Skip to content

P
pytensor
提交 baf12f54 authored 作者: Frederic's avatar Frederic
浏览文件
操作

Advance the new GpuConv

上级 50677701
隐藏空白字符变更
内嵌 并排
正在显示 包含 1088 行增加 和 3 行删除
theano/sandbox/gpuarray/conv.py 0 → 100644
浏览文件 @ baf12f54
import theano
from theano import gof
class GpuConv(gof.Op):
"""
Implement the batched and stacked 2d convolution on the gpu.
"""
@staticmethod
def logical_output_shape_2d(imshp, kshp, mode):
if mode == 'valid':
return imshp[0] - kshp[0] + 1, imshp[1] - kshp[1] + 1
if mode == 'full':
return imshp[0] + kshp[0] - 1, imshp[1] + kshp[1] - 1
raise ValueError(mode)
def __init__(self, border_mode,
subsample=(1, 1),
logical_img_hw=None,
logical_kern_hw=None,
logical_kern_align_top=True,
version=-1,
verbose=0,
kshp=None,
imshp=None,
max_threads_dim0=None):
"""
:param version: each version of c_code implement many kernel for the
convolution. By default we try to guess the best one.
You can force one version with this parameter. This
parameter is used by the tests.
:param verbose: for value of 1,2 and 3. Print more information during
the execution of the convolution. Mostly used for
optimization or debugging.
:param kshp: The size of the kernel. If provided, can genera
faster code. If the GpuConv op is automatically
inserted,
we take its value automatically from the Conv op.
:param imshp: The size of the image. Not used for code generation but
allow to select an experimental new version in another
repo.
:param max_threads_dim0: The maximum number of thread for the
block size dimensions 0 (blockDim.x) used by the
GPU function.
"""
self.border_mode = border_mode
self.subsample = subsample
if logical_img_hw is not None:
h, w = logical_img_hw
#TODO: reconsider this... since shapes are not given in
# constructor, maybe a multiplier + offset is a more
# appropriate way of passing this logical grid
logical_img_hw = tuple(logical_img_hw)
self.logical_img_hw = logical_img_hw
if logical_kern_hw is not None:
h, w = logical_kern_hw
#TODO: reconsider this... since shapes are not given in
# constructor, maybe a multiplier + offset is a more
# appropriate way of passing this logical grid
logical_kern_hw = tuple(logical_kern_hw)
self.logical_kern_hw = logical_kern_hw
self.logical_kern_align_top = logical_kern_align_top
self.version = version
self.verbose = verbose
self.kshp = kshp
self.imshp = imshp
self.max_threads_dim0 = max_threads_dim0
def __eq__(self, other):
return type(self) == type(other) \
and self.border_mode == other.border_mode \
and self.subsample == other.subsample \
and self.logical_img_hw == other.logical_img_hw \
and self.logical_kern_hw == other.logical_kern_hw \
and self.logical_kern_align_top == other.logical_kern_align_top \
and self.version == other.version \
and self.verbose == other.verbose \
and self.kshp == other.kshp\
and self.imshp == other.imshp\
and self.max_threads_dim0 == other.max_threads_dim0
def __setstate__(self, d):
self.__dict__.update(d)
if not hasattr(self, "imshp"):
self.imshp = None
if not hasattr(self, "max_threads_dim0"):
self.max_threads_dim0 = None
def __hash__(self):
# don't use hash(self.version) as hash(-1)==-2 and
# hash(-2)==-2 in python!
return hash(type(self)) \
^ hash(self.border_mode) \
^ hash(self.subsample) \
^ hash(self.logical_img_hw) \
^ hash(self.logical_kern_hw) \
^ hash(self.logical_kern_align_top) \
^ self.version \
^ hash(self.verbose) \
^ hash(self.kshp)\
^ hash(self.imshp)\
^ hash(self.max_threads_dim0)
def __str__(self):
return '%s{%s, %s, %s, %s, %s, %s, %s}' % (
self.__class__.__name__,
self.border_mode,
str(self.subsample),
str(self.logical_img_hw),
str(self.logical_kern_hw),
str(self.logical_kern_align_top),
str(self.imshp),
str(self.kshp))
def make_node(self, img, kern):
if img.type.ndim != 4:
raise TypeError('img must be 4D tensor')
if kern.type.ndim != 4:
raise TypeError('kern must be 4D tensor')
broadcastable = [img.type.broadcastable[0], kern.type.broadcastable[0],
False, False]
return Apply(self, [img, kern], [CudaNdarrayType(broadcastable)()])
def flops(self, inputs, outputs):
""" Useful with the hack in profilemode to print the MFlops"""
images, kerns = inputs
out, = outputs
assert images[1] == kerns[1]
flops = 0
if self.border_mode == "valid":
# nb mul and add by output pixel
flops = kerns[2] * kerns[3] * 2
# nb flops by output image
flops *= out[2] * out[3]
# nb patch multiplied
flops *= images[1] * kerns[0] * images[0]
else:
flops = (images[0] * kerns[0] * images[1] *
kerns[2] * kerns[3] *
images[2] * images[3] * 2)
return flops
def make_thunk(self, node, storage_map, compute_map, no_recycling):
node_ = copy.copy(node)
assert node.op is node_.op
if node_.op.max_threads_dim0 is None:
cuda = theano.sandbox.cuda
device_id = cuda.use.device_number
if device_id is None:
cuda.use("gpu",
force=False,
default_to_move_computation_to_gpu=False,
move_shared_float32_to_gpu=False,
enable_cuda=False,
test_driver=True)
device_id = cuda.use.device_number
cuda_ndarray = theano.sandbox.cuda.cuda_ndarray.cuda_ndarray
prop = cuda_ndarray.device_properties(device_id)
node_.op.max_threads_dim0 = prop['maxThreadsDim0']
return super(GpuConv, node_.op).make_thunk(node_, storage_map,
compute_map, no_recycling)
def c_compile_args(self):
nb = 0
if self.kshp is not None:
nb = self.kshp[1]
return ['-DTHEANO_KERN_WID=' + str(nb)] # ,'-g','-G']
def c_headers(self):
return ['cuda_ndarray.cuh', '<stdio.h>']
def c_code_cache_version(self):
# raise this whenever modifying any of the support_code_files
return (0, 20)
def c_support_code_apply(self, node, nodename):
# REMEMBER TO RAISE c_code_cache_version when changing any of
# these files
files = ['conv_kernel.cu', 'conv_full_kernel.cu', 'conv.cu']
codes = [open(os.path.join(os.path.split(__file__)[0], f)).read()
for f in files]
return reduce(str.__add__, codes)
def c_code(self, node, nodename, inp, out_, sub):
img, kern = inp
out, = out_
dx = self.subsample[0]
dy = self.subsample[1]
border_mode = self.border_mode
version = self.version
verbose = self.verbose
sub = sub.copy()
max_threads_dim0 = self.max_threads_dim0
if max_threads_dim0 is None:
raise NotImplementedError("GpuConv.c_code should not be called "
"directly. It should be called by "
"make_thunk() that add some information "
"related to the selected GPU.")
