元算子卷积层实现
元算子卷积层实现
元算子是jittor的关键概念，元算子的层次结构如下所示。
元算子由重索引算子，重索引化简算子和元素级算子组成。重索引算子，重索引化简算子都是一元算子。 重索引算子是其输入和输出之间的一对多映射。重索引简化算子是多对一映射。广播，填补, 切分算子是常见的重新索引算子。 而化简，累乘，累加算子是常见的索引化简算子。元素级算子是元算子的第三部分，与前两个相比，元素算级子可能包含多个输入。元素级算子的所有输入和输出形状必须相同，它们是一对一映射的。 例如，两个变量的加法是一个二进制的逐元素算子。
元算子的层级结构。元算子包含三类算子，重索引算子，重索引化简算子，元素级算子。元算子的反向传播算子还是元算子。元算子可以组成常用的深度学习算子。而这些深度学习算子又可以进一步组成深度学习模型。
上面演示了如何通过三个元算子实现矩阵乘法：
def matmul(a, b):
(n, m), k = a.shape, b.shape[-1]
a = a.broadcast([n,m,k], dims=[2])
b = b.broadcast([n,m,k], dims=[0])
return (a*b).sum(dim=1)
本文将展示如何使用元算子实现卷积。
首先，实现一个朴素的Python卷积：
import numpy as np
import os
def conv_naive(x, w):
N,H,W,C = x.shape
Kh, Kw, _C, Kc = w.shape
assert C==_C, (x.shape, w.shape)
y = np.zeros([N,H-Kh+1,W-Kw+1,Kc])
for i0 in range(N):
for i1 in range(H-Kh+1):
for i2 in range(W-Kw+1):
for i3 in range(Kh):
for i4 in range(Kw):
for i5 in range(C):
for i6 in range(Kc):
if i1-i3<0 or i2-i4<0 or i1-i3>=H or i2-i4>=W: continue
y[i0, i1, i2, i6] += x[i0, i1 + i3, i2 + i4, i5] * w[i3,i4,i5,i6]
return y
下载一个猫的图像，并使用conv_naive实现一个简单的水平滤波器。
# %matplotlib inline
import pylab as pl
img_path="/tmp/cat.jpg"
if not os.path.isfile(img_path):
!wget -O - 'https://upload.wikimedia.org/wikipedia/commons/thumb/4/4f/Felis_silvestris_catus_lying_on_rice_straw.jpg/220px-Felis_silvestris_catus_lying_on_rice_straw.jpg' > $img_path
img = pl.imread(img_path)
pl.subplot(121)
pl.imshow(img)
kernel = np.array([
[-1, -1, -1],
[0, 0, 0],
[1, 1, 1],
])
pl.subplot(122)
x = img[np.newaxis,:,:,:1].astype("float32")
w = kernel[:,:,np.newaxis,np.newaxis].astype("float32")
y = conv_naive(x, w)
print (x.shape, y.shape) # shape exists confusion
pl.imshow(y[0,:,:,0])
naive_conv运作良好。用jittor替换朴素实现。
import jittor as jt
def conv(x, w):
N,H,W,C = x.shape
Kh, Kw, _C, Kc = w.shape
assert C==_C
xx = x.reindex([N,H-Kh+1,W-Kw+1,Kh,Kw,C,Kc], [
'i0', # Nid
'i1+i3', # Hid+Khid
'i2+i4', # Wid+KWid
'i5', # Cid|
])
ww = w.broadcast_var(xx)
yy = xx*ww
y = yy.sum([3,4,5]) # Kh, Kw, c
return y
# Let's disable tuner. This will cause jittor not to use mkl for convolution
jt.flags.enable_tuner = 0
jx = jt.array(x)
jw = jt.array(w)
jy = conv(jx, jw).fetch_sync()
print (jx.shape, jy.shape)
pl.imshow(jy[0,:,:,0])
结果看起来一样。性能如何？
%time y = conv_naive(x, w)
%time jy = conv(jx, jw).fetch_sync()
可以看出jittor的实现要快得多。为什么这两个实现在数学上等效，而jittor的实现运行速度更快？将逐步进行解释：
首先，看一下jt.reindex的帮助文档。
help(jt.reindex)
可以扩展重索引操作，以便更好地理解：
