensor,
bias: Optional[Tensor] = None,
groups=1,
stride=1,
dilation=1,
padding: int | Tuple[int, ...] = 0,
acc_dtype: Optional[DTypeLike] = None,
) -> Tensor
Applies a convolution over a tensor with a given weight
and optional bias
.
Note
unlike PyTorch, this implementation is not limited to only 2d convolutions and instead works for any number of dimensions.
See: https://pytorch.org/docs/stable/generated/torch.nn.Conv2d.html
t = Tensor.arange(9).reshape(1, 1, 3, 3)
w = Tensor.ones(1, 1, 2, 2)
print(t.conv2d(w).numpy())
[[[[ 8. 12.]
[20. 24.]]]]
Source code in tinygrad/tensor.py
2045 2046 2047 2048 2049 2050 2051 2052 2053 2054 2055 2056 2057 2058 2059 2060 2061 2062 2063 2064 2065 2066 2067 2068 2069 2070 2071 2072 2073 2074 2075 2076 2077 2078 2079 2080 2081 2082 2083 2084 2085 2086 2087 2088 2089 2090 2091 2092 2093 2094 2095 2096 2097 2098 2099 2100 2101 2102 2103 2104 2105 2106 |
|
conv_transpose2d
¤
conv_transpose2d(
weight: Tensor,
bias: Optional[Tensor] = None,
groups=1,
stride=1,
dilation=1,
padding=0,
output_padding=0,
) -> Tensor
Applies a transposed convolution over a tensor with a given weight
and optional bias
.
Note
unlike PyTorch, this implementation is not limited to only 2d transposed convolutions and instead works for any number of dimensions.
See: https://pytorch.org/docs/stable/generated/torch.nn.ConvTranspose2d.html
t = Tensor.arange(9).reshape(1, 1, 3, 3)
w = Tensor.ones(1, 1, 2, 2)
print(t.conv_transpose2d(w).numpy())
[[[[ 0. 1. 3. 2.]
[ 3. 8. 12. 7.]
[ 9. 20. 24. 13.]
[ 6. 13. 15. 8.]]]]
Source code in tinygrad/tensor.py
2108 2109 2110 2111 2112 2113 2114 2115 2116 2117 2118 2119 2120 2121 2122 2123 2124 2125 2126 2127 2128 2129 2130 2131 2132 |
|
dot
¤
Performs dot product between two tensors.
If w
is 1-D, it's a sum product over the last axis of self
and w
.
If w
is N-D with N>=2, it's a sum product over the last axis of self
and the second-to-last axis of w
.
You can pass in the optional acc_dtype
keyword argument to control the data type of the accumulation.
a = Tensor([1, 2, 3])
b = Tensor([1, 1, 0])
print(a.dot(b).numpy())
3
a = Tensor([[1, 2], [3, 4]])
b = Tensor([[5, 6], [7, 8]])
print(a.dot(b).numpy())
[[19 22]
[43 50]]
Source code in tinygrad/tensor.py
2134 2135 2136 2137 2138 2139 2140 2141 2142 2143 2144 2145 2146 2147 2148 2149 2150 2151 2152 2153 2154 2155 2156 2157 2158 2159 2160 |
|
matmul
¤
Performs matrix multiplication between two tensors.
You can pass in the reverse
keyword argument to control the order of the matrix multiplication.
You can pass in the optional acc_dtype
keyword argument to control the data type of the accumulation.
a = Tensor([[1, 2], [3, 4]])
b = Tensor([[5, 6], [7, 8]])
print(a.matmul(b).numpy())
[[19 22]
[43 50]]
Source code in tinygrad/tensor.py
2162 2163 2164 2165 2166 2167 2168 2169 2170 2171 2172 2173 2174 2175 |
|
einsum
staticmethod
¤
einsum(
formula: str,
*operands: Tensor | Sequence[Tensor],
acc_dtype: Optional[DTypeLike] = None
) -> Tensor
Sums the product of the elements of the input tensors according to a formula based on the Einstein summation convention.
