Skip to content

P
pytensor
提交 9411992c authored 作者: Alexander Matyasko's avatar Alexander Matyasko
浏览文件
操作

Add Max3dPoolGradGrad and padding

Add Max3dPoolGradGrad which accept 5d input and pools over last 3 dimensions. Also add padding parameter and now accept tensor variables to make_node which removes necessity to check for constant args. Local optimization are updated accordingly which now work for 2d and 3d pooling.
上级 a43a8425
隐藏空白字符变更
内嵌 并排
正在显示 包含 155 行增加 和 77 行删除
theano/gpuarray/blas.py
浏览文件 @ 9411992c
...@@ -1542,14 +1542,16 @@ class GpuDownsampleFactorMaxGradGrad(CGpuKernelBase): ...@@ -1542,14 +1542,16 @@ class GpuDownsampleFactorMaxGradGrad(CGpuKernelBase):
Implement the grad of downsample with max on the gpu. Implement the grad of downsample with max on the gpu.
""" """
__props__ = ('ds', 'st', 'ignore_border') __props__ = ('ignore_border', 'mode', 'ndim')
def __init__(self, ds, st=None, ignore_border=False): def __init__(self, ignore_border, mode='max', ndim=2):
self.ds = tuple(ds) self.ndim = ndim
self.st = self.ds if st is None else tuple(st)
self.ignore_border = ignore_border self.ignore_border = ignore_border
self.mode = mode
CGpuKernelBase.__init__(self, ['pool_grad_grad.c'], CGpuKernelBase.__init__(self, ['pool_grad_grad.c'],
'APPLY_SPECIFIC(pool_grad_grad)') 'APPLY_SPECIFIC(pool_grad_grad)')
assert self.mode == 'max'
assert self.ndim in [2, 3]
def c_headers(self): def c_headers(self):
return ['gpuarray/types.h', 'gpuarray/array.h', 'gpuarray/kernel.h', return ['gpuarray/types.h', 'gpuarray/array.h', 'gpuarray/kernel.h',
...@@ -1559,31 +1561,29 @@ class GpuDownsampleFactorMaxGradGrad(CGpuKernelBase): ...@@ -1559,31 +1561,29 @@ class GpuDownsampleFactorMaxGradGrad(CGpuKernelBase):
def c_header_dirs(self): def c_header_dirs(self):
return [os.path.dirname(__file__), pygpu.get_include()] return [os.path.dirname(__file__), pygpu.get_include()]
def make_node(self, x, z, gx): def make_node(self, inp, out, out_grad, ws, stride, pad):
ctx_name = infer_context_name(x, z, gx) ctx_name = infer_context_name(inp, out, out_grad)
x = as_gpuarray_variable(x, ctx_name) inp = as_gpuarray_variable(inp, ctx_name)
z = as_gpuarray_variable(z, ctx_name) assert (inp.ndim in [4, 5])
gx = as_gpuarray_variable(gx, ctx_name) out = as_gpuarray_variable(out, ctx_name)
assert (out_grad.ndim in [4, 5])
out_grad = as_gpuarray_variable(out_grad, ctx_name)
assert(out.ndim in [4, 5])
assert (out_grad.ndim == inp.ndim)
assert (inp.ndim == out.ndim)
if x.type.ndim != 4: ws = as_tensor_variable(ws)
raise TypeError('x must be 4D tensor') stride = as_tensor_variable(stride)
if z.type.ndim != 4: pad = as_tensor_variable(pad)
raise TypeError('z must be 4D tensor') assert ws.type.ndim == stride.type.ndim and ws.type.ndim == pad.type.ndim
if gx.type.ndim != 4: assert ws.type.ndim == 1
raise TypeError('gx must be 4D tensor')
return Apply(self, [x, z, gx], [x.type()]) return Apply(self, [inp, out, out_grad, ws, stride, pad], [inp.type()])
def get_params(self, node): def get_params(self, node):
return node.inputs[0].type.context return node.inputs[0].type.context
def get_op_params(self):
ds0, ds1 = self.ds
st0, st1 = self.st
ignore_border = int(self.ignore_border)
return [('DS0', ds0), ('DS1', ds1), ('ST0', st0), ('ST1', st1),
('IGNORE_BORDER', ignore_border)]
@inplace_allocempty(GpuGemv, 0) @inplace_allocempty(GpuGemv, 0)
def local_inplace_gpuagemv(node, inputs): def local_inplace_gpuagemv(node, inputs):
......