sub.update(locals())
return """
//Mandatory args
const char *mode_str = "%(border_mode)s";
//Optional args
int version = %(version)s;
int verbose = %(verbose)s;
int dx = %(dx)s;
int dy = %(dy)s;
int mode;
if (strcmp(mode_str, "full") == 0)
{
mode = ConvMode_FULL;
}
else if (strcmp(mode_str, "valid") == 0)
{
mode = ConvMode_VALID;
}
else
{
PyErr_SetString(PyExc_ValueError,
"mode must be one of 'full' or 'valid'");
return NULL;
}
// TODO, make out be decref before we alloc out2!
CudaNdarray * out2 = (CudaNdarray *)CudaNdarray_Conv(%(img)s, %(kern)s,
%(out)s, mode,
dx, dy,
version, verbose,
%(max_threads_dim0)s);
Py_XDECREF(%(out)s);
%(out)s = out2;
if (%(out)s==NULL){
%(fail)s
}
""" % sub
theano/sandbox/gpuarray/opt.py
浏览文件 @ baf12f54
...@@ -9,8 +9,8 @@ from theano.gof import (local_optimizer, EquilibriumDB, ...@@ -9,8 +9,8 @@ from theano.gof import (local_optimizer, EquilibriumDB,
InconsistencyError, EquilibriumOptimizer) InconsistencyError, EquilibriumOptimizer)
from theano.gof.python25 import all, any from theano.gof.python25 import all, any
from theano.tensor.nnet.conv import ConvOp
from theano.sandbox.gpuarray.type import GpuArrayType from theano.sandbox.gpuarray.type import GpuArrayType
from theano.sandbox.gpuarray.basic_ops import (host_from_gpu, from theano.sandbox.gpuarray.basic_ops import (host_from_gpu,
gpu_from_host, gpu_from_host,
gpu_alloc, gpu_alloc,
...@@ -20,6 +20,7 @@ from theano.sandbox.gpuarray.basic_ops import (host_from_gpu, ...@@ -20,6 +20,7 @@ from theano.sandbox.gpuarray.basic_ops import (host_from_gpu,
GpuReshape, GpuReshape,
GpuEye) GpuEye)
from theano.sandbox.gpuarray.blas import gpu_dot22, GpuGemv, GpuGemm from theano.sandbox.gpuarray.blas import gpu_dot22, GpuGemv, GpuGemm
from theano.sandbox.gpuarray.conv import GpuConv
from theano.sandbox.gpuarray.nnet import (GpuCrossentropySoftmaxArgmax1HotWithBias, from theano.sandbox.gpuarray.nnet import (GpuCrossentropySoftmaxArgmax1HotWithBias,
GpuCrossentropySoftmax1HotWithBiasDx) GpuCrossentropySoftmax1HotWithBiasDx)
from theano.sandbox.gpuarray.elemwise import (GpuElemwise, _is_scalar, from theano.sandbox.gpuarray.elemwise import (GpuElemwise, _is_scalar,
...@@ -372,7 +373,7 @@ def local_gpu_conv(node): ...@@ -372,7 +373,7 @@ def local_gpu_conv(node):
if node.op == gpu_from_host: if node.op == gpu_from_host:
#gpu_from_host(conv) -> gpu_conv(gpu_from_host) #gpu_from_host(conv) -> gpu_conv(gpu_from_host)
host_input = node.inputs[0] host_input = node.inputs[0]
if host_input.owner and isinstance(host_input.owner.op, conv.ConvOp): if host_input.owner and isinstance(host_input.owner.op, ConvOp):
gpu_conv = GpuConvOp_from_ConvOp(host_input.owner.op) gpu_conv = GpuConvOp_from_ConvOp(host_input.owner.op)
if gpu_conv is None: if gpu_conv is None:
return return
...@@ -386,7 +387,7 @@ def local_gpu_conv(node): ...@@ -386,7 +387,7 @@ def local_gpu_conv(node):
# differently then the gpu ConvOp # differently then the gpu ConvOp
return [out] return [out]
if isinstance(node.op, conv.ConvOp): if isinstance(node.op, ConvOp):
#conv(host_from_gpu) -> host_from_gpu(gpu_conv) #conv(host_from_gpu) -> host_from_gpu(gpu_conv)
img, kern = node.inputs img, kern = node.inputs
img_on_gpu = (img.owner and img.owner.op == host_from_gpu) img_on_gpu = (img.owner and img.owner.op == host_from_gpu)
......