xx = x.reindex([N,H-Kh+1,W-Kw+1,Kh,Kw,C,Kc], [
'i0', # Nid
'i1+i3', # Hid+Khid
'i2+i4', # Wid+KWid
'i5', # Cid
])
ww = w.broadcast_var(xx)
yy = xx*ww
y = yy.sum([3,4,5]) # Kh, Kw, c
扩展后：
shape = [N,H+Kh-1,W+Kw-1,Kh,Kw,C,Kc]
# expansion of x.reindex
xx = np.zeros(shape, x.dtype)
for i0 in range(shape[0]):
for i1 in range(shape[1]):
for i2 in range(shape[2]):
for i3 in range(shape[3]):
for i4 in range(shape[4]):
for i5 in range(shape[5]):
for i6 in range(shape[6]):
if is_overflow(i0,i1,i2,i3,i4,i5,i6):
xx[i0,i1,...,in] = 0
else:
xx[i0,i1,i2,i3,i4,i5,i6] = x[i0,i1+i3,i2+i4,i5]
# expansion of w.broadcast_var(xx)
ww = np.zeros(shape, x.dtype)
for i0 in range(shape[0]):
for i1 in range(shape[1]):
for i2 in range(shape[2]):
for i3 in range(shape[3]):
for i4 in range(shape[4]):
for i5 in range(shape[5]):
for i6 in range(shape[6]):
ww[i0,i1,i2,i3,i4,i5,i6] = w[i3,i4,i5,i6]
# expansion of xx*ww
yy = np.zeros(shape, x.dtype)
for i0 in range(shape[0]):
for i1 in range(shape[1]):
for i2 in range(shape[2]):
for i3 in range(shape[3]):
for i4 in range(shape[4]):
for i5 in range(shape[5]):
for i6 in range(shape[6]):
yy[i0,i1,i2,i3,i4,i5,i6] = xx[i0,i1,i2,i3,i4,i5,i6] * ww[i0,i1,i2,i3,i4,i5,i6]
# expansion of yy.sum([3,4,5])
shape2 = [N,H-Kh+1,W-Kw+1,Kc]
y = np.zeros(shape2, x.dtype)
for i0 in range(shape[0]):
for i1 in range(shape[1]):
for i2 in range(shape[2]):
for i3 in range(shape[3]):
for i4 in range(shape[4]):
for i5 in range(shape[5]):
for i6 in range(shape[6]):
y[i0,i1,i2,i6] += yy[i0,i1,i2,i3,i4,i5,i6]
循环融合后：
shape2 = [N,H-Kh+1,W-Kw+1,Kc]
y = np.zeros(shape2, x.dtype)
for i0 in range(shape[0]):
for i1 in range(shape[1]):
for i2 in range(shape[2]):
for i3 in range(shape[3]):
for i4 in range(shape[4]):
for i5 in range(shape[5]):
for i6 in range(shape[6]):
if not is_overflow(i0,i1,i2,i3,i4,i5,i6):
y[i0,i1,i2,i6] += x[i0,i1+i3,i2+i4,i5] * w[i3,i4,i5,i6]
这是就元算子的优化技巧，它可以将多个算子融合为一个复杂的融合算子，包括许多卷积的变化（例如group conv，separate conv等）。
jittor会尝试将融合算子优化得尽可能快。尝试一些优化（将形状作为常量编译到内核中），并编译到底层的c++内核代码中。
jt.flags.compile_options={"compile_shapes":1}
with jt.profile_scope() as report:
jy = conv(jx, jw).fetch_sync()
jt.flags.compile_options={}
print(f"Time: {float(report[1][4])/1e6}ms")
with open(report[1][1], 'r') as f:
print(f.read())
比之前的实现还要更快！ 从输出中，可以看一看func0的函数定义，这是卷积内核的主要代码，该内核代码是即时生成的。因为编译器知道内核的形状，所以使用了更多的优化方法。
Jittor简单演示了元算子的使用，并不是真正的性能测试，所以使用了比较小的数据规模进行测试，如果需要性能测试，打开jt.flags.enable_tuner = 1，会启动使用专门的硬件库加速。