theano/gpuarray/opt.py
浏览文件 @ 9411992c
...@@ -1595,28 +1595,26 @@ def local_gpua_lift_abstractconv_graph(op, context_name, inputs, outputs): ...@@ -1595,28 +1595,26 @@ def local_gpua_lift_abstractconv_graph(op, context_name, inputs, outputs):
@op_lifter([pool.DownsampleFactorMaxGradGrad]) @op_lifter([pool.DownsampleFactorMaxGradGrad])
@register_opt2([pool.DownsampleFactorMaxGradGrad]) @register_opt2([pool.DownsampleFactorMaxGradGrad])
def local_gpu_downsample_factor_max_grad_grad(op, ctx_name, inputs, outputs): def local_gpu_downsample_factor_max_grad_grad(op, ctx_name, inputs, outputs):
from theano.sandbox.cuda.opt import _check_constant_args_pool
assert op.__props__ == ('ignore_border', 'mode', 'ndim') assert op.__props__ == ('ignore_border', 'mode', 'ndim')
inp, out, out_grad, ws, st, pad = inputs inp, out, out_grad, ws, stride, pad = inputs
nd = op.ndim nd = op.ndim
ret = _check_constant_args_pool(nd, ws, st, pad, op) if nd not in (2, 3):
if ret is None:
return
ws, st, pad = ret
if nd != 2 or max(pad) != 0 or op.mode != 'max':
return return
inp = gpu_contiguous(as_gpuarray_variable(inp, ctx_name)) inp = gpu_contiguous(as_gpuarray_variable(inp, ctx_name))
out = gpu_contiguous(as_gpuarray_variable(out, ctx_name)) out = gpu_contiguous(as_gpuarray_variable(out, ctx_name))
out_grad = gpu_contiguous(as_gpuarray_variable(out_grad, ctx_name)) out_grad = gpu_contiguous(as_gpuarray_variable(out_grad, ctx_name))
op = GpuDownsampleFactorMaxGradGrad(ws, st, op.ignore_border)
op = GpuDownsampleFactorMaxGradGrad(op.ignore_border, op.mode, op.ndim)
if inp.ndim == nd + 2: if inp.ndim == nd + 2:
return op(inp, out, out_grad) return op(inp, out, out_grad, ws, stride, pad)
else: else:
inp_4D = pad_dims(inp, 2, 2) # reshape to 4D or 5D with 2 non-pooling dimensions
out_4D = pad_dims(out, 2, 2) inp_padded = pad_dims(inp, 2, nd)
out_grad_4D = pad_dims(out_grad, 2, 2) out_padded = pad_dims(out, 2, nd)
output_4D = op(inp_4D, out_4D, out_grad_4D) out_grad_padded = pad_dims(out_grad, 2, nd)
return unpad_dims(output_4D, inp, 2, 2) ret_padded = op(inp_padded, out_padded, out_grad_padded,
ws, stride, pad)
return unpad_dims(ret_padded, inp, 2, nd)
@register_opt("low_memory") @register_opt("low_memory")
......