theano/sandbox/gpuarray/tests/test_conv_cuda_ndarray.py 0 → 100644
浏览文件 @ baf12f54
"""
Tests for GPU convolution
"""
import sys
import time
import unittest
import numpy
from nose.plugins.skip import SkipTest
imported_scipy_convolve2d = False
try:
from scipy.signal import convolve2d
imported_scipy_convolve2d = True
except ImportError:
pass
import theano
from theano import tensor
from theano.gof.python25 import any
from theano.tests.unittest_tools import seed_rng
# We let that import do the init of the back-end if needed.
from theano.sandbox.gpuarray.tests.test_basic_ops import (mode_with_gpu,
mode_without_gpu)
from theano.sandbox.gpuarray.type import GpuArrayType
import pygpu
gftensor4 = GpuArrayType('float32', [False] * 4)
device_id = theano.sandbox.cuda.use.device_number
# TODO do with with the new back-end.
from theano.sandbox.cuda import cuda_ndarray
cuda_ndarray = theano.sandbox.cuda.cuda_ndarray.cuda_ndarray
device_prop = cuda_ndarray.device_properties(device_id)
def py_conv_valid_numpy(img, kern):
assert img.shape[1] == kern.shape[1]
outshp = (img.shape[0], kern.shape[0],
img.shape[2] - kern.shape[2] + 1,
img.shape[3] - kern.shape[3] + 1)
out = numpy.zeros(outshp, dtype='float32')
for b in xrange(out.shape[0]):
for k in xrange(out.shape[1]):
for rr in xrange(out.shape[2]):
for cc in xrange(out.shape[3]):
#rr, cc is the upper-left corner of img patches
imgpatch = img[b, :, rr:rr + kern.shape[2],
cc:cc + kern.shape[3]]
innerprod = (imgpatch[:, ::-1, ::-1] *
kern[k, :, :, :]).sum()
out[b, k, rr, cc] = innerprod
return out
def py_conv_full_numpy(img, kern):
# manually pad the img with zeros all around, and then run it
# through py_conv_valid
pad_rows = 2 * (kern.shape[2] - 1) + img.shape[2]
pad_cols = 2 * (kern.shape[3] - 1) + img.shape[3]
padded_img = numpy.zeros((img.shape[0], img.shape[1], pad_rows, pad_cols),
dtype=img.dtype)
padded_img[:, :, kern.shape[2] - 1: kern.shape[2] - 1 + img.shape[2],
kern.shape[3] - 1: kern.shape[3] - 1 + img.shape[3]] = img
return py_conv_valid_numpy(padded_img, kern)
def py_conv(img, kern, mode, subsample):
"""
use a scipy or numpy implementation depending is scipy is available.
The scipy version is faster.
"""
if imported_scipy_convolve2d:
return py_conv_scipy(img, kern, mode, subsample)
elif mode == 'valid':
return py_conv_valid_numpy(img, kern)[:, :, ::subsample[0],
::subsample[1]]
elif mode == 'full':
return py_conv_full_numpy(img, kern)[:, :, ::subsample[0],
::subsample[1]]
else:
raise Exception("Can't execute this kernel.")
def py_conv_scipy(img, kern, mode, subsample):
assert img.shape[1] == kern.shape[1]
if mode == 'valid':
outshp = (img.shape[0], kern.shape[0],
img.shape[2] - kern.shape[2] + 1,
img.shape[3] - kern.shape[3] + 1)
else:
outshp = (img.shape[0], kern.shape[0],
img.shape[2] + kern.shape[2] - 1,
img.shape[3] + kern.shape[3] - 1)
out = numpy.zeros(outshp, dtype='float32')
for b in xrange(out.shape[0]):
for k in xrange(out.shape[1]):
for s in xrange(img.shape[1]):
out[b, k, :, :] += convolve2d(img[b, s, :, :],
kern[k, s, :, :],
mode)
return out[:, :, ::subsample[0], ::subsample[1]]
def _params_allgood_header():
print "ishape kshape #Mflops CPU Mflops GPU Mflops Speedup"
def _params_allgood(ishape, kshape, mode, subsample=(1, 1), img_stride=(1, 1),
kern_stride=(1, 1), version=-1, verbose=0, random=True,
print_=None, id=None, rtol=1e-5, atol=1e-8,
nb_iter=0, ones=False, compile_kshp=None):
#
# This function is the core of several of the big unit-test drivers,
# but it can also be used very directly on its own to test a specific
# kind of convolution.
#
# See `test_example` (above) for an example of how to use this directly.
#
# :param kshape: (4d)The shape of the kernel at run time.
# :param compile_kshp: (2d) hardcode the shape of the kernel in
# the generated code This is supposed to be
# faster, but we need to check That we raise
# an error if the input have the wrong shape.
#
if ones:
assert not random
npy_img = theano._asarray(numpy.ones(ishape), dtype='float32')
npy_kern = -theano._asarray(numpy.ones(kshape), dtype='float32')
elif random:
npy_img = theano._asarray(numpy.random.rand(*ishape) + 1,
dtype='float32')
npy_kern = theano._asarray(numpy.random.rand(*kshape) - 2,
dtype='float32')
else:
npy_img = theano._asarray(numpy.arange(
numpy.prod(ishape)).reshape(ishape), dtype='float32') + 1
npy_kern = -(theano._asarray(numpy.arange(
numpy.prod(kshape)).reshape(kshape), dtype='float32') + 1)
img = pygpu.array(npy_img)
kern = pygpu.array(npy_kern)
#we take the stride after the transfert as we make c_contiguous
#data on the GPU.
