forked from tinygrad/tinygrad
Marcel BischoffandGitHub
e2f833f58f
* checkpoint * fixing pow * undo pow * backward max on GPU and CPU rewrite * indentation * changing seed for curiosity * max replaced equality * undo seed * rebase * fixed tests * merge error
251 lines
7.4 KiB
Python
251 lines
7.4 KiB
Python
import warnings
|
|
import numpy as np
|
|
from .tensor import Function, register
|
|
|
|
# ************* unary ops *************
|
|
|
|
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
|
|
return grad_output * (input >= 0)
|
|
register('relu', ReLU)
|
|
|
|
class Log(Function):
|
|
@staticmethod
|
|
def forward(ctx, input):
|
|
ctx.save_for_backward(input)
|
|
return np.log(input)
|
|
|
|
@staticmethod
|
|
def backward(ctx, grad_output):
|
|
input, = ctx.saved_tensors
|
|
return grad_output / input
|
|
register('log', Log)
|
|
|
|
class Exp(Function):
|
|
@staticmethod
|
|
def forward(ctx, input):
|
|
ret = np.exp(input)
|
|
ctx.save_for_backward(ret)
|
|
return ret
|
|
|
|
@staticmethod
|
|
def backward(ctx, grad_output):
|
|
ret, = ctx.saved_tensors
|
|
return grad_output * ret
|
|
register('exp', Exp)
|
|
|
|
# ************* binary ops *************
|
|
|
|
def unbroadcast(out, in_sh):
|
|
# adjoint operation to broadcast is sum. Need to sum all axis with 1 = in_sh[i] < out.shape[i]
|
|
sum_axis = tuple([i for i in range(len(in_sh)) if in_sh[i]==1 and out.shape[i]>1]) if in_sh != (1,) else None
|
|
return out.sum(axis=sum_axis).reshape(in_sh)
|
|
|
|
class Add(Function):
|
|
@staticmethod
|
|
def forward(ctx, x, y):
|
|
ctx.save_for_backward(x.shape, y.shape)
|
|
return x+y
|
|
|
|
@staticmethod
|
|
def backward(ctx, grad_output):
|
|
shape_x, shape_y = ctx.saved_tensors
|
|
return unbroadcast(grad_output, shape_x), unbroadcast(grad_output, shape_y)
|
|
register('add', Add)
|
|
|
|
class Sub(Function):
|
|
@staticmethod
|
|
def forward(ctx, x, y):
|
|
ctx.save_for_backward(x.shape, y.shape)
|
|
return x-y
|
|
|
|
@staticmethod
|
|
def backward(ctx, grad_output):
|
|
shape_x, shape_y = ctx.saved_tensors
|
|
return unbroadcast(grad_output, shape_x), unbroadcast(-grad_output, shape_y)
|
|
register('sub', Sub)
|
|
|
|
class Mul(Function):
|
|
@staticmethod
|
|
def forward(ctx, x, y):
|
|
ctx.save_for_backward(x, y)
|
|
return x*y
|
|
|
|
@staticmethod
|
|
def backward(ctx, grad_output):
|
|
x,y = ctx.saved_tensors
|
|
return unbroadcast(y*grad_output, x.shape), unbroadcast(x*grad_output, y.shape)
|
|
register('mul', Mul)
|
|
|
|
class Pow(Function):
|
|
@staticmethod
|
|
def forward(ctx, x, y):
|
|
ctx.save_for_backward(x, y)
|
|
return x ** y
|
|
|
|
@staticmethod
|
|
def backward(ctx, grad_output):
|
|
x,y = ctx.saved_tensors
|
|
return unbroadcast(y * (x**(y-1.0)) * grad_output, x.shape), \
|
|
unbroadcast((x**y) * np.log(x) * grad_output, y.shape)
|
|
register('pow', Pow)
|
|
|
|
# ************* reduce ops *************
|
|
|
|
class Sum(Function):
|
|
@staticmethod
|
|
def forward(ctx, input, axis=None):
|
|
ctx.save_for_backward(input, axis)
|
|
return np.array([input.sum()]) if axis is None else input.sum(axis=axis)
|
|
|
|
@staticmethod
|
|
def backward(ctx, grad_output):
|
|
input, axis = ctx.saved_tensors
|
|
axis = [axis] if type(axis) is int else axis
|
|
shape = [1 if axis is None or i in axis else input.shape[i] for i in range(len(input.shape))]
|
|
return grad_output.reshape(shape) + np.zeros_like(input)
|
|
register('sum', Sum)
|
|
|
|
class Max(Function):
|
|
@staticmethod
|
|
def forward(ctx, inp, axis=None):
|
|
axis = [axis] if type(axis) == int else axis
|
|
ret = np.amax(inp, axis=None if axis is None else tuple(axis), keepdims=True)
|
|
ctx.save_for_backward(inp, axis, ret)
|
|
if axis is not None:
|
|
ret = ret.reshape([inp.shape[i] for i in range(len(inp.shape)) if i not in axis])
|
|
return ret
|
|
|
|
@staticmethod
|
|
def backward(ctx, grad_output):
|
|
input, axis, ret = ctx.saved_tensors
|
|
shape = [1 if axis is None or i in axis else input.shape[i] for i in range(len(input.shape))]
|
|
ret2 = (input==ret.reshape(shape))
|