theano/gpuarray/pool_grad_grad.c
浏览文件 @ 9411992c
#section kernels #section kernels
#kernel pool_grad_grad_kernel : size, size, size, size, size, size, size, *, *, *, size, size, size, size, * : #kernel max_pool2d_grad_grad_kernel : size, size, size, size, size, size, size, *, *, *, size, size, size, size, size, size, * :
KERNEL void pool_grad_grad_kernel(const ga_size nthreads, KERNEL void max_pool2d_grad_grad_kernel(const ga_size nthreads,
const ga_size num, const ga_size channels, const ga_size pooled_height, const ga_size num, const ga_size channels, const ga_size pooled_height,
const ga_size pooled_width, const ga_size height, const ga_size width, const ga_size pooled_width, const ga_size height, const ga_size width,
GLOBAL_MEM const DTYPE_i0 *x, GLOBAL_MEM const DTYPE_i1 *z, GLOBAL_MEM const DTYPE_i2 *gx, GLOBAL_MEM const DTYPE_i0 *x, GLOBAL_MEM const DTYPE_i1 *z, GLOBAL_MEM const DTYPE_i2 *gx,
const ga_size kernel_h, const ga_size kernel_w, const ga_size stride_h, const ga_size stride_w, const ga_size kernel_h, const ga_size kernel_w, const ga_size stride_h, const ga_size stride_w,
const ga_size pad_h, const ga_size pad_w,
GLOBAL_MEM DTYPE_o0 *gz) GLOBAL_MEM DTYPE_o0 *gz)
{ {
// grid stride looping // grid stride looping
...@@ -16,9 +17,11 @@ KERNEL void pool_grad_grad_kernel(const ga_size nthreads, ...@@ -16,9 +17,11 @@ KERNEL void pool_grad_grad_kernel(const ga_size nthreads,
const ga_size ph = (index / pooled_width) % pooled_height; const ga_size ph = (index / pooled_width) % pooled_height;
const ga_size c = (index / pooled_width / pooled_height) % channels; const ga_size c = (index / pooled_width / pooled_height) % channels;
const ga_size n = (index / pooled_width / pooled_height / channels); const ga_size n = (index / pooled_width / pooled_height / channels);
const ga_size hstart = ph*stride_h; ga_int hstart = static_cast<ga_int>(ph*stride_h) - static_cast<ga_int>(pad_h);
hstart = max(hstart, 0);
const ga_size hend = min(hstart + kernel_h, height); const ga_size hend = min(hstart + kernel_h, height);
const ga_size wstart = pw*stride_w; ga_int wstart = static_cast<ga_int>(pw*stride_w) - static_cast<ga_int>(pad_w);
wstart = max(wstart, 0);
const ga_size wend = min(wstart + kernel_w, width); const ga_size wend = min(wstart + kernel_w, width);
const ga_size offset = (n*channels + c) * height * width; const ga_size offset = (n*channels + c) * height * width;
...@@ -38,44 +41,102 @@ KERNEL void pool_grad_grad_kernel(const ga_size nthreads, ...@@ -38,44 +41,102 @@ KERNEL void pool_grad_grad_kernel(const ga_size nthreads,
gz[index] = gradient; gz[index] = gradient;
} }
} }
#kernel max_pool3d_grad_grad_kernel : size, size, size, size, size, size, size, size, size, *, *, *, size, size, size, size, size, size, size, size, size, * :
KERNEL void max_pool3d_grad_grad_kernel(const ga_size nthreads,
const ga_size num, const ga_size channels, const ga_size pooled_depth,
const ga_size pooled_height, const ga_size pooled_width,
const ga_size depth, const ga_size height, const ga_size width,
GLOBAL_MEM const DTYPE_i0 *x, GLOBAL_MEM const DTYPE_i1 *z, GLOBAL_MEM const DTYPE_i2 *gx,