if img_stride != (1, 1):
img = img[:, :, ::img_stride[0], ::img_stride[1]]
npy_img = npy_img[:, :, ::img_stride[0], ::img_stride[1]]
if kern_stride != (1, 1):
kern = kern[:, :, ::kern_stride[0], ::kern_stride[1]]
npy_kern = npy_kern[:, :, ::kern_stride[0], ::kern_stride[1]]
t2 = None
rval = True
try:
t0 = time.time()
cpuval = py_conv(npy_img, npy_kern, mode, subsample)
t1 = time.time()
i = gftensor4()
k = gftensor4()
op = theano.sandbox.cuda.blas.GpuConv(border_mode=mode,
subsample=subsample,
version=version,
verbose=verbose,
kshp=compile_kshp)(i, k)
f = theano.function([i, k], op, mode=mode_with_gpu)
gpuval = f(img, kern)
t2 = time.time()
for i in range(nb_iter):
gpuval2 = f(img, kern)
assert numpy.allclose(numpy.asarray(gpuval),
numpy.asarray(gpuval2))
assert (numpy.asarray(gpuval) == numpy.asarray(gpuval2)).all()
gpuval = numpy.asarray(gpuval)
if gpuval.shape != cpuval.shape:
print >> sys.stdout, "ERROR: shape mismatch",
print >> sys.stdout, gpuval.shape, cpuval.shape
rval = False
if rval:
rval = numpy.allclose(cpuval, gpuval, rtol=rtol)
assert numpy.all(numpy.isfinite(gpuval))
except NotImplementedError, e:
print >> sys.stdout, '_params_allgood Failed allclose', e
rval = False
if (t2 is not None):
if mode == 'valid':
approx_fp = cpuval.size * ishape[1] * kshape[2] * kshape[3] * 2
else:
approx_fp = (ishape[0] * kshape[0] * kshape[1] * kshape[2] *
kshape[3] * ishape[2] * ishape[3] * 2)
approx_fp /= 1e6
cpu_mflops = approx_fp / (t1 - t0)
gpu_mflops = approx_fp / (t2 - t1)
if verbose > 0:
print >> sys.stdout, '%15s' % str(ishape), '%15s' % str(kshape),
print >> sys.stdout, '%12.5f %7.2f %7.2f %7.1f' % (approx_fp,
cpu_mflops, gpu_mflops, (t1 - t0) / (t2 - t1))
if not rval:
print >> sys.stdout, ('test_' + mode + ' id=' + str(id) +
' FAILED for ishape, kshape, mode, subsample,' +
' img_stride, kern_stride, version', ishape,
kshape, mode, subsample, img_stride, kern_stride,
version)
diff = cpuval - gpuval
diffabs = numpy.absolute(diff)
pr_diff = diffabs / numpy.absolute(cpuval)
nb_close = (diffabs <= (atol + rtol * numpy.absolute(gpuval))).sum()
print "max absolute diff:", (diffabs.max(), "avg abs diff:",
numpy.average(diffabs))
print "median abs diff:", (numpy.median(diffabs), "nb close:",
nb_close, "/", diff.size)
print "max relatif diff:", (pr_diff.max(), "avg rel diff:",
numpy.average(pr_diff))
if not rval and print_ != False:
if npy_img.shape[0] > 5:
print "img", npy_img[0]
print "kern", npy_kern[0]
print "gpu", gpuval[0][0]
print "cpu", cpuval[0][0]
print "diff", diff[0][0]
else:
print "img", npy_img
print "kern", npy_kern
print "gpu", gpuval
print "cpu", cpuval
print "diff", diff
return rval
def exec_conv(version, shapes, verbose, random, mode,
print_=None, rtol=1e-5, ones=False):
if verbose > 0:
_params_allgood_header()
nb_failed = 0
nb_tests = 0
failed_version = set()
failed_id = []
# I put -1 in case we forget to add version in the test to.
for ver in version:
for id, (ishape, kshape, subshape,
istride, kstride) in enumerate(shapes):
ret = False
try:
ret = _params_allgood(ishape,
kshape,
mode,
subsample=subshape,
img_stride=istride,
kern_stride=kstride,
version=ver,
verbose=verbose,
random=random,
id=id,
print_=print_,
rtol=rtol,
ones=ones)
except Exception, e:
print ver, id, (ishape, kshape, subshape, istride, kstride)
print e
pass
if not ret:
failed_version.add(ver)
failed_id.append(id)
nb_failed += 1
nb_tests += 1
if nb_failed > 0:
print "nb_failed", nb_failed, "on", nb_tests,
print "failed_version", failed_version, "failed_id", failed_id
assert nb_failed == 0, nb_failed
else:
print 'Executed', nb_tests, 'different shapes'
def get_basic_shapes():
#basic test of image and kernel shape
return [((1, 1, 1, 1), (1, 1, 1, 1), (1, 1), (1, 1), (1, 1)),
((1, 1, 2, 2), (1, 1, 2, 2), (1, 1), (1, 1), (1, 1)),
((1, 1, 3, 3), (1, 1, 2, 2), (1, 1), (1, 1), (1, 1)),
#basic test for unsquare kernel and image
((1, 1, 2, 4), (1, 1, 2, 2), (1, 1), (1, 1), (1, 1)),
((1, 1, 3, 4), (1, 1, 2, 2), (1, 1), (1, 1), (1, 1)),
((1, 1, 4, 3), (1, 1, 2, 2), (1, 1), (1, 1), (1, 1)),
((1, 1, 4, 4), (1, 1, 3, 2), (1, 1), (1, 1), (1, 1)),
((1, 1, 4, 4), (1, 1, 2, 3), (1, 1), (1, 1), (1, 1))]
def get_shapes(imshp=(1, 1), kshp=(1, 1), subsample=(1, 1),
img_stride=(1, 1), kern_stride=(1, 1)):
""" all possible case if we one or more of stack size, batch size,
nkern. We use the gived image shape, kernel shape and subsmaple
shape."""
return [
#stack only
((1, 2) + imshp, (1, 2) + kshp, subsample, img_stride, kern_stride),
#batch only
((3, 1) + imshp, (1, 1) + kshp, subsample, img_stride, kern_stride),
#nkern only
((1, 1) + imshp, (2, 1) + kshp, subsample, img_stride, kern_stride),
#batch and nkern
((3, 1) + imshp, (2, 1) + kshp, subsample, img_stride, kern_stride),
#batch and stack
((3, 2) + imshp, (1, 2) + kshp, subsample, img_stride, kern_stride),
#stack and nkern
((1, 2) + imshp, (2, 2) + kshp, subsample, img_stride, kern_stride),
#batch, nkern and stack
((2, 2) + imshp, (2, 2) + kshp, subsample, img_stride, kern_stride),
#batch, nkern and stack
((3, 2) + imshp, (4, 2) + kshp, subsample, img_stride, kern_stride)
]
def get_shapes2(scales_img=(1, 1), scales_kern=(1, 1), subsample=(1, 1),
img_stride=(1, 1), kern_stride=(1, 1)):
#basic test of stack, batch and nkern paramter
shapes = get_shapes((1 * scales_img[0], 1 * scales_img[1]),
(1 * scales_kern[0], 1 * scales_kern[1]),
subsample, img_stride, kern_stride)
#basic test of stack, batch and nkern paramter with image and kernel shape
shapes += get_shapes((2 * scales_img[0], 2 * scales_img[1]),
(2 * scales_kern[0], 2 * scales_kern[1]),
subsample, img_stride, kern_stride)
#basic test of stack, batch and nkern paramter with image and kernel shape
shapes += get_shapes((3 * scales_img[0], 3 * scales_img[1]),
(2 * scales_kern[0], 2 * scales_kern[1]),
subsample, img_stride, kern_stride)
#basic test of stack, batch and nkern paramter with not square image.