See: https://pytorch.org/docs/stable/generated/torch.einsum.html
x = Tensor([[1, 2], [3, 4]])
y = Tensor([[5, 6], [7, 8]])
print(Tensor.einsum("ij,ij->", x, y).numpy())
70
Source code in tinygrad/tensor.py
1930 1931 1932 1933 1934 1935 1936 1937 1938 1939 1940 1941 1942 1943 1944 1945 1946 1947 1948 1949 1950 1951 1952 1953 1954 1955 1956 1957 1958 1959 1960 1961 1962 1963 1964 1965 1966 1967 1968 1969 1970 1971 1972 1973 |
|
cumsum
¤
Computes the cumulative sum of the tensor along the specified axis
.
t = Tensor.ones(2, 3)
print(t.numpy())
[[1. 1. 1.]
[1. 1. 1.]]
print(t.cumsum(1).numpy())
[[1. 2. 3.]
[1. 2. 3.]]
Source code in tinygrad/tensor.py
2196 2197 2198 2199 2200 2201 2202 2203 2204 2205 2206 2207 2208 |
|
cummax
¤
Computes the cumulative max of the tensor along the specified axis
.
t = Tensor([0, 1, -1, 2, -2, 3, -3])
print(t.numpy())
[ 0 1 -1 2 -2 3 -3]
print(t.cummax(0).numpy())
[0 1 1 2 2 3 3]
Source code in tinygrad/tensor.py
2210 2211 2212 2213 2214 2215 2216 2217 2218 2219 2220 2221 2222 |
|
triu
¤
Returns the upper triangular part of the tensor, the other elements are set to 0.
The argument diagonal
determines which diagonal is on the boundary. diagonal = 0
means the main diagonal.
Positive diagonal
means above the main diagonal, and negative diagonal
means below the main diagonal.
t = Tensor([[1, 2, 3, 4], [5, 6, 7, 8], [9, 10, 11, 12]])
print(t.numpy())
[[ 1 2 3 4]
[ 5 6 7 8]
[ 9 10 11 12]]
print(t.triu(diagonal=0).numpy())
[[ 1 2 3 4]
[ 0 6 7 8]
[ 0 0 11 12]]
print(t.triu(diagonal=1).numpy())
[[ 0 2 3 4]
[ 0 0 7 8]
[ 0 0 0 12]]
print(t.triu(diagonal=-1).numpy())
[[ 1 2 3 4]
[ 5 6 7 8]
[ 0 10 11 12]]
Source code in tinygrad/tensor.py
2234 2235 2236 2237 2238 2239 2240 2241 2242 2243 2244 2245 2246 2247 2248 2249 2250 2251 2252 2253 2254 2255 |
|
tril
¤
Returns the lower triangular part of the tensor, the other elements are set to 0.
The argument diagonal
determines which diagonal is on the boundary. diagonal = 0
means the main diagonal.
Positive diagonal
means above the main diagonal, and negative diagonal
means below the main diagonal.
t = Tensor([[1, 2, 3, 4], [5, 6, 7, 8], [9, 10, 11, 12]])
print(t.numpy())
[[ 1 2 3 4]
[ 5 6 7 8]
[ 9 10 11 12]]
print(t.tril(diagonal=0).numpy())
[[ 1 0 0 0]
[ 5 6 0 0]
[ 9 10 11 0]]
print(t.tril(diagonal=1).numpy())
[[ 1 2 0 0]
[ 5 6 7 0]
[ 9 10 11 12]]
print(t.tril(diagonal=-1).numpy())
[[ 0 0 0 0]
[ 5 0 0 0]
[ 9 10 0 0]]
Source code in tinygrad/tensor.py
2257 2258 2259 2260 2261 2262 2263 2264 2265 2266 2267 2268 2269 2270 2271 2272 2273 2274 2275 2276 2277 2278 |
|
interpolate
¤
Downsamples or Upsamples to the input size
, accepts 0 to N batch dimensions.
The interpolation algorithm is selected with mode
which currently only supports linear
, nearest
and nearest-exact
.
