def test_convolution(): imgSize = 32 filterSize = 5 padding = 2 color = 1 imgNum = 1 filterNum = 64 stride = 1 modulesX = 1 + int(((2 * padding + imgSize - filterSize) / float(stride))) print 'Modules X', modulesX img = gpuarray.to_gpu(np.ones((imgSize * imgSize * color, imgNum)).astype(np.float32)) filter = gpuarray.to_gpu(np.ones((filterSize * filterSize * color, filterNum)).astype(np.float32)) target = gpuarray.to_gpu(np.ones((modulesX * modulesX * filterNum, imgNum)).astype(np.float32)) print 'standard output for convolution' print convolve2d(np.ones((imgSize, imgSize)).astype(np.float32), np.ones((filterSize, filterSize)).astype(np.float32),'valid') cudaconv2.convFilterActs(img, filter, target, imgSize, modulesX, modulesX, -padding, stride, color, 1, 0.0, 1.0) print 'pycuda output for convolution' atarget = target.get() print atarget
def fprop(self, input, output):
cudaconv2.convFilterActs(input, self.filter, output, self.imgSize, self.outputSize,
self.outputSize, -self.padding, self.stride, self.numColor, 1)
self.tmp = gpuarray.empty((self.numFilter, self.get_single_img_size() * self.batchSize/self.numFilter), dtype=np.float32)
gpu_copy_to(output, self.tmp)
add_vec_to_rows(self.tmp, self.bias)
gpu_copy_to(self.tmp, output)
def fprop(self, input, output, train=TRAIN):
cudaconv2.convFilterActs(input, self.weight, output, self.imgSize,
self.outputSize, self.outputSize,
-self.padding, self.stride, self.numColor, 1)
self.tmp = gpuarray.empty(
(self.numFilter,
self.get_single_img_size() * self.batchSize / self.numFilter),
dtype=np.float32)
gpu_copy_to(output, self.tmp)
add_vec_to_rows(self.tmp, self.bias)
gpu_copy_to(self.tmp, output)
if PFout:
print_matrix(output, self.name)
def convFilterActs(input, weight, output, bias, padding, stride):
from distbase import cuda_base
image_y = input.shape[ConvDataLayout.HEIGHT]
output_y = output.shape[ConvDataLayout.HEIGHT]
output_x = output.shape[ConvDataLayout.WIDTH]
color = input.shape[ConvDataLayout.CHANNEL]
cudaconv2.convFilterActs(input, weight, output, image_y, output_y, output_x, padding, stride,
color, 1)
batch_size = output.shape[ConvDataLayout.BATCH]
channel = output.shape[ConvDataLayout.CHANNEL]
# bias term
cuda_base.add_vec_to_rows(output.reshape((channel, output_y * output_x * batch_size)), bias)
def fprop(self, input, output, train=TRAIN):
#np.save('input.arr', input.get())
#np.save('weight.arr', self.weight.wt.get())
cudaconv2.convFilterActs(input, self.weight.wt, output, self.img_size, self.outputSize,
self.outputSize, -self.padding, self.stride, self.numColor, 1)
#util.log_info('%s', output.get().mean())
self.tmp = gpuarray.empty((self.numFilter,
self.get_single_img_size() * self.batch_size / self.numFilter),
dtype=np.float32)
gpu_copy_to(output, self.tmp)
add_vec_to_rows(self.tmp, self.bias.wt)
gpu_copy_to(self.tmp, output)
if PFout:
print_matrix(output, self.name)
def fprop(self, input, output, train=TRAIN):
#np.save('input.arr', input.get())
#np.save('weight.arr', self.weight.wt.get())
cudaconv2.convFilterActs(input, self.weight.wt, output, self.img_size, self.outputSize,
self.outputSize, -self.padding, self.stride, self.numColor, 1)
#util.log_info('%s', output.get().mean())
self.tmp = gpuarray.empty((self.numFilter,
self.get_single_img_size() * self.batch_size / self.numFilter),
dtype=np.float32)
gpu_copy_to(output, self.tmp)
add_vec_to_rows(self.tmp, self.bias.wt)
gpu_copy_to(self.tmp, output)
if PFout:
print_matrix(output, self.name)
padding = 2
color = 1
imgNum = 1
filterNum = 64
stride = 1
modulesX = 1 + int(((2 * padding + imgSize - filterSize) / float(stride)))
img = gpuarray.to_gpu(np.ones((imgSize * imgSize * color, imgNum)).astype(np.float32))
filter = gpuarray.to_gpu(np.ones((filterSize * filterSize * color, filterNum)).astype(np.float32))
target = gpuarray.to_gpu(np.ones((modulesX * modulesX * filterNum, imgNum)).astype(np.float32))
print 'standard output for convolution'
print convolve2d(np.ones((imgSize, imgSize)).astype(np.float32), np.ones((filterSize, filterSize)).astype(np.float32),'valid')
cudaconv2.convFilterActs(img, filter, target, imgSize, modulesX, modulesX, -padding, stride, color, 1, 0.0, 1.0)
print 'pycuda output for convolution'
print target.get()
#from pycuda.compiler import *
#mod = SourceModule(open('foo.cu').read(), no_extern_c=True, include_dirs=['/home/justin/guppy/include'])
#kernel = mod.get_function('kernel')
#def i(x): return np.int32(x)
#
#grid = (1, 32 * 32 * 64 / (4 * 8), 1)
#blocks = (32, 4, 1)
#kernel(img, filter, target, i(1), i(64), i(32), i(32), i(5), i(-2), i(1), i(32), i(32), i(1),
# np.float32(0.0), np.float32(1.0), np.int32(True), block=blocks, grid=grid)
imgNum = 1
filterNum = 64
stride = 1
modulesX = 1 + int(((2 * padding + imgSize - filterSize) / float(stride)))
print 'Modules X', modulesX
img = gpuarray.to_gpu(np.ones((imgSize * imgSize * color, imgNum)).astype(np.float32))
filter = gpuarray.to_gpu(np.ones((filterSize * filterSize * color, filterNum)).astype(np.float32))
target = gpuarray.to_gpu(np.ones((modulesX * modulesX * filterNum, imgNum)).astype(np.float32))
print 'standard output for convolution'
print convolve2d(np.ones((imgSize, imgSize)).astype(np.float32), np.ones((filterSize, filterSize)).astype(np.float32),'valid')
cudaconv2.convFilterActs(img, filter, target, imgSize, modulesX, modulesX, -padding, stride, color, 1, 0.0, 1.0)
print 'pycuda output for convolution'
atarget = target.get()
print atarget
#for i in range(atarget.shape[0]):
# print atarget[i, 0]
#from pycuda.compiler import *
#mod = SourceModule(open('foo.cu').read(), no_extern_c=True, include_dirs=['/home/justin/guppy/include'])
#kernel = mod.get_function('kernel')
#def i(x): return np.int32(x)
#
#grid = (1, 32 * 32 * 64 / (4 * 8), 1)