shapes += get_shapes((4 * scales_img[0], 3 * scales_img[1]),
(2 * scales_kern[0], 2 * scales_kern[1]),
subsample, img_stride, kern_stride)
#basic test of stack, batch and nkern paramter with not square image.
shapes += get_shapes((3 * scales_img[0], 4 * scales_img[1]),
(2 * scales_kern[0], 2 * scales_kern[1]),
subsample, img_stride, kern_stride)
#basic test of stack, batch and nkern paramter with not square kernel.
shapes += get_shapes((4 * scales_img[0], 4 * scales_img[1]),
(3 * scales_kern[0], 2 * scales_kern[1]),
subsample, img_stride, kern_stride)
#basic test of stack, batch and nkern paramter with not square kernel.
shapes += get_shapes((4 * scales_img[0], 4 * scales_img[1]),
(2 * scales_kern[0], 3 * scales_kern[1]),
subsample, img_stride, kern_stride)
return shapes
def get_valid_shapes():
# img shape, kern shape, subsample shape
shapes = get_basic_shapes()
shapes += get_shapes2()
#test image stride
shapes += get_shapes2(scales_img=(2, 2), img_stride=(1, 2))
shapes += get_shapes2(scales_img=(2, 2), img_stride=(2, 1))
shapes += get_shapes2(scales_img=(2, 2), img_stride=(2, 2))
shapes += get_shapes2(scales_img=(2, 2), img_stride=(-1, -1))
shapes += get_shapes2(scales_img=(2, 2), kern_stride=(-1, -1))
#test subsample done in a separate fct
shapes += [
#other test
((2, 1, 2, 2), (1, 1, 2, 2), (1, 1), (1, 1), (1, 1))
, ((3, 2, 4, 4), (4, 2, 4, 4), (1, 1), (1, 1), (1, 1))
, ((4, 1, 10, 10), (1, 1, 2, 2), (1, 1), (1, 1), (1, 1))
, ((1, 1, 4, 4), (1, 1, 2, 3), (1, 1), (1, 1), (1, 1))
, ((4, 1, 10, 10), (1, 1, 2, 3), (1, 1), (1, 1), (1, 1))
, ((4, 1, 10, 10), (1, 1, 2, 10), (1, 1), (1, 1), (1, 1))
, ((4, 1, 20, 10), (1, 1, 2, 10), (1, 1), (1, 1), (1, 1))
, ((3, 2, 8, 8), (4, 2, 4, 4), (1, 1), (1, 1), (1, 1)) #stack, nkern, bsize
, ((3, 2, 8, 6), (4, 2, 4, 4), (1, 1), (1, 1), (1, 1)) #stack, nkern, bsize, non-square image
, ((3, 2, 8, 6), (4, 2, 4, 3), (1, 1), (1, 1), (1, 1)) #stack, nkern, bsize, non-square image, non-square kern
, ((3, 2, 8, 6), (4, 2, 4, 6), (1, 1), (1, 1), (1, 1)) #stack, nkern, bsize ,non-square image, non-square kern, kernsize==imgsize on one dim
, ((16, 5, 64, 64), (8, 5, 8, 8), (1, 1), (1, 1), (1, 1)) # a big one
, ((16, 1, 28, 28), (20, 1, 5, 5), (1, 1), (1, 1), (1, 1)) # MNIST LeNET layer 1
, ((20, 16, 32, 32), (1, 16, 28, 28), (1, 1), (1, 1), (1, 1)) # layer 1 backprop to weights
, ((60,20,28,28), (10,20,5,5), (1, 1), (2,2), (1, 1))#added a test case that fail from test_nnet.py.test_conv_nnet2
, ((10,5,28,28), (10,5,5,5), (1, 1), (2,2), (1, 1))#test precedent but reduced that triger the error
#Test more than maxThreadsDim0
, ((2,4,13,1050), (3,4,10, 11), (1, 1), (1, 1), (1, 1))
, ((2,4,1050,13), (3,4,10, 11), (1, 1), (1, 1), (1, 1))
]
shapes += [ ((60,1,28,28),(20,1,5,5), (1, 1), (1, 1), (1, 1))#test_lenet_28 1 layers
, ((60,20,12,12),(30,20,5,5), (1, 1), (1, 1), (1, 1))#test_lenet_28 2 layers
, ((60,30,8,8),(20,30,5,5), (1, 1), (1, 1), (1, 1))#test_lenet_28 bprop 1 full
, ((20,60,12,12),(30,60,8,8), (1, 1), (1, 1), (1, 1))#test_lenet_28 bprop 2 valid
# , ((1,60,28,28),(20,60,24,24), (1, 1), (1, 1), (1, 1))#test_lenet_28 bprop 2 valid
, ((10,1,64,64),(20,1,7,7), (1, 1), (1, 1), (1, 1))#test_lenet_64 1 layers
, ((10,20,29,29),(30,20,7,7), (1, 1), (1, 1), (1, 1))#test_lenet_64 2 layers
, ((10,30,23,23),(20,30,7,7), (1, 1), (1, 1), (1, 1))#test_lenet_64 full
# , ((20,10,29,29),(30,10,23,23), (1, 1), (1, 1), (1, 1))#test_lenet_64 bprop 1
# , ((1,10,64,64),(20,10,58,58), (1, 1), (1, 1), (1, 1))#test_lenet_64 bprop 2
]
return shapes
def test_valid_0_2():
seed_rng()
shapes = get_valid_shapes()
version = [0, 2]
verbose = 0
random = True
print_ = False
ones = False
if ones:
random = False
shapes2 = []
for id, (ishape, kshape, subshape, istride, kstride) in enumerate(shapes):
oshape = [ishape[0]] + [kshape[0]] + list(numpy.asarray(ishape[2:]) -
numpy.asarray(kshape[2:]) +
numpy.asarray([1, 1]))
if oshape[3] > device_prop['maxThreadsDim0']:
continue
if ishape[1] > 1:
continue
if ((numpy.prod(ishape[2:]) + numpy.prod(kshape[2:])) * 4 >
(16 * 1024 - 150)):
continue
if subshape == (1, 1):