To run bilinear
or trilinear
, pass in a 2D or 3D size.
t = Tensor([[1, 2, 3, 4], [21, 22, 23, 24], [41, 42, 43, 44]])
print(t.numpy())
[[ 1 2 3 4]
[21 22 23 24]
[41 42 43 44]]
print(t.interpolate(size=(2,3), mode="linear").numpy())
[[ 6 7 8]
[36 37 38]]
Source code in tinygrad/tensor.py
2280 2281 2282 2283 2284 2285 2286 2287 2288 2289 2290 2291 2292 2293 2294 2295 2296 2297 2298 2299 2300 2301 2302 2303 2304 2305 2306 2307 2308 2309 2310 |
|
scatter
¤
scatter(
dim: int,
index: Tensor,
src: Union[Tensor, ConstType],
reduce: Union[
None, Literal["multiply"], Literal["add"]
] = None,
) -> Tensor
Scatters src
values along an axis specified by dim
.
Apply add
or multiply
reduction operation with reduce
.
src = Tensor.arange(1, 11).reshape(2, 5)
print(src.numpy())
[[ 1 2 3 4 5]
[ 6 7 8 9 10]]
index = Tensor([[0, 1, 2, 0]])
print(Tensor.zeros(3, 5, dtype=src.dtype).scatter(0, index, src).numpy())
[[1 0 0 4 0]
[0 2 0 0 0]
[0 0 3 0 0]]
index = Tensor([[0, 1, 2], [0, 1, 4]])
print(Tensor.zeros(3, 5, dtype=src.dtype).scatter(1, index, src).numpy())
[[1 2 3 0 0]
[6 7 0 0 8]
[0 0 0 0 0]]
print(Tensor.full((2, 4), 2.0).scatter(1, Tensor([[2], [3]]), 1.23, reduce='multiply').numpy())
[[2. 2. 2.46 2. ]
[2. 2. 2. 2.46]]
print(Tensor.full((2, 4), 2.0).scatter(1, Tensor([[2], [3]]), 1.23, reduce='add').numpy())
[[2. 2. 3.23 2. ]
[2. 2. 2. 3.23]]
Source code in tinygrad/tensor.py
2312 2313 2314 2315 2316 2317 2318 2319 2320 2321 2322 2323 2324 2325 2326 2327 2328 2329 2330 2331 2332 2333 2334 2335 2336 2337 2338 2339 2340 2341 2342 2343 2344 2345 2346 |
|
Neural Network (functional)¤
linear
¤
Applies a linear transformation to self
using weight
and bias
.
See: https://pytorch.org/docs/stable/generated/torch.nn.Linear.html
t = Tensor([[1, 2], [3, 4]])
weight = Tensor([[1, 2], [3, 4]])
bias = Tensor([1, 2])
print(t.linear(weight, bias).numpy())
[[ 8 12]
[16 24]]
Source code in tinygrad/tensor.py
3269 3270 3271 3272 3273 3274 3275 3276 3277 3278 3279 3280 3281 3282 3283 |
|
sequential
¤
Applies a sequence of functions to self
chaining the output of each function to the input of the next.
t = Tensor([1, 2, 3])
print(t.sequential([lambda x: x * 2, lambda x: x + 1]).numpy())
[3 5 7]
Source code in tinygrad/tensor.py
3285 3286 3287 3288 3289 3290 3291 3292 3293 3294 |
|
layernorm
¤
Applies Layer Normalization over a mini-batch of inputs.
- Described: https://paperswithcode.com/method/layer-normalization
- Paper: https://arxiv.org/abs/1607.06450v1
t = Tensor.randn(8, 10, 16) * 2 + 8
print(t.mean().item(), t.std().item())
7.923057556152344 2.0072731971740723
t = t.layernorm()
print(t.mean().item(), t.std().item())
-2.184478153921532e-09 1.0003893375396729
Source code in tinygrad/tensor.py
3296 3297 3298 3299 3300 3301 3302 3303 3304 3305 3306 3307 3308 3309 3310 3311 3312 3313 |
|
batchnorm
¤
batchnorm(
weight: Optional[Tensor],
bias: Optional[Tensor],
mean: Tensor,
invstd: Tensor,
axis: Union[int, Tuple[int, ...]] = 1,
) -> Tensor
Applies Batch Normalization over a mini-batch of inputs.