const ga_size kernel_d, const ga_size kernel_h, const ga_size kernel_w,
const ga_size stride_d, const ga_size stride_h, const ga_size stride_w,
const ga_size pad_d, const ga_size pad_h, const ga_size pad_w,
GLOBAL_MEM DTYPE_o0 *gz)
{
// grid stride looping
for (ga_size index = GID_0 * LDIM_0 + LID_0;
index < nthreads; index += LDIM_0 * GDIM_0) {
const ga_size pw = index % pooled_width;
const ga_size ph = (index / pooled_width) % pooled_height;
const ga_size pd = (index / pooled_width / pooled_height) % pooled_depth;
const ga_size c = (index / pooled_width / pooled_height / pooled_depth) % channels;
const ga_size n = (index / pooled_width / pooled_height / pooled_depth / channels);
ga_int dstart = static_cast<ga_int>(pd*stride_d) - static_cast<ga_int>(pad_d);
dstart = max(dstart, 0);
const ga_size dend = min(dstart + kernel_d, depth);
ga_int hstart = static_cast<ga_int>(ph*stride_h) - static_cast<ga_int>(pad_h);
hstart = max(hstart, 0);
const ga_size hend = min(hstart + kernel_h, height);
ga_int wstart = static_cast<ga_int>(pw*stride_w) - static_cast<ga_int>(pad_w);
wstart = max(wstart, 0);
const ga_size wend = min(wstart + kernel_w, width);
const ga_size offset = (n*channels + c) * depth * height * width;
const DTYPE_i0* x_slice = x + offset;
const DTYPE_i2* gx_slice = gx + offset;
DTYPE_o0 gradient = 0;
for (ga_size d=dstart; d < dend; ++d) {
for (ga_size h=hstart; h < hend; ++h) {
for (ga_size w=wstart; w < wend; ++w) {
// maximum in the region
if (z[index] == x_slice[(d * height + h) * width + w]) {
gradient += gx_slice[(d * height + h)* width + w];
}
}
}
}
gz[index] = gradient;
}
}
#section support_code_struct #section support_code_struct
int APPLY_SPECIFIC(pool_grad_grad)(PyGpuArrayObject *x, int APPLY_SPECIFIC(pool_grad_grad)(PyGpuArrayObject *x,
PyGpuArrayObject *z, PyGpuArrayObject *z,
PyGpuArrayObject *gx, PyGpuArrayObject *gx,
PyArrayObject *ws,
PyArrayObject *stride,
PyArrayObject *pad,
PyGpuArrayObject **gz, PyGpuArrayObject **gz,
PyGpuContextObject *ctx) { PyGpuContextObject *ctx) {
if (PyGpuArray_NDIM(x) != 4 if (!GpuArray_IS_C_CONTIGUOUS(&x->ga)
|| PyGpuArray_NDIM(z) != 4 || !GpuArray_IS_C_CONTIGUOUS(&z->ga)
|| PyGpuArray_NDIM(gx) != 4) || !GpuArray_IS_C_CONTIGUOUS(&gx->ga))
{ {
PyErr_SetString(PyExc_ValueError, "GpuDownsampleFactorMaxGradGrad: rank error"); PyErr_Format(PyExc_ValueError,
"GpuPoolingGradGrad: requires data to be C-contiguous");
return 1; return 1;
} }
if (NULL == *gz || theano_size_check(*gz, 4, PyGpuArray_DIMS(z), z->ga.typecode)) size_t ndims = PyArray_DIM(ws, 0);
if (PyGpuArray_NDIM(x) != ndims + 2
|| PyGpuArray_NDIM(z) != ndims + 2
|| PyGpuArray_NDIM(gx) != ndims + 2)
{ {
Py_XDECREF(*gz); PyErr_SetString(PyExc_ValueError, "GpuPoolingGradGrad: rank error");
*gz = pygpu_zeros(4, PyGpuArray_DIMS(z), return 1;
z->ga.typecode, GA_C_ORDER,
ctx, Py_None);
if (NULL == *gz)
{
PyErr_SetString(PyExc_RuntimeError,
"GpuDownsampleFactorMaxGradGrad: failed to allocate memory");
return 1;
}
} }
if (!GpuArray_IS_C_CONTIGUOUS(&x->ga) if (theano_prep_output(gz, PyGpuArray_NDIM(z), PyGpuArray_DIMS(z),
