Skip to content

P
pytensor
提交 195e49c7 authored 作者: Frederic Bastien's avatar Frederic Bastien
浏览文件
操作

implemented dx,dy for ConvOp in the not unrolled case.

Make a version of the code clean. better test of dx,dy
上级 4eea8a4b
隐藏空白字符变更
内嵌 并排
正在显示 包含 31 行增加 和 32 行删除
theano/sandbox/conv.py
浏览文件 @ 195e49c7
......@@ -55,12 +55,8 @@ class ConvOp(Op):
else:
print "OPTIMISATION WARNING: in ConvOp.__init__() unroll_kern(%s) should be 0 or a multiple of nkern(%s)We revert it to 1. This won't change the result, but may make it slower."%(str(self.unroll_kern),str(self.nkern))
self.unroll_kern=1
print "OPTIMISATION WARNING: in ConvOp.__init__() unroll_kern(%s) should be 0 or a multiple of nkern(%s)We revert it to 1. This won't change the result, but may make it slower."%(str(self.unroll_kern),str(self.nkern))
if (self.dx!=1 or self.dy!=1) and self.unroll_batch==0 and self.unroll_kern==0:
print "WARNING: dx!=1 or dy!=1 is only supported with unrolling! We will unroll by 1"
self.unroll_kern=1
self.unroll_batch=1
if (self.dx!=1 or self.dy!=1):
print "WARNING: dx(%d)!=1 or dy(%d)!=1. The gradient is not implemented for those case."
self.outshp = getFilterOutShp(self.imshp, kshp, (dx,dy), output_mode)
self.out_mode = output_mode
if not self.out_mode in ["valid", "full"]:
......@@ -95,7 +91,7 @@ class ConvOp(Op):
raise Exception("The image and the kernel must have the same type."
"inputs(%s), kerns(%s)"%(inputs.dtype, kerns.dtype))
output = tensor.tensor(dtype=inputs.type.dtype,
broadcastable=[False]*outdim,
broadcastable=[False]*outdim,
name="ConvOp_Output");
return gof.Apply(self, [inputs, kerns], [output])
......@@ -134,8 +130,8 @@ class ConvOp(Op):
* inputs needs to be a 4D tensor. Couldn't get 3D to work
* will crash if filter the same size as input image
"""
assert self.dx==1 and self.dy==1#We didn't implemented the grad for that case. Can this be done?
if self.dx!=1 or self.dy!=1:
raise NotImplementedError("I don't know how to implement the grad when dx!=1 or dy!=1! Is this possible?")
####### Determine gradient on kernels ########
if inputs.ndim == 3:
......@@ -150,7 +146,7 @@ class ConvOp(Op):
(bsize, nkern) = (self.imshp[0], self.nkern)
imshp = N.hstack((self.bsize, self.imshp[1:]))
kshp = self.outshp
un_b = self.unroll_batch
un_b = self.unroll_batch
un_k = self.unroll_kern
elif self.out_mode == 'full':
(img, filters) = (newgz, newin)
......@@ -245,7 +241,7 @@ using namespace std;
self.unroll_kern)
#TODO: should we choose the unroll size automatically with the bigger divisor under 5?
if self.out_mode == 'valid':
if self.out_mode == 'valid' and self.dx==0 and self.dy==0:
# print "return gemm version"
return _conv_op_code_valid_gemm % d
else:
......@@ -395,8 +391,11 @@ if ((!%(z)s)
}
int Os[2];
if (mode == FULL) {Os[0] = dim_im[0]+dim_ker[0]-1; Os[1] = dim_im[1]+dim_ker[1]-1;}
else {Os[0] = dim_im[0]-dim_ker[0]+1; Os[1] = dim_im[1]-dim_ker[1]+1;}
Os[0]=%(self_outshp0)s;
Os[1]=%(self_outshp1)s;
//I keep the formula to calculte Os in case we need it in the futur.
//if (mode == FULL) {Os[0] = (int)ceil((dim_im[0]+dim_ker[0]-1)/float(%(self_dx)s)); Os[1] = ceil((dim_im[1]+dim_ker[1]-1)/float(%(self_dy)s));}
//else {Os[0] = (int)ceil((dim_im[0]-dim_ker[0]+1)/float(%(self_dx)s)); Os[1] = (int)ceil((dim_im[1]-dim_ker[1]+1)/float(%(self_dy)s));}
for(int b=0;b< %(self_bsize)s;b++){
for(int n_kern=0;n_kern<%(self_nkern)s;n_kern++){
......@@ -417,12 +416,14 @@ for(int b=0;b< %(self_bsize)s;b++){