shapes2.append((ishape, kshape, subshape, istride, kstride))
shapes = shapes2
exec_conv(version, shapes, verbose, random, 'valid',
print_=print_, ones=ones, rtol=1.1e-5)
def test_valid_1_3_11_12():
seed_rng()
shapes = get_valid_shapes()
version = [1, 3, 11, 12]
verbose = 0
random = True
print_ = False
ones = False
if ones:
random = False
shapes2 = []
for id, (ishape, kshape, subshape, istride, kstride) in enumerate(shapes):
oshape = [ishape[0]] + [kshape[0]] + list(numpy.asarray(ishape[2:]) -
numpy.asarray(kshape[2:]) +
numpy.asarray([1, 1]))
if oshape[3] > device_prop['maxThreadsDim0']:
continue
if ((numpy.prod(ishape[2:]) + numpy.prod(kshape[2:])) * 4 >
(16 * 1024 - 150)):
continue
if subshape == (1, 1):
shapes2.append((ishape, kshape, subshape, istride, kstride))
shapes = shapes2
exec_conv(version, shapes, verbose, random, 'valid',
print_=print_, ones=ones, rtol=1.1e-5)
def test_valid_4():
seed_rng()
shapes = get_valid_shapes()
version = [4]
verbose = 0
random = True
print_ = False
ones = False
if ones:
random = False
shapes2 = []
for id, (ishape, kshape, subshape, istride, kstride) in enumerate(shapes):
oshape = [ishape[0]] + [kshape[0]] + list(numpy.asarray(ishape[2:]) -
numpy.asarray(kshape[2:]) +
numpy.asarray([1, 1]))
if oshape[3] > device_prop['maxThreadsDim0']:
continue
if ishape[1] > 1:
continue
if ((kshape[2] * ishape[3] * 4 + numpy.prod(kshape[2:]) * 4) >
(16 * 1024 - 150)):
continue
if subshape == (1, 1):
shapes2.append((ishape, kshape, subshape, istride, kstride))
shapes = shapes2
exec_conv(version, shapes, verbose, random, 'valid',
print_=print_, ones=ones, rtol=1.1e-5)
def test_valid_5():
seed_rng()
shapes = get_valid_shapes()
version = [5]
verbose = 0
random = True
print_ = False
ones = False
if ones:
random = False
shapes2 = []
# print len(shapes)
for id, (ishape, kshape, subshape, istride, kstride) in enumerate(shapes):
oshape = [ishape[0]] + [kshape[0]] + list(numpy.asarray(ishape[2:]) -
numpy.asarray(kshape[2:]) +
numpy.asarray([1, 1]))
if oshape[3] > device_prop['maxThreadsDim0']:
continue
if ((kshape[2] * ishape[3] * 4 + numpy.prod(kshape[2:]) * 4) >
(16 * 1024 - 150)):
continue
if subshape == (1, 1):
shapes2.append((ishape, kshape, subshape, istride, kstride))
shapes = shapes2
# print len(shapes2)
exec_conv(version, shapes, verbose, random, 'valid',
print_=print_, ones=ones, rtol=1.1e-5)
def test_valid_7_8_13():
seed_rng()
shapes = get_valid_shapes()
# This is to test the "new" lower shared memory usage.
shapes.append(((10, 30, 60, 60), (20, 30, 40, 40),
(1, 1), (1, 1), (1, 1)))
version = [7, 8, 13]
verbose = 0
random = True
print_ = False
ones = False
if ones:
random = False
shapes2 = []
# print len(shapes)
for id, (ishape, kshape, subshape, istride, kstride) in enumerate(shapes):
oshape = [ishape[0]] + [kshape[0]] + list(numpy.asarray(ishape[2:]) -
numpy.asarray(kshape[2:]) +
numpy.asarray([1, 1]))
if oshape[2] * oshape[3] > device_prop['maxThreadsDim0']:
continue
if max(numpy.prod(ishape[2:]) * 4 + 2 * kshape[3] * 4,
oshape[2] * oshape[3] * 4 * 2) > (16 * 1024 - 150):
continue
if subshape == (1, 1):
shapes2.append((ishape, kshape, subshape, istride, kstride))
shapes = shapes2
# print len(shapes2)
exec_conv(version, shapes, verbose, random, 'valid',
print_=print_, ones=ones, rtol=1.1e-5)
def test_valid_9_10():
seed_rng()
shapes = get_valid_shapes()
version = [9, 10]
verbose = 0
random = True
print_ = False
ones = False
if ones:
random = False
shapes2 = []
# print len(shapes)
for id, (ishape, kshape, subshape, istride, kstride) in enumerate(shapes):
oshape = [ishape[0]] + [kshape[0]] + list(numpy.asarray(ishape[2:]) -
numpy.asarray(kshape[2:]) +
numpy.asarray([1, 1]))
if oshape[3] > device_prop['maxThreadsDim0']:
continue
if (kshape[3] * 4 + ishape[3]) > (16 * 1024 - 150):
continue
if subshape == (1, 1):
shapes2.append((ishape, kshape, subshape, istride, kstride))
shapes = shapes2
# print len(shapes2)
exec_conv(version, shapes, verbose, random, 'valid',
print_=print_, ones=ones, rtol=1.1e-5)
def test_valid():
seed_rng()
shapes = get_valid_shapes()
#shapes=shapes[400:426]