|| !GpuArray_IS_C_CONTIGUOUS(&z->ga) z->ga.typecode, GA_C_ORDER, ctx) != 0)
|| !GpuArray_IS_C_CONTIGUOUS(&gx->ga)
|| !GpuArray_IS_C_CONTIGUOUS(&(*gz)->ga))
{ {
PyErr_Format(PyExc_ValueError, PyErr_SetString(PyExc_RuntimeError,
"GpuDownsampleFactorMaxGradGrad: requires data to be C-contiguous"); "GpuPoolingGradGrad: failed to allocate memory");
return 1; return 1;
} }
{ {
// scope for running kernel // scope for running kernel
size_t w[3];
size_t s[3];
size_t p[3];
for(int i = 0; i < ndims; i++) {
w[i] = *((npy_intp*)PyArray_GETPTR1(ws, i));
s[i] = *((npy_intp*)PyArray_GETPTR1(stride, i));
p[i] = *((npy_intp*)PyArray_GETPTR1(pad, i));
}
size_t max_threads_dim; size_t max_threads_dim;
int err; int err;
const size_t* z_dims = PyGpuArray_DIMS(z); const size_t* z_dims = PyGpuArray_DIMS(z);
...@@ -87,22 +148,41 @@ int APPLY_SPECIFIC(pool_grad_grad)(PyGpuArrayObject *x, ...@@ -87,22 +148,41 @@ int APPLY_SPECIFIC(pool_grad_grad)(PyGpuArrayObject *x,
PyErr_SetString(PyExc_RuntimeError, "Could not fetch max_threads_dims"); PyErr_SetString(PyExc_RuntimeError, "Could not fetch max_threads_dims");
return 1; return 1;
} }
size_t num_kernels = z_dims[0] * z_dims[1] * z_dims[2] * z_dims[3];
size_t threads_per_block = max_threads_dim; size_t threads_per_block = max_threads_dim;
size_t n_blocks = (num_kernels + threads_per_block - 1) / threads_per_block;
err = pool_grad_grad_kernel_call(1, &n_blocks, &threads_per_block, 0, if (ndims == 2) {
num_kernels, size_t num_kernels = z_dims[0] * z_dims[1] * z_dims[2] * z_dims[3];
z_dims[0], z_dims[1], z_dims[2], z_dims[3], size_t n_blocks = (num_kernels + threads_per_block - 1) / threads_per_block;
x_dims[2], x_dims[3], err = max_pool2d_grad_grad_kernel_call(1, &n_blocks, &threads_per_block, 0,
x->ga.data, z->ga.data, gx->ga.data, num_kernels,
DS0, DS1, ST0, ST1, z_dims[0], z_dims[1], z_dims[2], z_dims[3],
(*gz)->ga.data); x_dims[2], x_dims[3],
if (err != GA_NO_ERROR) { x->ga.data, z->ga.data, gx->ga.data,
PyErr_Format(PyExc_RuntimeError, w[0], w[1], s[0], s[1], p[0], p[1],
"GpuDownsampleFactorMaxGradGrad: %s.", (*gz)->ga.data);
GpuKernel_error(&k_pool_grad_grad_kernel, err)); if (err != GA_NO_ERROR) {
return 1; PyErr_Format(PyExc_RuntimeError,
"GpuPoolingGradGrad: max_pool2d_grad_grad_kernel %s.",
GpuKernel_error(&k_max_pool2d_grad_grad_kernel, err));
return 1;
}
}
else if (ndims == 3) {
size_t num_kernels = z_dims[0] * z_dims[1] * z_dims[2] * z_dims[3] * z_dims[4];
size_t n_blocks = (num_kernels + threads_per_block - 1) / threads_per_block;
err = max_pool3d_grad_grad_kernel_call(1, &n_blocks, &threads_per_block, 0,
num_kernels,
z_dims[0], z_dims[1], z_dims[2], z_dims[3], z_dims[4],
x_dims[2], x_dims[3], x_dims[4],
x->ga.data, z->ga.data, gx->ga.data,
w[0], w[1], w[2], s[0], s[1], s[2], p[0], p[1], p[2],
(*gz)->ga.data);
if (err != GA_NO_ERROR) {
PyErr_Format(PyExc_RuntimeError,
"GpuPoolingGradGrad: max_pool3d_grad_grad_kernel %s.",
GpuKernel_error(&k_max_pool3d_grad_grad_kernel, err));
return 1;
}
} }
} }
return 0; return 0;
......
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