|
div = ret2.sum(axis=None if axis is None else tuple(axis), keepdims=True)
|
|
return ret2*grad_output.reshape(shape)/div
|
|
register('max', Max)
|
|
|
|
# ************* movement ops *************
|
|
|
|
def inner_slice(x, arg):
|
|
padding = [(max(0, -p[0]), max(0, p[1]-x.shape[i])) for i,p in enumerate(arg)]
|
|
x = np.pad(x, padding)
|
|
slicee = [(p[0] + padding[i][0], p[1] + padding[i][0]) for i,p in enumerate(arg)]
|
|
return x[tuple([slice(x[0], x[1], None) for x in slicee])]
|
|
|
|
class Slice(Function):
|
|
@staticmethod
|
|
def forward(ctx, x, arg=None):
|
|
ctx.save_for_backward(x.shape)
|
|
return inner_slice(x, arg)
|
|
|
|
@staticmethod
|
|
def backward(ctx, grad_output):
|
|
shape, = ctx.saved_tensors
|
|
narg = [(0-p[0], grad_output.shape[i]+(shape[i]-p[1])) for i,p in enumerate(ctx.arg)]
|
|
return inner_slice(grad_output, narg)
|
|
register('slice', Slice)
|
|
|
|
class Reshape(Function):
|
|
@staticmethod
|
|
def forward(ctx, x, shape):
|
|
ctx.save_for_backward(x.shape)
|
|
return x.reshape(shape)
|
|
|
|
@staticmethod
|
|
def backward(ctx, grad_output):
|
|
in_shape, = ctx.saved_tensors
|
|
return grad_output.reshape(in_shape)
|
|
register('reshape', Reshape)
|
|
|
|
class Transpose(Function):
|
|
@staticmethod
|
|
def forward(ctx, x, order):
|
|
ctx.save_for_backward(order)
|
|
return np.transpose(x, order)
|
|
|
|
@staticmethod
|
|
def backward(ctx, x):
|
|
return np.transpose(x, np.argsort(ctx.order))
|
|
register('transpose', Transpose)
|
|
|
|
# ************* processing ops *************
|
|
|
|
class Matmul(Function):
|
|
@staticmethod
|
|
def forward(ctx, input, weight):
|
|
ctx.save_for_backward(input, weight)
|
|
return input @ weight
|
|
|
|
@staticmethod
|
|
def backward(ctx, grad_output):
|
|
input, weight = ctx.saved_tensors
|
|
grad_input = grad_output @ np.swapaxes(weight, -2, -1)
|
|
grad_weight = np.swapaxes(input, -2, -1) @ grad_output
|
|
return grad_input, grad_weight
|
|
register('matmul', Matmul)
|
|
|
|
class Conv2D(Function):
|
|
@staticmethod
|
|
def forward(ctx, x, w, stride=1, groups=1):
|
|
if type(ctx.stride) == int:
|
|
ctx.stride = (ctx.stride, ctx.stride)
|
|
cout,cin,H,W = w.shape
|
|
ys,xs = ctx.stride
|
|
bs,cin_ = x.shape[0], x.shape[1]
|
|
oy,ox = (x.shape[2]-(H-ys))//ys, (x.shape[3]-(W-xs))//xs
|
|
assert cin*ctx.groups == cin_
|
|
assert cout % ctx.groups == 0
|
|
rcout = cout//ctx.groups
|
|
|
|
gx = x.reshape(bs,ctx.groups,cin,x.shape[2],x.shape[3])
|
|
tx = np.lib.stride_tricks.as_strided(gx,
|
|
shape=(bs, ctx.groups, cin, oy, ox, H, W),
|
|
strides=(*gx.strides[0:3], gx.strides[3]*ys, gx.strides[4]*xs, *gx.strides[3:5]),
|
|
writeable=False,
|
|
)
|
|
tw = w.reshape(ctx.groups, rcout, cin, H, W)
|
|
ctx.save_for_backward(tx, tw, x.shape)
|
|
|
|
ret = np.zeros((bs,ctx.groups,oy,ox,rcout),dtype=x.dtype)
|
|
for g in range(ctx.groups):
|
|
#ijYXyx,kjyx -> iYXk ->ikYX
|
|
ret[:,g] += np.tensordot(tx[:,g], tw[g], ((1,4,5),(1,2,3)))
|
|
return np.moveaxis(ret,4,2).reshape(bs, cout, oy, ox)
|
|
|
|
@staticmethod
|
|
def backward(ctx, grad_output):
|
|
bs,_,oy,ox = grad_output.shape
|
|
tx, tw, x_shape = ctx.saved_tensors
|
|
_,rcout,cin,H,W = tw.shape
|
|
ys,xs = ctx.stride
|
|
OY,OX = x_shape[2:4]
|
|
|
|
ggg = grad_output.reshape(bs,ctx.groups,rcout,oy,ox)
|
|
|
|
gdw = np.zeros((ctx.groups,rcout,cin,H,W), dtype=tx.dtype)
|
|
for g in range(ctx.groups):
|
|
#'ikYX,ijYXyx -> kjyx'
|
|
gdw[g] += np.tensordot(ggg[:,g], tx[:,g], ((0,2,3),(0,2,3)))
|
|
|
|
# needs to be optimized
|
|
gdx = np.zeros((bs,ctx.groups,cin,OY,OX), dtype=tx.dtype)
|
|
for k in range(oy*ox):
|
|
Y, X = k//ox, k%ox
|
|
iY,iX = Y*ys, X*xs
|
|
#gdx[:,:,: , iY:iY+H, iX:iX+W] += np.einsum('igk,gkjyx->igjyx', ggg[:,:,:,Y,X], tw)
|
|
for g in range(ctx.groups):
|
|
tg = np.dot(ggg[:,g,:,Y,X].reshape(bs, -1), tw[g].reshape(rcout, -1))
|
|
gdx[:, g, :, iY:iY+H, iX:iX+W] += tg.reshape((bs, cin, H, W))
|
|
|
|
return gdx.reshape((bs, ctx.groups*cin, OY, OX)), gdw.reshape((ctx.groups*rcout, cin, H, W))
|
|
register('conv2d', Conv2D)
|
|
|