# I put -1 in case we forget to add version in the test to.
# I put -2 to test the reference version.
version = [-2, -1, 6]
verbose = 0
# version=[1]
random = True
print_ = False
ones = False
if ones:
random = False
exec_conv(version, shapes, verbose, random, 'valid',
print_=print_, ones=ones, rtol=1.1e-5)
def test_full():
seed_rng()
shapes = get_basic_shapes()
shapes += get_shapes2()
#test image stride
shapes += get_shapes2(scales_img=(2, 2), img_stride=(1, 2))
shapes += get_shapes2(scales_img=(2, 2), img_stride=(2, 1))
shapes += get_shapes2(scales_img=(2, 2), img_stride=(2, 2))
shapes += get_shapes2(scales_img=(2, 2), img_stride=(-1, -1))
shapes += get_shapes2(scales_img=(2, 2), kern_stride=(-1, -1))
#test subsample done in a separate fct
shapes += [
#other test
((2, 1, 2, 2), (1, 1, 2, 2), (1, 1), (1, 1), (1, 1))
, ((3, 2, 4, 4), (4, 2, 4, 4), (1, 1), (1, 1), (1, 1))
, ((4, 1, 10, 10), (1, 1, 2, 2), (1, 1), (1, 1), (1, 1))
, ((1, 1, 4, 4), (1, 1, 2, 3), (1, 1), (1, 1), (1, 1))
, ((4, 1, 10, 10), (1, 1, 2, 3), (1, 1), (1, 1), (1, 1))
, ((4, 1, 10, 10), (1, 1, 2, 10), (1, 1), (1, 1), (1, 1))
, ((4, 1, 20, 10), (1, 1, 2, 10), (1, 1), (1, 1), (1, 1))
, ((3, 2, 8, 8), (4, 2, 4, 4), (1, 1), (1, 1), (1, 1)) #stack, nkern, bsize
, ((3, 2, 8, 6), (4, 2, 4, 4), (1, 1), (1, 1), (1, 1)) #stack, nkern, bsize, non-square image
, ((3, 2, 8, 6), (4, 2, 4, 3), (1, 1), (1, 1), (1, 1)) #stack, nkern, bsize, non-square image, non-square kern
, ((3, 2, 8, 6), (4, 2, 4, 6), (1, 1), (1, 1), (1, 1)) #stack, nkern, bsize ,non-square image, non-square kern, kernsize==imgsize on one dim
, ((16, 5, 64, 64), (8, 5, 8, 8), (1, 1), (1, 1), (1, 1)) # a big one
, ((16, 1, 28, 28), (20, 1, 5, 5), (1, 1), (1, 1), (1, 1)) # MNIST LeNET layer 1
, ((20, 16, 32, 32), (1, 16, 28, 28), (1, 1), (1, 1), (1, 1)) # layer 1 backprop to weights
#other test
, ((3, 1, 1, 1), (2, 1, 5, 3), (1, 1), (1, 1), (1, 1))#kernel bigger then image
, ((3, 2, 1, 1), (4, 2, 1, 1), (1, 1), (1, 1), (1, 1))
, ((3, 2, 4, 4), (4, 2, 2, 6), (1, 1), (1, 1), (1, 1))
, ((3, 2, 4, 4), (4, 2, 8, 6), (1, 1), (1, 1), (1, 1))#kernel bigger then image
, ((4, 2, 10, 10), (3, 2, 2, 12), (1, 1), (1, 1), (1, 1))
]
shapes += [
# ((60,1,28,28),(20,1,5,5), (1, 1), (1, 1), (1, 1))#test_lenet_28 1 layers
# , ((60,20,12,12),(30,20,5,5), (1, 1), (1, 1), (1, 1))#test_lenet_28 2 layers
((60,30,8,8),(20,30,5,5), (1, 1), (1, 1), (1, 1))#test_lenet_28 bprop 1 full
# , ((20,60,12,12),(30,60,8,8), (1, 1), (1, 1), (1, 1))#test_lenet_28 bprop 2 valid
# , ((1,60,28,28),(20,60,24,24), (1, 1), (1, 1), (1, 1))#test_lenet_28 bprop 2 valid
# , ((10,1,64,64),(20,1,7,7), (1, 1), (1, 1), (1, 1))#test_lenet_64 1 layers
# , ((10,20,29,29),(30,20,7,7), (1, 1), (1, 1), (1, 1))#test_lenet_64 2 layers
, ((10,30,23,23),(20,30,7,7), (1, 1), (1, 1), (1, 1))#test_lenet_64 full
# , ((20,10,29,29),(30,10,23,23), (1, 1), (1, 1), (1, 1))#test_lenet_64 bprop 1
# , ((1,10,64,64),(20,10,58,58), (1, 1), (1, 1), (1, 1))#test_lenet_64 bprop 2
#Test more than maxThreadsDim0
, ((2,4,13,1050), (3,4,10, 11), (1, 1), (1, 1), (1, 1))
, ((2,4,1050,13), (3,4,10, 11), (1, 1), (1, 1), (1, 1))
]
# shapes=shapes[:277]
version = [-2, -1, 0, 1, 2, 3, 4, 5]
verbose = 0
# version=[4]
random = True
exec_conv(version, shapes, verbose, random, 'full')
def test_subsample():
seed_rng()
# implement when
shapes = [((1, 1, 1, 1), (1, 1, 1, 1), (1, 1), (1, 1), (1, 1)),
((1, 1, 1, 1), (1, 1, 1, 1), (2, 2), (1, 1), (1, 1)),
((4, 2, 10, 10), (3, 2, 2, 2), (1, 3), (1, 1), (1, 1)),
((4, 2, 10, 10), (3, 2, 2, 2), (3, 3), (1, 1), (1, 1)),
((4, 2, 10, 10), (3, 2, 2, 2), (3, 1), (1, 1), (1, 1))
]
shapes += get_shapes2(scales_img=(2, 2), subsample=(1, 1))
shapes += get_shapes2(scales_img=(2, 2), subsample=(1, 2))
shapes += get_shapes2(scales_img=(2, 2), subsample=(2, 1))
shapes += get_shapes2(scales_img=(2, 2), subsample=(2, 2))
#We put only the version that implement the subsample to make the test faster.