- Described: https://paperswithcode.com/method/batch-normalization
- Paper: https://arxiv.org/abs/1502.03167
t = Tensor.randn(8, 4, 16, 16) * 2 + 8
print(t.mean().item(), t.std().item())
8.030435562133789 1.9699469804763794
t = t.batchnorm(None, None, t.mean(axis=(0,2,3)), t.var(axis=(0,2,3)).add(1e-5).rsqrt())
print(t.mean().item(), t.std().item())
1.7121278688136954e-06 0.9998164176940918
Source code in tinygrad/tensor.py
3315 3316 3317 3318 3319 3320 3321 3322 3323 3324 3325 3326 3327 3328 3329 3330 3331 3332 3333 3334 3335 3336 |
|
dropout
¤
dropout(p=0.5) -> Tensor
Applies dropout to self
.
Note
dropout is only applied when Tensor.training
is True
.
- Described: https://paperswithcode.com/method/dropout
- Paper: https://jmlr.org/papers/v15/srivastava14a.html
Tensor.manual_seed(42)
t = Tensor.randn(2, 2)
with Tensor.train():
print(t.dropout().numpy())
[[ 0. 2.17 ]
[ 0. -0.1682]]
Source code in tinygrad/tensor.py
3338 3339 3340 3341 3342 3343 3344 3345 3346 3347 3348 3349 3350 3351 3352 3353 3354 3355 |
|
one_hot
¤
Converts self
to a one-hot tensor.
num_classes
defaults to -1, which means num_classes will be inferred as max(self) + 1.
t = Tensor([0, 1, 3, 3, 4])
print(t.one_hot(5).numpy())
[[1 0 0 0 0]
[0 1 0 0 0]
[0 0 0 1 0]
[0 0 0 1 0]
[0 0 0 0 1]]
Source code in tinygrad/tensor.py
3357 3358 3359 3360 3361 3362 3363 3364 3365 3366 3367 3368 3369 |
|
scaled_dot_product_attention
¤
scaled_dot_product_attention(
key: Tensor,
value: Tensor,
attn_mask: Optional[Tensor] = None,
dropout_p: float = 0.0,
is_causal: bool = False,
) -> Tensor
Computes scaled dot-product attention.
self
is the query tensor, key
is the key tensor, and value
is the value tensor.
- Described: https://paperswithcode.com/method/scaled
- Paper: https://arxiv.org/abs/1706.03762v7
q = Tensor.randn(2, 4, 8)
k = Tensor.randn(2, 4, 8)
v = Tensor.randn(2, 4, 8)
print(q.scaled_dot_product_attention(k, v).numpy())
[[[-0.1425 -0.1433 -0.3625 0.8853 -0.3129 1.0271 -0.0019 0.2445]
[-0.7137 0.2617 1.1393 0.692 0.0461 0.1132 0.391 -0.3563]
[ 0.4718 0.6791 0.8956 0.9387 -0.7198 0.753 0.5702 0.2661]
[-1.0183 0.005 0.9208 0.6447 0.2658 0.0411 0.2314 -0.4636]]
[[ 0.2928 -0.3364 -0.1937 -0.0755 -0.6196 -0.7339 0.8431 -0.3794]
[ 0.5915 0.3565 -0.6987 0.241 0.2624 -0.1074 -0.3026 -0.3574]
[ 0.3176 -0.4436 -0.3136 -0.5334 -0.5756 -0.851 0.9595 -0.4201]
[ 0.4378 0.0234 -0.0984 0.4847 -0.3579 -0.3998 0.3781 -0.2338]]]
Source code in tinygrad/tensor.py
3371 3372 3373 3374 3375 3376 3377 3378 3379 3380 3381 3382 3383 3384 3385 3386 3387 3388 3389 3390 3391 3392 |
|
binary_crossentropy
¤
Computes the binary cross-entropy loss between self
and Y
.