int new_m;
for (int m=0; m < Os[0]; m++) {
for (int iter_m=0; iter_m < Os[0]; iter_m++) {
// Reposition index into input image based on requested output size
if (mode == FULL) new_m = m ;
else new_m = (m+dim_ker[0]-1);
int pos_m = iter_m*%(self_dx)s;//The position of the patch in the image
if (mode == FULL) new_m = pos_m ;
else new_m = (pos_m+dim_ker[0]-1);
for (int n=0; n < Os[1]; n++) { // loop over columns
for (int iter_n=0; iter_n < Os[1]; iter_n++) { // loop over columns
int pos_n=iter_n*%(self_dy)s;
%(type)s sum=0;
// Sum over kernel, if index into image is out of bounds
......@@ -440,7 +441,7 @@ for(int b=0;b< %(self_bsize)s;b++){
}else{
//do the part where kernel is to the right of the img
int k=0,max_k=max((int)(n-dim_im[1])+1,0);
int k=0,max_k=max((int)(pos_n-dim_im[1])+1,0);
if(fill_value!=0){
for(k=0;k<max_k;k++){
......@@ -449,9 +450,9 @@ for(int b=0;b< %(self_bsize)s;b++){
}else {k=max_k;}
//do the part where the kernel is on the img
max_k=min(n+1,(int)dim_ker[1]);
max_k=min(pos_n+1,(int)dim_ker[1]);
const %(type)s * idx_in=&in[ind0*dim_im[1]];
for (int ind1=n-k; k<max_k; k++,ind1--) {
for (int ind1=pos_n-k; k<max_k; k++,ind1--) {
sum+= idx_hvals[k] * idx_in[ind1];
}
//do the part to the left of the img
......@@ -461,14 +462,13 @@ for(int b=0;b< %(self_bsize)s;b++){
}else{
const %(type)s* idx_in=&in[ind0*dim_im[1]]; //JB: should be dim_im[1] right? (was dim_im[0])
const %(type)s* idx_hvals=&hvals[j*dim_ker[1]];
int new_n = (n+dim_ker[1]-1);
int new_n = (pos_n+dim_ker[1]-1);
for (int k=0,last=new_n; k < dim_ker[1]; k++,last--) {
sum+=idx_hvals[k]*idx_in[last];
}
}
}//for j
out[m*dim_zz[1]+n] %(affectation)s sum;
out[iter_m*dim_zz[1]+iter_n] %(affectation)s sum;
}//for n
}//for m
}//for stack_size
......@@ -770,7 +770,11 @@ if(%(img2d)s->nd==2){
img2d_dim[1]=%(img2d)s->dimensions[1];
img2d_dim[0]=%(img2d)s->dimensions[0];
}else {
PyErr_SetString(PyExc_ValueError, "img don't have a good shape");
std:stringstream temp;
temp << "nddim="<<%(img2d)s->nd;
std::string param = temp.str();
PyErr_SetString(PyExc_ValueError,
("img don't have a good shape. " + param).c_str());
%(fail)s;
}
......@@ -784,11 +788,7 @@ if(%(filtersflipped)s->nd==3){
kerns_dim[1]=%(filtersflipped)s->dimensions[1];
kerns_dim[0]=%(filtersflipped)s->dimensions[0];
}else{
std:stringstream temp;
temp << "nddim="<<%(filtersflipped)s->nd;
std::string param = temp.str();
PyErr_SetString(PyExc_ValueError,
("kernel don't have a good shape. " + param).c_str());
PyErr_SetString(PyExc_ValueError, "kernel don't have a good shape");
%(fail)s;
}
......
theano/sandbox/test_conv.py
浏览文件 @ 195e49c7
......@@ -311,10 +311,10 @@ class TestConvOp(unittest.TestCase):
# fixed parameters
# test multiple configuration at the same time
bsizes = [6,6] # batch size
imshp_starts = [(1,13,14),(1,4,3)]
imshp_starts = [(1,13,14),(1,4,5)]
kshpss = ([[5,6],[7,4]],[[2,2],[2,2]])
nkernss = [[20,40],[2,2]] # per output pixel
ssizess = [[(1,1),(2,2)],[(1,1),(2,2)]]
ssizess = [[(1,1),(1,2)],[(1,1),(2,2)]]
convmodes = ['valid','full']
do_convolve2=True
unroll = [(0,0),(1,1),(2,2),(3,2)]#(batch,kern)
......@@ -417,7 +417,6 @@ class TestConvOp(unittest.TestCase):
d=N.asarray(ntot)/tpytot
print 'speed up py theano(ConvOp) vs convolve2d: %.3fx'%d.mean(),d
def test_ConvOpGrad(self):
"""
test the gradient in float and double
......
Markdown 格式
0% 或
您添加了 0 人 到此讨论。请谨慎行事。
请先完成此评论的编辑！
请 注册 或者 后发表评论