version_valid = [-2, -1, 1, 3, 11, 12]
version_full = [-2, -1]
verbose = 0
random = True
print_ = False
ones = False
if ones:
random = False
exec_conv(version_valid, shapes, verbose, random, 'valid',
print_=print_, ones=ones)
exec_conv(version_full, shapes, verbose, random, 'full',
print_=print_, ones=ones)
class TestConv2DGPU(unittest.TestCase):
def test_logical_shapes(self):
seed_rng()
for stride in range(1, 4):
kshp = (10, 2, 10, 10)
featshp = (3, 10, 11, 11)
a = tensor.ftensor4()
A = tensor.ftensor4()
# Need to transpose first two dimensions of kernel, and reverse
# index kernel image dims (for correlation)
kernel_rotated = tensor.transpose(A, axes=[1, 0, 2, 3])
featshp_logical = (featshp[0], featshp[1], featshp[2] * stride,
featshp[3] * stride)
kshp_rotated = (kshp[1], kshp[0], kshp[2], kshp[3])
#print featshp, kshp_rotated, featshp_logical[1:], kshp[2:]
image_estimate = tensor.nnet.conv2d(a, kernel_rotated,
border_mode='full',
image_shape=featshp,
filter_shape=kshp_rotated,
imshp_logical=featshp_logical[1:],
kshp_logical=kshp[2:])
func = theano.function([a, A], image_estimate, mode=mode_with_gpu)
#theano.printing.debugprint(func,)
assert any([isinstance(node.op, theano.sandbox.cuda.blas.GpuConv)
for node in func.maker.fgraph.toposort()])
a_in = numpy.random.randn(*featshp).astype("float32")
A_in = numpy.random.randn(*kshp).astype("float32")
func(a_in, A_in)
def test_invalid_input_shape(self):
"""
Tests that when the shape gived at build time is not the same as
run time we raise an error
"""
seed_rng()
verbose = 0
random = True
print_ = False
ones = False
if ones:
random = False
global mode_with_gpu
mode_with_gpu_orig = mode_with_gpu
try:
if theano.config.mode in ['DebugMode', 'DEBUG_MODE']:
mode_with_gpu = theano.compile.mode.get_mode(
'FAST_RUN').including('gpu')
for mode in ['valid', 'full']:
for shapes in [((3, 2, 8, 8), (4, 2, 5, 5), (8, 8)),
((3, 2, 8, 8), (4, 2, 5, 5), (5, 8)),
#((3, 2, 8, 8), (4, 2, 5, 5), (8, 5)),
# We use only the number of columns.
]:
self.assertRaises(ValueError, _params_allgood,
shapes[0], shapes[1],
verbose=verbose, random=random,
mode=mode,
print_=print_, ones=ones,
compile_kshp=shapes[2])
finally:
mode_with_gpu = mode_with_gpu_orig
def benchmark():
shapes_valid = [
#test_lenet_28 shape
((20, 60,12,12), (30,60,8,8), (1, 1), (1, 1), (1, 1))#valid
,((60, 20,12,12), (30,20,5,5), (1, 1), (1, 1), (1, 1))#valid
,((60, 1,28,28), (20,1,5,5), (1, 1), (1, 1), (1, 1))#valid
,((1, 60,28,28), (20,60,24,24), (1, 1), (1, 1), (1, 1))#valid
#test_lenet_32 shape
,((20, 60,14,14), (30,60,10,10), (1, 1), (1, 1), (1, 1))#valid
,((60, 20,14,14), (30,20,5,5), (1, 1), (1, 1), (1, 1))#valid
,((60, 1,32,32), (20,1,5,5), (1, 1), (1, 1), (1, 1))#valid
,((1, 60,32,32), (20,60,28,28), (1, 1), (1, 1), (1, 1))#valid
#test_lenet_64 shape
,((10, 20,29,29), (30,20,7,7), (1, 1), (1, 1), (1, 1))#valid
,((20, 10,29,29), (30,10,23,23), (1, 1), (1, 1), (1, 1))#valid
,((10, 1,64,64), (20,1,7,7), (1, 1), (1, 1), (1, 1))#valid
,((1, 10,64,64), (20,10,58,58), (1, 1), (1, 1), (1, 1))#valid
#test_lenet_108 shape
,((10, 20,51,51), (30,20,7,7), (1, 1), (1, 1), (1, 1))#valid
,((20, 10,51,51), (30,10,45,45), (1, 1), (1, 1), (1, 1))#valid
,((10, 1,108,108), (20,1,7,7), (1, 1), (1, 1), (1, 1))#valid
,((1, 10,108,108), (20,10,102,102), (1, 1), (1, 1), (1, 1))#valid
#test_lenet_256 shape
,((2, 20,124,124), (30,20,9,9), (1, 1), (1, 1), (1, 1))#valid
,((20, 2,124,124), (30,2,116,116), (1, 1), (1, 1), (1, 1))#valid
,((2, 1,256,256), (20,1,9,9), (1, 1), (1, 1), (1, 1))#valid
,((1, 2,256,256), (20,2,248,248), (1, 1), (1, 1), (1, 1))#valid
]
shapes_full = [
#test_lenet_28 shape
((60, 30,8,8), (20, 30, 5, 5), (1, 1), (1, 1), (1, 1))#full
#test_lenet_32 shape
,((60, 30,10,10), (20, 30, 5, 5), (1, 1), (1, 1), (1, 1))#full conv_full_patch_stack_padded' N=1
#test_lenet_64 shape
,((10, 30,23,23), (20, 30, 7, 7), (1, 1), (1, 1), (1, 1))#full conv_full_patch_stack_padded' N=3
#test_lenet_108 shape
,((10, 30,45,45), (20, 30, 7, 7), (1, 1), (1, 1), (1, 1))#full 'conv_full_patch_stack_padded' N=9
#test_lenet_256 shape
,((2, 30,116,116), (20, 30, 9,9), (1, 1), (1, 1), (1, 1))#full conv_reference_full
]
# shapes_valid=shapes_valid[-1:]
# shapes_full=shapes_full[-1:]
version = [-1]
verbose = 1
random = True
exec_conv(version, shapes_valid, verbose, random, 'valid',
print_=None, rtol=1e-3)
exec_conv(version, shapes_full, verbose, random, 'full')
def test_stack_rows_segfault_070312():
seed_rng()
# 07/03/2012
# Running this unittest with cuda-memcheck exposes an illegal read.
# THEANO_FLAGS=device=gpu cuda-memcheck nosetests \
# test_conv_cuda_ndarray.py:test_stack_rows_segfault_070312
img = theano.shared(numpy.random.rand(1, 80, 96, 96).astype('float32'))
kern = theano.shared(numpy.random.rand(1, 80, 9, 9).astype('float32'))
out = theano.shared(numpy.random.rand(1, 2, 2, 3).astype('float32'))
op = theano.tensor.nnet.conv.ConvOp(imshp=(80, 96, 96), kshp=(9, 9),
nkern=1, bsize=1)
f = theano.function([], [], updates=[(out, op(img, kern))], mode=mode_with_gpu)
f()
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