See: https://pytorch.org/docs/stable/generated/torch.nn.BCELoss.html
t = Tensor([0.1, 0.9, 0.2])
Y = Tensor([0, 1, 0])
print(t.binary_crossentropy(Y).item())
0.14462155103683472
Source code in tinygrad/tensor.py
3399 3400 3401 3402 3403 3404 3405 3406 3407 3408 3409 3410 3411 |
|
binary_crossentropy_logits
¤
Computes the binary cross-entropy loss between self
and Y
where self
is logits.
See: https://pytorch.org/docs/stable/generated/torch.nn.BCEWithLogitsLoss.html
t = Tensor([-1, 2, -3])
Y = Tensor([0, 1, 0])
print(t.binary_crossentropy_logits(Y).item())
0.16292567551136017
Source code in tinygrad/tensor.py
3413 3414 3415 3416 3417 3418 3419 3420 3421 3422 3423 3424 3425 |
|
sparse_categorical_crossentropy
¤
sparse_categorical_crossentropy(
Y: Tensor,
ignore_index: int = -1,
label_smoothing=0.0,
reduction: ReductionStr = "mean",
) -> Tensor
Computes the sparse categorical cross-entropy loss between self
and Y
.
Note
self
is logits and Y
is the target labels.
NOTE: unlike PyTorch, this function expects the class axis to be -1
See: https://pytorch.org/docs/stable/generated/torch.nn.CrossEntropyLoss.html
t = Tensor([[-1, 2, -3], [1, -2, 3]])
Y = Tensor([1, 2])
print(t.sparse_categorical_crossentropy(Y).item())
0.09391524642705917
Source code in tinygrad/tensor.py
3427 3428 3429 3430 3431 3432 3433 3434 3435 3436 3437 3438 3439 3440 3441 3442 3443 3444 3445 3446 3447 3448 3449 3450 |
|
cross_entropy
¤
cross_entropy(
Y: Tensor,
reduction: ReductionStr = "mean",
label_smoothing: float = 0.0,
) -> Tensor
Compute the cross entropy loss between input logits and target.
Note
self
are logits and Y
are the target labels or class probabilities.
See: https://pytorch.org/docs/stable/generated/torch.nn.functional.cross_entropy.html
t = Tensor([[-1, 2, -3], [1, -2, 3]])
Y = Tensor([1, 2])
print(t.cross_entropy(Y).item())
0.09391524642705917
t = Tensor([[-1, 2, -3], [1, -2, 3]])
Y = Tensor([1, 2])
print(t.cross_entropy(Y, reduction='none').numpy())
[0.055 0.1328]
Source code in tinygrad/tensor.py
3452 3453 3454 3455 3456 3457 3458 3459 3460 3461 3462 3463 3464 3465 3466 3467 3468 3469 3470 3471 3472 3473 3474 3475 |
|
nll_loss
¤
nll_loss(
Y: Tensor,
weight: Optional[Tensor] = None,
ignore_index: Optional[int] = None,
reduction: ReductionStr = "mean",
) -> Tensor
Compute the negative log likelihood loss between log-probabilities and target labels.
Note
self
is log-probabilities and Y
is the Y labels or class probabilities.
See: https://pytorch.org/docs/stable/generated/torch.nn.functional.nll_loss.html
t = Tensor([[-1, 2, -3], [1, -2, 3]])
Y = Tensor([1, 2])
print(t.log_softmax().nll_loss(Y).item())
0.09391524642705917
t = Tensor([[-1, 2, -3], [1, -2, 3]])
Y = Tensor([1, 2])
print(t.log_softmax().nll_loss(Y, reduction='none').numpy())
[0.055 0.1328]
Source code in tinygrad/tensor.py
3477 3478 3479 3480 3481 3482 3483 3484 3485 3486 3487 3488 3489 3490 3491 3492 3493 3494 3495 3496 3497 3498 3499 |
|